452 pieces of evidence, each with its source. 98 are named deployments or case studies; the rest are surveys (161), studies (87), filings (47), incidents (34) and benchmarks (25). Most evidence so far: Banking & financial services.
Inc42 investor survey (85+ institutional investors)
~63% see some froth in Indian AI valuations; vertical AI applications seen as top value-capture segment (39%)
Survey in Inc42's Q3 analysis; Indian AI funding was $438M across 35 deals in Q3 2026, up 265% YoY.
Gartner (supply chain practice)
63% expect significant performance improvement within three years; 25% have redesigned jobs because of AI
Gartner 2026 CPO Survey of 213 senior procurement leaders and a C-Suite AI Survey of 1,303 respondents (134 in procurement), both closed April 2026. Headline is that role redesign lags expectations.
Bessemer Venture Partners with Bain & Company (reported by TechTarget)
~50% cutting or planning cuts; 8-13% average reduction; 73% of provider cuts in revenue cycle and billing; 71% of payer cuts in claims
secondary: TechTarget on the same 226-executive survey. 67% of providers use autonomous agents in revenue cycle, 4% in clinical work; 36% of providers delayed or cancelled clinical AI because of regulatory uncertainty. Pharma least likely to cut (44%). Planned, not realised, cuts.
PwC US
No quantified saving; assurance leader: 'It's not that drastic yet'; nearly $1.5B accelerated AI investment
secondary: CFO Dive. Enterprise generative AI tools in use on audits since October 2025 (document review, committee-communication analysis, process documentation); the executive denies three months of work has dropped to one. A candid counterweight to vendor claims.
Brookings (Humlum and Vestergaard)
Where employers encourage use and provide tools and training, 93% of workers used chatbots and 19% report saving over an hour a day vs ~41% use with no employer action; 39% of workers in encouraging workplaces had training
Time savings are self-reported; study finds no earnings or hours effect still. Brookings summary of the Danish study; page opened.
Boston Consulting Group (via Consulting.us)
48.5% generate measurable value; future-built share 5% to 7.5%; time to impact about 11 months vs 17 months for laggards
secondary: trade-press report of BCG's 2026 index (n=1,330). The jump from 2025 partly reflects a reframed segmentation. Source: research report of 14 Oct 2026.
Concentrix (via Pulse 2.0)
Revenue $2.45B (-1.2%, -0.5% constant currency); $1.05B non-cash goodwill impairment; GAAP operating loss $910.3M; FY26 guide $9.827B-$9.877B
secondary: Press-release coverage. About half of revenue now comes from business won and deployed in the last three years, per the CEO. Impairment was tied to the recent trading range and market capitalisation.
AWS / Strand Partners (Vietnam)
26% of businesses (about 245,000), up from 18% in 2025
AWS-commissioned study by Strand Partners surveying 1,000 business leaders and 1,000 members of the public in Vietnam, as reported by Viet Nam News on 1 Oct 2026 and VietnamPlus on 6 Oct 2026. Primary AWS page not found; fieldwork dates not published.
AWS / Strand Partners (Vietnam)
Financial services adoption 41% vs 26% across all sectors; 79% of adopters reported productivity gains; healthcare adoption 23%
Sector cut of the AWS/Strand Vietnam study.
AWS / Strand Partners (Vietnam)
38% aware of agentic AI; 8% embedded agents in core workflows; 19% experimenting or piloting; 81% of agentic AI users reported productivity gains
Agentic AI cut of the AWS/Strand Vietnam study.
AWS / Strand Partners (Vietnam)
54% cite digital/AI skills shortages as a barrier; 87% of employers say reskilling is important; 26% of employees trained in past year (19% in 2025)
Skills and training findings from the AWS/Strand Vietnam study. Relevant to executive education demand.
Techcom Securities (TCBS), Vietnam
Over 460 employees use Kiro; development time cut by nearly 25%; 70% of AI-generated code passed first-round reviews
TCBS deployed AWS's Kiro agentic development environment, as reported in Viet Nam News coverage of the AWS Vietnam report launch. Results are as reported alongside the AWS release.
OmiGroup (Vietnam)
Identified 15 of 17 molecular mechanisms linked to Type-2 diabetes, versus one each for two alternative AI approaches
AI-powered chronic disease research tool built on AWS; result as reported in the AWS Vietnam launch coverage.
NoteX (Vietnam)
30% reduction in AI processing costs and latency; over one million users across Vietnam and 20+ countries in first year
NoteX voice translation platform on AWS (Bedrock and SageMaker), as reported by Viet Nam News; vendor-reported.
Diaflow (Vietnam)
Over 40,000 users (75% in the US); more than 100 million AI tasks executed
Enterprise AI agent orchestration service built on Amazon Bedrock, as reported in Viet Nam News.
AWS / Strand Partners (Vietnam)
Of adopters: 61% still experimenting; 71% broke even or better; 23% formal AI strategy; 13% clearly defined, consistently applied ROI framework; 46% lack reliable way to measure returns
Re-verified wording via Vietnam News (1 Oct 2026) and VietnamPlus/VNA (6 Oct 2026); 1,000 business leaders and 1,000 public. Fieldwork dates not published in any copy seen. AWS-commissioned, self-reported by respondents. Primary AWS/Strand page not found by search.
Accenture
FY26 revenue $74.18B (+6% USD, +5% local); bookings $84.5B; Q4 revenue $18.68B; headcount about 814,000
Press release exhibit (8-K) for the year ended 31 Aug 2026. It reports no quantified AI bookings or revenue, only qualitative AI strategy. Q4 bookings $22.2B (book-to-bill 1.2).
Accenture (earnings call, via Benzinga/Webull)
Fixed-price incl. outcome-based 65% of bookings; managed services bookings record $12.8B; over 400 clients started advanced AI work; nearly 110,000 AI and data staff
secondary: Call transcript summary (page date taken from the 10/01 stamp and content naming FY26). CFO said pricing was stable for the year but lower in many areas in Q4; AI efficiencies are being passed to clients and, she said, more than offset by new work. Management would not confirm an analyst's 5% headcount-growth figure.
Globe Telecom (Philippines)
PHP1.2bn ($19.2m) measurable benefit as of Aug 2026; >PHP825m incremental revenue; PHP368m cost avoidance; 2.8x return
Company-reported via TelecomTV. Investment base and calculation method not disclosed in the article, so the 2.8x cannot be checked. One of few Asian firms publicly reporting a financial AI return.
Bessemer Venture Partners with Bain & Company
226 executives, 65 use cases: avg 3.5x return; 54% material ROI in year one; revenue-cycle 4.0x; pharma clinical development 2.3x
Investor-and-consultant survey of providers, payers and pharma; self-reported by executives and sponsored by a VC with a stake in the category. Returns met or exceeded expectations in about 40% of use cases. Sample fell from 408 the prior year. Treat as upper-bound sentiment, not audited ROI.
Concentrix (via Parameter.io)
Stock down about 11% premarket to $22.11, near 52-week low of $19.12; Q4 constant-currency revenue -3% to -5%
secondary: Market-news summary. CEO said over 50% of revenue is now AI-influenced or newly transformed client work and that the firm is 'aggressively disrupting our own traditional business'; management tied the weak outlook to faster client AI adoption. Adjusted EPS $2.92 beat the $2.71 consensus.
Revalize / TEAM LEWIS / OpinionRoute (You Can't Scale AI on Ambition Alone)
49% measure AI ROI rigorously; 50% track informally or inconsistently; 72% say they can attribute outcomes to AI; 500 decision makers (DACH, UK, US), fielded 7-17 Jul 2026
secondary: reprint of the vendor press release (Revalize sells CPQ/PLM/CAD software), so vendor-sponsored. C-suite leaders were more than twice as likely as senior managers to report advanced AI adoption.
Revalize / TEAM LEWIS / OpinionRoute (You Can't Scale AI on Ambition Alone)
No measurable financial impact: 3% to 13% over the year; broad or advanced adoption 36% (Q1 2026) to 60% about six months later; 44% accelerating AI to offset tariff and cost pressure
secondary: reprint of the vendor press release; vendor-sponsored survey of 500 manufacturing decision makers at firms with 100+ employees. Adoption and the 'no impact' share are both rising, which the release does not reconcile.
Gartner (IT practice)
70% of enterprises predicted to abandon agentic AI built through vendor forward-deployed engineering by 2028
Gartner prediction, not survey data: no sample or method given. Argues engagements fail structurally (scope, incentives, governance, exit terms) before technically, and warns of 'FDE washing' where ordinary consulting is sold at premium rates. Relevant to buyers of outsourced AI builds.
Agoda / Macrame Consulting (7 Asian markets)
Vietnam: 63% save 7+ hrs/week, 43% save 10+; all markets 55% save 7+ hrs (18% in 2025); 79% require human approval before deployment
Survey of 800+ developers and engineering leaders, fieldwork June-Aug 2026, commissioned by Agoda (vendor-commissioned, self-reported time savings). Vietnam News 25 Sep 2026.
KPMG
Nearly 6 in 10 leaders report measurable business value; value cited: productivity 55%, faster decisions 49%, customer and employee experience 38%, financial performance 37% (n=314 C-suite at $1B+ US firms, fielded 24 Jul-25 Aug 2026)
Self-reported survey by an advisory firm; primary KPMG page opened. Exact share behind 'nearly 6 in 10' not given.
KPMG
62% building or deploying agents (53% prior quarter); multi-agent 25% (6%); significant workforce adoption 44% (23% last quarter, 10% a year ago); agent controls built in 30% vs 43% two quarters earlier; 74% include cost reviews in AI approvals; 43% use token budgets
Self-reported (n=314, $1B+ US firms); primary page opened.
KPMG
Banking n=100 ($1B+ US): 65% report measurable value (productivity and efficiency 97%, financial 67%); 80% cite data readiness and access as the top barrier to agents (63% prior quarter); 75% bar agents from sensitive data without human oversight (51%)
Self-reported sector cut fielded 10 Aug-14 Sep 2026; primary page opened.
KPMG
Asset management and private equity n=100: 81% report measurable value; average financial value realised $179M and 47% report over $100M; agent deployment 30% (19% in Q2); agent pilots 34% (52% in Q2); 98% upskilling
Self-reported dollar values from large firms, no method or baseline given; primary page opened. Why pilots fell is not explained in the source.
US Census Bureau Business Trends and Outlook Survey (analysis by StartBusinessByState)
21.1% (+/-0.9) for surveys 27 Jul-6 Sep 2026 vs 12.3% for 3 Nov-14 Dec 2025 (+71%)
secondary: third-party analysis of official Census BTOS data (experimental product); date is the release of the latest survey wave, analysis downloaded 28 Sep 2026. Census asks whether the firm used AI in the past two weeks, so this measures use, not value.
US Census Bureau Business Trends and Outlook Survey (analysis by StartBusinessByState)
15.2% (+/-0.8) vs 8.5% in late 2025 (+80%); latest single wave 15.4%
secondary: third-party analysis of Census BTOS data; date is the latest wave release. Use rate over the previous two weeks, not outcomes.
US Census Bureau Business Trends and Outlook Survey (analysis by StartBusinessByState)
10.4% (+/-1.0) vs 7.0% in late 2025 (+48%); latest single wave 13.0%; truck transportation 8.1%
secondary: third-party analysis of Census BTOS data. Lowest of the industries in this set.
US Census Bureau Business Trends and Outlook Survey (analysis by StartBusinessByState)
16.5% (+/-3.9) vs 15.4% in late 2025 (+7%, within margin of error); latest single wave 11.8%; agriculture (excl. farms) single wave 6.2%
secondary: third-party analysis of Census BTOS data. Utilities started higher than most sectors but did not move; agriculture figure excludes farms and no three-wave average was published.
EY US
PCAOB inspection findings rate 5% (28% in 2024); AI tools given to 100% of US audit professionals; $1B technology and talent investment
secondary: Accounting Today on EY's quality report. EY credits AI and process change; the page cannot isolate AI's contribution, and EY expects agentic AI to support end-to-end audit by 2028.
PwC 2026 Annual Corporate Directors Survey (via Bloomberg Law)
71% name AI as top skill to strengthen; 82% rate AI-risk/performance information fair, poor or unavailable
secondary: Poll of public-company directors run March to May 2026 (sample size not stated). Over two-thirds cite cybersecurity, data privacy and IP as their main AI worries; 58% say boards interacted more with the CISO.
Gartner
$2.67T in 2026, +49.5% YoY; AI infrastructure $1.48T; AI services $576B; AI software $462B; GenAI models $28B
Analyst forecast. Infrastructure is over half of spend, while spend on generative AI models themselves is about 1% of the total. Gartner says demand stays 'inelastic' to memory-related price rises.
NYU SPS Tisch Center, RateGain and HEDNA
More than half use or are procuring generative AI; fewer than 1 in 10 cut manual work by over 30%; 80%+ of commercial teams spend 1-2 days a week on manual reporting
Covers 270+ hotel brands and 58,000+ properties in 53 countries, survey period 1 Dec 2024 to 30 Nov 2025. RateGain is a hospitality software vendor and co-sponsor.
NYU SPS Tisch Center, RateGain and HEDNA
55% no or only minor change to distribution strategy; OTAs generate close to twice the bookings of hotel-owned digital channels
Same report. AI-originated search contributes a measurable share of reservations but no figure was given. Mid-sized chains show the most reporting automation and governance maturity.
Sony Bank / Fujitsu (Japan)
Development period -30%; man-hours -40%; impact assessment up to -90%; integration test execution up to -90%; code generation rate 99%
Company-reported in joint Sony Bank/Fujitsu release; measured vs conventional methods as of July 2026, live production work. No independent verification or method disclosed; 'up to' figures are best cases.
AWS / Strand Partners (Realizing Korea's AI Potential 2026)
58% adoption (from 48%); 81% report productivity gains; about 14 hours a week saved; about 60% report revenue growth
secondary: VentureSquare reporting an AWS-commissioned survey of 1,000 corporate leaders. Vendor-sponsored, self-reported savings.
AWS / Strand Partners (Realizing Korea's AI Potential 2026)
47% cannot measure AI ROI; 27% formal comprehensive strategy; 67% call AI a top priority
secondary: VentureSquare reporting an AWS-commissioned survey. Productivity claims in the same study are self-reported.
Deloitte and DocuSign
37% average time savings for legal teams; 29% cost savings from labour and outside counsel; agentic workflows nearly 30% higher ROI
Vendor-co-sponsored survey of over 1,100 leaders (DocuSign sells agreement software). Self-reported averages; page notes outcomes depend on data quality and human oversight. 61% still use manual processes to surface insights from stored contracts.
Houzz (2026 U.S. State of AI in Construction & Design Industry)
52% of firms use AI (up 20 points); average 4.7 hours saved per week; Houzz estimates $244,000 annual productivity benefit per firm
secondary: Builder magazine report of a Houzz survey; sample size and fieldwork dates not stated. The $244,000 is a Houzz estimate, not a measured figure; 95% of respondents report some concern about AI. Most use is sales, marketing and admin (estimates and proposals 87% of adopters).
Gartner (customer service practice)
27% would retry a chatbot after a bad experience; 7% used one in their last service interaction; 49% say they would have if offered
Gartner survey of 3,566 B2B and B2C customers fielded Feb-Mar 2026. Cited elsewhere in the release: 87% say access to a human agent is essential when GenAI is used. A quality-risk counterpoint for service-desk automation.
WPP
Headcount 97,388 at 30 Jun 2026 after about 11,000 cuts since start of 2025; up to 1,000 more reported
secondary: Reuters report carried by The Straits Times affiliate NST, citing the Financial Times; WPP did not comment and Reuters could not verify the new figure. Page links AI to the plan to merge Ogilvy, VML and AKQA into 'WPP Creative' but gives no AI-specific saving.
REVA Air Ambulance (via Ramp)
Invoice processing from 15-20 minutes to under 3; close accelerated about two weeks
Vendor-reported: Ramp blog quoting the controller. Books close by the 4th or 5th of the month versus nearly three weeks late before. Self-reported before-and-after comparison by the vendor; no independent audit or stated baseline method.
Snapdocs; Hospital Association of Oregon; MakeStickers (via Ramp)
Snapdocs reconciliation 5-6 hours to under 30 minutes; Oregon AP batch about 10 hours to minutes; MakeStickers 8-10 hours a week saved
Vendor-reported: Ramp blog compilation of customer quotes on its bill-pay product. Shows the scale of claimed gains in small finance teams, but all figures are self-reported with no methodology.
Morgan Stanley AI Governance Report 2026 (with ISS board data)
200 executives (revenue over $100M); 41% full AI model inventory; 71% require approved tools only; 68% of S&P 500 have an AI-skilled non-executive director
Survey fielded April 2026 plus ISS board data as of June 2026. 56% name data risk as the top AI risk; about 90% are defining who owns AI governance but only just under half say ownership is fully allocated. Only 24% of S&P 500 boards have three or more AI-skilled non-executives; Energy has none.
McKinsey & Company
37% attribute some EBIT impact to AI (flat); about 6% high performers; 14% saw an AI-linked workforce decline; 39% expect headcount cuts next year
McKinsey 'State of AI in 2026: On the road to ROI', 1,719 respondents in 97 nations, fielded May-June 2026. Self-reported; high performers are those attributing at least 5% of EBIT to AI, and nearly three-quarters of them have redesigned workflows.
McKinsey & Company
About 20% constrained by AI operating costs including tokens; 60% plan higher AI investment; 32% skipped a software purchase by building in-house
Same McKinsey survey (1,719 respondents, May-June 2026). 40% of $1B+ firms are scaling agents versus 22% of smaller firms, so scale effects favour large companies. Self-reported survey data.
IDC (Smith and Villars blog)
61% exceeded 2025 cloud AI budget (83% of those by under 15%); token prices down about 50% YoY yet bills not falling; 1 in 5 CEOs has a separate AI budget line
IDC blog citing an April 2026 survey (Middle East, Turkiye, Africa) and a May 2026 survey of 1,448 IT leaders, where difficulty budgeting token- and inference-based pricing was the top vendor-pricing barrier. Also repeats reports that Uber used its full-year AI budget by mid-April (secondary within IDC).
Gartner (finance practice)
Qualitative: early agent pilots 'more often fail from unclear controls than from weak technology'
Gartner analyst Q&A (Alex Levine), not a new survey. Recommends a contained, reversible first workflow, defined data access and human review, a sandbox, and judging pilots by governance readiness (traceability, failure log) rather than autonomy or ROI. Opinion and advice; no quantified outcome.
AWS / Strand Partners (Unlocking Malaysia's AI Potential 2026)
38% consistent use (from 27%); 67% of adopters on basic tools; 19% have a formal scale-up strategy
secondary: TechNode Global reporting an AWS-commissioned study. Sample size not stated. Financial services 53%, manufacturing 50%.
AWS / Strand Partners (Unlocking Malaysia's AI Potential 2026)
30% clear accountability; 27% regular monitoring or audits; 18% documented escalation; financial services 39% formal governance vs 24% overall
secondary: TechNode Global reporting an AWS-commissioned study. 47% want practical frameworks or templates.
RICS (AI in commercial property and construction report 2026)
Construction: about two-thirds use AI, 39% piloting, 19% regular use (up from 12%), 4% widespread; 3,148 respondents across two monitors
RICS Q1 2026 Global Construction Monitor (1,883) and Commercial Property Monitor (1,265). Top construction barrier is lack of skilled personnel; privacy and security concern rose from 22% to 30%. The report gives no ROI figures; the UK productivity question drew mostly 'not sure'.
RICS (AI in commercial property and construction report 2026)
Almost two-thirds not prepared or minimally prepared; 43% not aligned with RICS AI standard; 26% unaware of it; 37% plan to raise AI investment
Construction monitor respondents (1,883). Advanced readiness is described as extremely rare. The report is internally inconsistent on the share with no investment plans (about a quarter in one place, about a fifth in another).
Oreopoulos Liut Sungu Low (EdWorkingPaper 26-1552)
6,997 grade 6-8 students in 20 Hamilton County schools, practice week 23-27 Mar 2026: no significant effect on the combined delayed score; in the mastery arm AI students scored 40.2% vs 37.0% on a practised exercise (+3.1 points, p about 0.065, only suggestive)
Measured randomised trial, primary PDF opened. One-week practice and one-week delayed test only.
Insurance Journal / Gallagher study
AI-related lawsuits +978% 2021-2025 and +137% in 2024-2025 (Gallagher); patent 11.9%, copyright 11.2%, privacy 10.2% of cases
secondary: Trade-press feature quoting a Lathrop GPM attorney that an 'industry-wide reaction' is under way and rapid adoption of AI exclusions across most lines is expected. Verisk could not say how many insurers have adopted the ISO forms. Berkley's absolute exclusion covers D&O, E&O and fiduciary products. Claims not excluded may still be covered by 'silent' AI cover.
Tracxn data via Crowdfund Insider
2026 YTD $4.1B across 23 disclosed equity rounds; Kling AI about 68% of total; historical Vietnam AI funding $19M
Regional benchmark as of July 2026: funding is dominated by Singapore-linked late-stage rounds; Vietnam, Malaysia, Indonesia and Thailand are small.
Gartner (supply chain practice)
67% of supply chain digital investment goes to AI; over half of CSCOs unsure what return they get
secondary: ESM Magazine report on two Gartner surveys (394 professionals at firms above $250M revenue, Nov 2025-Feb 2026; 135 senior leaders, Jan-Apr 2026). The primary Gartner release was not located.
CNH Industrial (Farmer Pulse Report)
62% see moderate or extremely high value; 74% adopt to cut input costs; 56% cite cost and 36% ROI uncertainty as barriers; 217 North American farmers
Vendor-sponsored survey of 217 US and Canadian farmers and ranchers, May 2026. Covers precision technology (89% use auto-guidance), not AI specifically; included as the agriculture adoption baseline. Perceived value, not measured returns.
Ivey Business School (Birkinshaw Raz Walters)
14 tools reviewed (7 public edtech, 7 faculty pilots); cites a 2026 UK study with 95% of students using GenAI
secondary: commentary article with qualitative review, no outcome metrics; cited UK study not identified.
Bank of England staff (reported by Staffing Industry Analysts citing Bloomberg)
Vacancies in most AI-exposed occupations -15% over three years (medium exposure -10%, low -6%; customer service and admin more than -20%); software and IT consulting contribution to annual productivity growth up tenfold vs pre-Covid decade to 0.1 percentage points in 2023-25
secondary: trade article on two BoE staff blog posts; the posts were not opened and the article names no dataset (not confirmed to be the Decision Maker Panel). Observational association, not causal. Governor Bailey expects a J-curve.
AWS / Strand Partners (Unlocking Indonesia's AI Potential 2026)
40% adoption; 75% of adopters report productivity gains; 62% attribute revenue growth to AI; SMEs 38%
Vendor-sponsored study of 1,000 business leaders and 1,000 members of the public, published on an AWS page. Financial services 58% adoption.
AWS / Strand Partners (Unlocking Indonesia's AI Potential 2026)
56% of adopters still exploring; 12% fully integrated; 24% formal strategy; 21% data governance; 56% cite skills shortage
Vendor-sponsored survey of 1,000 business leaders. Among SME adopters only 10% are fully integrated.
hibank (Indonesia)
End-of-day processing time down 30%; infrastructure usage down 25%
Vendor-reported on an AWS page (Kiro is an AWS product); no independent verification or baseline given.
Rey (Indonesia healthtech)
Claim decisions from 1-2 weeks to under 1 day; over 7,000 claims a day; over 500,000 insured members
Vendor-reported on an AWS page; baseline and methodology not published.
Shopify
AI-driven traffic and orders about 3x YoY; AI channels bring new buyers at nearly 2x the rate of other channels
secondary: Retail TouchPoints on Shopify's Q2 2026 earnings call. Company-reported and not independently verified.
Shopify
Nearly 34m Sidekick conversations in Q2; over 36,000 custom apps (12,000 in Q1); 8% more new merchants reach five orders in 15 days
secondary: Retail TouchPoints on Shopify's Q2 2026 earnings call. Company-reported usage metrics, not merchant profit outcomes.
TCS, Infosys, HCLTech (via Wright Research)
Q1 FY27 YoY: TCS +2.7%, Infosys +2.8%, HCLTech +2.6%; Infosys AI at 8.2% of revenue; HCLTech 25-30% more effort for same revenue
secondary: Analyst blog summarising company results and Nasscom FY26 estimates (sector about $315B, +6.1%). Reports TCS passing 10-15% of productivity savings to clients at signing, analysts modelling 2-3% annual application-services deflation, and fresher hiring falling from about 600,000 in FY22 to 120,000 in FY25 (Xpheno). Several figures are blog-reported, not from filings.
Ramp AI Index (FRED series)
56.13% in Aug 2026 vs 44.98% in Aug 2025
Share of Ramp customer businesses with paid AI subscriptions. Ramp customers skew tech-forward and the index counts paid adoption only, so it runs well above Census-style estimates. Series last updated 29 Sep 2026.
Ramp AI Index (FRED series)
Anthropic 43.79% vs 16.20% a year earlier; OpenAI 39.76% vs 40.97%
Share of Ramp businesses paying each provider. Anthropic overtook the Feb 2026 level of 27.5% within six months; OpenAI drifted down from 41.2% in Feb 2026. Evidence on vendor concentration and switching, not on outcomes. Ramp customer base is not representative of all firms.
Ramp AI Index (FRED series)
Large businesses 59.04% (44.04% a year earlier); finance and insurance 73.71% (60.10%)
Paid-adoption shares for Ramp customers by segment, August 2026 vs August 2025. Large firms have caught up with the overall average (56%) and finance and insurance sits well above it.
Ramp AI Index (FRED series)
Information sector 80.92% in Aug 2026 vs 72.42% in Aug 2025
Highest-adoption sector in the Ramp series, up 8.5 points in a year. Ramp customers skew tech-forward and the index counts paid subscriptions only, so it overstates adoption in the wider economy.
Stanford Digital Economy Lab (Brynjolfsson Chandar Chen)
Employment of ages 22-25 in AI-exposed occupations 19% below counterfactual as of June 2026, vs 15% in July 2025 vintage; no widespread economy-wide displacement
Descriptive, not causal per the authors; ADP payroll data through June 2026. Estimates attenuate when controlling for education.
Stanford Digital Economy Lab (Brynjolfsson Chandar Chen)
Levels: ages 22-25 in two most-exposed quintiles fell about 11% Nov 2022 to Jun 2026; three least-exposed quintiles grew about 10%
Gap stems mainly from reduced hiring rather than separations; adjustment shows in employment not base pay.
Forrester (reported by People Matters)
55% of employers regret laying off workers for AI; Forrester expects half of AI-attributed layoffs to be quietly reversed (Predictions 2026)
secondary: reported by People Matters and The Next Web (2026-08-23); sample size not reported by either.
Gartner (reported by People Matters)
Gartner projects half of companies that cut customer service headcount due to AI will need to rehire for similar functions by 2027
secondary: reported by People Matters 2026-07-31; I did not open the Gartner release so its date, wording and sample are unconfirmed; forecast not observed.
Federal Reserve Bank of St. Louis (Ozkan Kalyani Sullivan)
About 490,000 S&P Global earnings-call transcripts for 5,198 US listed firms 2000-2025: AI reached roughly 15% of productivity sentences by end-2025; 95% of AI productivity sentences refer to future gains vs about three-quarters of non-AI ones; 95% say productivity is rising
Measured text analysis (LLM classifier on sentences) of what executives say, not of realised productivity; primary page opened; no firms named. Useful as evidence that call rhetoric runs ahead of reported results.
UPS (Q2 2026 earnings call, CEO Carol Tome)
Cost per piece about 28% lower in automated buildings; 68.5% of US volume through automated buildings (64% a year earlier); 50 million labour hours eliminated across 2025-2026
secondary: Supply Chain Dive report of the earnings call; self-reported by UPS. Automation here is largely robotics and sortation rather than AI specifically. About 78,000 operational positions tied to Amazon volume are being eliminated; UPS expects to close nearly 150 buildings.
Verisk ISO endorsements (as reported by Shumaker, Loop & Kendrick, National Law Review)
Forms CG 40 47, CG 40 48, CG 35 08; reported effective January 2026
secondary: Law-firm alert says ISO generative-AI exclusion forms reached the market in January 2026, and that broad 'arising out of' wording may let insurers deny claims where AI is only loosely connected. Berkley also has an 'absolute' AI exclusion in specialty lines. The article names no denied claims and gives no dollar figures.
DBS Bank (Singapore)
digibot: 9 in 10 queries resolved without follow-up call (H1 2026); DBS Joy: calls/emails down 7%, satisfaction up 17%, 10M+ users
Self-reported in DBS press release (republished by The Asian Banker). No handling-time or cost saving stated; the 7% is volume of calls/emails, not time.
UPS (Q2 2026 results)
Automation removed $4.5bn of related expense; network reconfiguration to deliver $3bn benefit by end 2026; US Domestic operating profit $16m vs $916m; net income $604m vs $1.28bn
secondary: Transport Topics report of UPS results. Savings are company-reported. Shows efficiency gains coinciding with sharply lower profit as about 2 million lower-quality Amazon packages a day were deliberately shed.
Lemonade
LAE ratio 5% vs ~9% industry; gross loss ratio 60%; in-force premium $1.43B (+32.5%)
secondary: MarketBeat earnings summary. Management attributes the LAE ratio to Lemonade OS driving AI across claims. Claims automation rate was not quantified in this summary. Gross loss ratio included 7 points of favourable prior-period development, so do not read it as AI-only effect. AI-native insurer, not directly comparable to incumbents.
Gallup
65% positive productivity effect; 45% for 1-2 work purposes, 66% for 3-4, 78% for 5-6, 90% for 7 or more
Gallup survey of employees in organisations that have implemented AI (May 2026 data); self-reported effects.
Gallup
36% strongly agree manager supports team AI use; those who do are 9.3x as likely to say AI has transformed how work gets done; leaders frequent use 51% vs 17% in 2023, ICs 26% vs 9%
Gallup (page last updated 2026-09-30). Multipliers are associations from self-reported survey data.
Politecnico di Milano (JAMA Network Open)
903 FDA-authorised AI devices analysed; recall risk higher where clinical evidence undocumented
University news item; no numeric effect size on the page. Misuse also frequently linked to recalls. Authors call for integrated post-market surveillance across EU and non-EU adverse-event sources.
Deloitte (governance professionals survey, via CIO Dive)
51% of boards lack AI rules/guidance; 47% do not facilitate AI use; 8% use approved AI tools for committee work
secondary: Trade press summary of a Deloitte survey of corporate secretaries and in-house counsel (sample size not stated). 77% of boards had director AI briefings in the past six months, so education is ahead of policy.
Telefonica Deutschland (via Telecoms.com)
Up to 1,100 staff (about 14%); EUR 185M annual savings from 2028; Q1 revenue -8.6%
secondary: Trade press. Restructuring charges of EUR 265M in Q2 and up to EUR 155M in H2; about 60 of 800 shops close. Company says it is 'intensifying the use of AI in key processes' in sales and customer service, but revenue was already falling 1.4% (2024) and 3.8% (2025), so AI is one driver among several.
HKT (Hong Kong Business AI Adoption Survey)
67% implemented, piloting or planning; 82% have 12-month investment plan; 14% of adopters use agentic AI; SMEs 19% plan agentic vs nearly 40% of large firms
Telecom-vendor survey of nearly 300 IT leaders, Q1 2026; no independent research firm named. Hong Kong, not mainland China.
US FDA list (reported by AI2Work)
1,524 authorisations to 30 Mar 2026; radiology 76%; 6 devices (1.6%) cite randomised data
secondary: AI2Work editorial summarising the FDA list, Imaging Wire and two JAMA-family studies; it gives no primary citations and says figures were not independently verified. Use as directional until checked against the FDA list.
Bank of America (reported by CIO Dive)
200,000+ employees use AI-enabled tools; 400,000+ prompts a day; 300+ approved AI use cases, 114 of them generative; 34 fully implemented; CFO says manual work reduced, no savings figure
secondary: CIO Dive summary of Q2 2026 earnings-call remarks (calls believed 14-15 July; the article, dated 20 July, gives only weekdays). Transcript not opened. Company-reported usage counts, not cost or headcount results.
JPMorgan Chase (reported by CIO Dive)
Nearly 1,000 live AI use cases across risk, fraud, marketing and document reading; Dimon says AI is expensive, he does not expect it to lift margins soon and customers will be the main beneficiaries ('You don't uniquely benefit from AI')
secondary: CIO Dive summary of earnings-call remarks; transcript not opened. Company-reported; no savings figure.
Citigroup; BNY; Wells Fargo (reported by CIO Dive)
Citi: nearly 9 in 10 employees use its AI tools; BNY and Wells Fargo gave no quantified results (Wells launched an 'AI Teammate' for advisers); across the five banks the article reports no headcount reduction, cost saving or productivity percentage
secondary: CIO Dive summary; transcripts not opened. Shows that even in 2026 bank disclosures are usage counts, not P&L effect.
UK Office for National Statistics (Business Insights and Conditions Survey, wave 159)
About 12% (Sept 2023) to about 35% (June 2026) of businesses with 10+ employees; 28% of 0-9 employee firms vs 49% of 250+
Official statistics, fieldwork 5 to 28 June 2026. Information and communication near 60%, construction 13%. Large language models used by 18% of businesses.
UK Office for National Statistics (Business Insights and Conditions Survey, wave 159)
Average AI technologies per adopter 1.4 to 1.6; about 6% of operations-focused adopters report headcount decrease; 41% of 10+ firms report no barriers
Extensive use is reported by 17% of the smallest firms and 9% of the largest. Lack of expertise most evident at 100 to 249 employees (about 18%).
FPT Corporation (Vietnam, via Vietstock)
Group revenue VND 26T (+12.6%); pre-tax profit VND 5.7T (+18.1%); overseas digital transformation revenue +23.8% to VND 9.6T
secondary: Vietnamese business press on H1 results. FPT says AI Factory clusters in Vietnam and Japan exceeded 90% utilisation and turned profitable in Q2 2026 (company-reported) and that it signed 14 overseas IT projects above $10M each.
UnitedHealth Group (via VectorShift call summary)
96% first-pass approval; 70% of employed Optum Health providers on ambient AI (target 90%+); targets 80% of prior auths real-time by end-2027
secondary: third-party call summary; one prior-auth volume figure (~69,000 administrative hours) is garbled in the transcript and should be verified. Management also said operating cost ratio rose partly due to AI investment and that clinicians stay involved in any non-approval. Claimed 90% reduction in clinician burnout is unsubstantiated.
AWS / Strand Partners (Unlocking Thailand's AI Potential 2026)
43% adoption (from 32%); 74% basic, 17% intermediate, 9% advanced; 84% report productivity gains; 71% report revenue gains averaging 19%
secondary: The Nation Thailand reporting an AWS-sponsored survey of 1,000 business leaders. Vendor-sponsored, self-reported.
AWS / Strand Partners (Unlocking Thailand's AI Potential 2026)
19% enforced AI policies, 47% informal, 34% none; 21% formal AI strategy; 51% cite skills shortage
secondary: The Nation Thailand reporting an AWS-sponsored survey. 24% have heard of agentic AI and 19% feel ready to adopt.
KASIKORN Business-Technology Group (Thailand)
Engineering effort cut 2-3x on legacy COBOL modernisation
secondary: The Nation Thailand reporting an AWS-sponsored study. Vendor-reported; baseline not stated.
Osotspa (Thailand)
About 1,450 queries a month across finance, legal and marketing
secondary: The Nation Thailand reporting an AWS-sponsored study. Usage volume only; no cost or time saving given.
US Senate (reported by Becker's Payer)
Letters dated 14 Jul 2026; OIG found 12% of skilled-nursing prior-auth requests denied across 19 MA insurers
secondary: letters ask each insurer for an inventory of predictive technologies and whether AI can make final denials. Cites June 2026 OIG reports on June 2024 prior-auth data. Response deadline 28 July; outcome not captured here.
UnitedHealth Group (reported by Motley Fool)
Close to $1B expected cost reductions; denials from missing info -68%; appeals -88%; call volume -25%
secondary: Motley Fool summary of company-reported early results from 1,000+ AI use cases and a ~$1.5B budget. Figures are management claims; the article does not give measurement bases. Pairs with the senators' scrutiny of the same sector.
Thryv (reported by Chain Store Age)
561 US SMB decision-makers surveyed April 2026: AI adoption 66% (55% a year earlier); 70% say AI increased revenue, 55% say it cut costs; 92% of users say it saves time; 53% spend at least $100 a month
secondary and vendor-run: Thryv sells an AI marketing platform; respondents were from a third-party panel, mostly $1M-1.9M revenue firms. Self-reported with no baseline; compare Census BTOS (about 20%) and DSIT (16%) for how definitions and samples change the headline.
Thryv (reported by Chain Store Age)
70% say they need more training to use AI effectively; 57% rely mainly on YouTube and social media to learn; 46% would pick AI software over hiring for equally capable work (38% in 2025)
secondary and vendor-run survey, self-reported (n=561 US SMBs, April 2026).
HMRC (Phase 3 Copilot evaluation)
3,000 randomly allocated licences, Sep-Dec 2024: self-reported saving of 2-3% of the week (about 60 minutes) after cutting about 20% for non-users and response bias; 83% used the licence; satisfaction 7.1/10; 1,364 survey responses (45%)
Primary gov.uk page opened. Time saving is self-reported; Viva usage data and task exercise gave no time results on the page. The up-to-£50M analyst estimate is unclear on timing and is not used as a metric.
Allianz Partners
1,500 to 1,800 jobs cut over 12 to 18 months, mostly call centres
secondary: Insurance Business reporting CEO Tomas Kunzmann's statement at a Munich event. Cuts via severance, early retirement and voluntary leave after six months of works-council talks. Percentage-of-division figures in other press are unsourced.
Japan Ministry of Internal Affairs and Communications (2026 White Paper on Information and Communications)
86.4% Japan vs 90.9% US, 91.6% Germany, 98.1% China; 68.9% of Japanese firms plan use, up from 49.7%
Online survey Jan to Feb 2026: 515 Japanese companies and 309 in each comparison country. Use is for 'some work', a low bar for depth.
Japan Ministry of Internal Affairs and Communications (2026 White Paper on Information and Communications)
Nearly 30% no organisational effort, higher than US, Germany and China; Japan lags on skills environment
Same Jan to Feb 2026 survey (515 Japanese firms). Indicates usage without redesign is common.
Fang Yuan Zhang Donati Sarvary (arXiv; MSI)
Pre-sale service chatbot: sales +16.3% and conversion +21.7% vs control (44,614 consumers); part of seven randomised field experiments Sep 2023-Jun 2024 on a cross-border e-commerce platform of mostly small sellers
Measured randomised field experiments, primary arXiv page opened; platform not named. Top-line sales only: labour cost, ROI and profit not measured; short horizon; workflows chosen by managers.
Fang Yuan Zhang Donati Sarvary (arXiv; MSI)
Product descriptions sales +2.1% (4.77M consumers); search query refinement +2.9%; push messages +1.6% (not significant, ~40% exposed); Google ad titles -4.5% (not significant, un-tuned model); summed gain about $4.63-5.18 per consumer a year, 5.5-6.2% of per-user revenue growth
Measured randomised field experiments (primary page); the aggregate assumes linear additivity and covers top-line sales only, with no cost or ROI. Returns fell 3.9% for descriptions and 11.4% for push messages.
Deloitte (AI in manufacturing)
84% report measurable value; about 20% average KPI improvement potential; about 20% of use cases scaled across sites; 140+ manufacturers
Deloitte survey of more than 140 manufacturing organisations. Adoption by area: quality 62%, production 57%, logistics/supply chain 49%. Technology use: ML/deep learning 42%, GenAI and agentic 40%, physical AI 18%. The 20% KPI figure is potential, not realised, improvement.
Deloitte (AI in manufacturing)
Barriers: implementation cost 43%, technical expertise 35%, resistance to change 35%, regulation 34%, data quality 30%. Risks: operational disruption 79%, cybersecurity 51%
Same Deloitte survey of 140+ manufacturers. System or model reliability is cited as a risk by only 15%, well behind operational disruption.
Concentrix, Teleperformance (via LetsDataScience)
Concentrix -21% premarket; Teleperformance -11% to -13% in sympathy
secondary: Aggregator note on 29-30 June 2026 trading. Investors framed it as conversational AI displacing human support, some calling the sub-sector 'uninvestable'; Concentrix itself pointed to offshoring and spend pullbacks. Notes Teleperformance lost close to a third of its value in one session in early 2024 after Klarna's chatbot claim.
Concentrix
Q2 revenue $2,462.5M (+0.6% constant currency); operating margin 3.9% vs 6.1%; FY26 guide cut to $9.925B-$10.025B
Q2 FY26 results exhibit (8-K). Company headline claims iX Suite deals up 400% YoY but discloses no AI revenue. Operating income fell to $95.4M from $148.3M (per press summary). Management blamed faster client offshoring and selective spend pullbacks.
Ford Motor Company (reported by Bloomberg, summarised by AI Weekly)
About 350 inspectors rehired; AI vision system reportedly ran 150 million inspections at 17 North American sites and flagged 400,000 issues
secondary: AI Weekly summary of a Bloomberg report; the primary Bloomberg article was not opened. Reported gaps: AI did not preserve institutional expertise or train junior staff. Ford is also expanding AI inspection (sixfold at Kentucky Truck Plant per WardsAuto); defect types missed and total rehiring were not specified.
Cushman & Wakefield
Q1 operating and administrative expenses about $337m, up $30.9m (10%); about 600 people with enterprise Claude access (anonymous broker estimate)
secondary: Bisnow. Company attributed cost growth to salaries and technology; the AI share is not separated. Company added token limits and a business-case process.
JLL
About 25% of employees daily users; about 2 hours a week saved (user estimate); 100,000+ users and 35m+ prompts on its generative AI model
secondary: Bisnow. Time savings are user estimates, not measured. JLL Spark has invested over $445m in nearly 60 startups since 2017.
CBRE
Projected 25% research cost reduction; nearly $1.7bn on hardware and software in 2025, up $300m
secondary: Bisnow quoting CEO Robert Sulentic. A projection, not a realised result.
Walker & Dunlop
About 10 minutes versus roughly a week for a manual analyst; several results missed
secondary: Bisnow, describing an internal test. No dollar spend disclosed.
JLL
92% piloting AI; 5% achieved most programme goals
secondary: reported in Bisnow as industry context; sample size and fieldwork dates not stated in the article.
Hawaii Pacific Health Medical Group (JMIR Medical Informatics)
High users: time in notes per appointment 7.14 to 5.63 min (-21%); burnout 59.5% to 50.6% (p=0.34, not significant)
Prospective 3-month quality-improvement pilot, Dec 2024 to Feb 2025, of the Abridge scribe integrated with Epic: 79 of 80 providers completed, 25,600+ notes, 23 specialties; 57% were high users. Pajama time -13% (not significant). Self-audit of 494 notes: 86.6% rated accurate; about 84% needed edits to under a quarter of content. Authors note self-reported data and a non-validated burnout measure.
Schneider Electric / World Economic Forum Global Lighthouse Network
On-time delivery 61% to 97%; lead times down up to 35%; $43m backorders eliminated
Named as a Customer Centricity Lighthouse in the June 2026 WEF release. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Saudi Aramco / World Economic Forum Global Lighthouse Network
Production volumes +26%; product quality +43%; OEE +44%; CO2 emissions -16%
More than 50 use cases at the Hawiyah NGL plant. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Rockwell Automation / World Economic Forum Global Lighthouse Network
Units per person-hour +43%; defects -35%; time-to-competency -67%; 1,000+ SKUs and 20,000+ changeovers a year
More than 50 solutions deployed. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
DCM Shriram / World Economic Forum Global Lighthouse Network
EBITDA +11 percentage points; power costs -32%; material costs -15%; CO2 -14%
45 solutions deployed at the Gujarat site. One of the few Lighthouse results stated in financial terms (EBITDA). Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
CIMC Reefer Containers / World Economic Forum Global Lighthouse Network
Manufacturing lead times -32%; defects -47%; conversion costs -24%
More than 50 digital and AI applications at the Jiaozhou site. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Hitachi Vantara Manufacturing / World Economic Forum Global Lighthouse Network
Inventory -50%; order-to-ship lead time -77%
US plant (Norman) listed by the WEF in the Productivity category. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Unilever / World Economic Forum Global Lighthouse Network
Response times -72%; SKUs +50%; minimum order quantities -40%; service levels 99%
Supply Chain Resilience category, Haridwar, India. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
RRS Qingdao Smart Logistics Park / World Economic Forum Global Lighthouse Network
Transportation costs -23%; inventory turnover +40%; response time +48%
Supply Chain Resilience category. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Fujian Ningde Nuclear Power / World Economic Forum Global Lighthouse Network
Human error -71%; profit per employee +50%; digital talent more than 10x; 45 use cases
Talent category, Fuding, China. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
China Merchants Heavy Industry / World Economic Forum Global Lighthouse Network
Throughput 2.6x; dock time -52%; first-pass assembly yield +25%
More than 25 digital solutions at the Haimen site. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Sidley Austin (law-firm blog) on EU AI Omnibus
Parliament vote 16 Jun 2026; adoption targeted before 2 Aug 2026; SME relief extended to small mid-caps
secondary: Law-firm note says Parliament adopted the AI Omnibus on 16 June 2026 with Council approval outstanding and both institutions aiming to finish before 2 August 2026. The AI Office would supervise systems built on a provider's own general-purpose models; national authorities keep sector competence (e.g. financial institutions). Provisional-deal date is given as 7 May here versus 6 May in the Gibson Dunn alert.
Accenture (via 01net.it press-release reprint)
Q3 revenue $18.7B (+3% local currency); bookings $19.3B vs $19.7B; headcount about 799,000
secondary: Reprint of the quarterly press release for the quarter ended 31 May 2026; no AI-specific metrics in the text. Headcount rose to about 814,000 by year-end.
Evident Insights (reported by Insurance Business)
65 disclosed use cases: productivity 75%, revenue uplift 2%; 49% single-activity tools; 3 of 30 insurers disclose enterprise ROI
secondary: Insurance Business on the second annual index of 30 large North American and European insurers. Ranking: Allianz, AXA, Manulife, Zurich (up from 12th), Liberty Mutual. Claims is 28% of use cases, underwriting and pricing 17%; only 8% show improved decision quality across end-to-end workflows.
KCCI Economic Research Institute
52.7% SME vs 66.5% large-firm workers; gap narrows to about 4 points after controlling for support and skills
secondary: SBS Star reporting. Survey of 3,000 workers. Active internal encouragement lifted use by 15.5 points; subscription subsidies by 8.1 points.
KPMG
Technology sector n=100 ($1B+ US, fielded 4 May-2 Jun 2026): barriers to agent deployment data readiness and access 75%, human oversight skills 67%, system complexity 31%, workforce resistance 30%; employee adoption of agents 68%
Self-reported; primary PDF opened but chart labels were ambiguous so deployment-stage percentages and the spend figure ($294M vs $269M conflict) were not used.
IBM Cost of a Data Breach 2025 (as reported by Matters.AI)
$4.63M average shadow-AI breach, $670k above standard; 20% of breaches; 97% lacked AI access controls
secondary: Vendor blog restating IBM's 2025 report (3,470 interviews across 600 breached organisations): shadow AI featured in 20% of breaches, took 247 days to identify, exposed customer PII in 65% of cases, and 63% of organisations lacked an AI governance policy. IBM's own page was not confirmed to carry these figures; check against the report before quoting.
Gartner (finance practice)
63% say 2025 AI implementation was slower than expected; 66% of adopters cite efficiency as a top benefit
Gartner survey of 204 finance leaders fielded March 2026, released at its Finance Symposium. Analyst Marco Steecker warns CFOs not to mistake pilot and tool counts for value and says too many initiatives target incremental gains rather than material business problems. Analyst-firm survey; the realised-versus-expected value chart has no readable values on the page.
Salesforce
220,000 leads worked autonomously; $42M pipeline in Q1 FY27; SDR agent Piper delivers 45% more pipeline than traditional web agents
Vendor self-reported on its Q1 FY27 earnings call (27 May 2026). Piper (from the Qualified acquisition) engages about 50% of salesforce.com traffic. Pipeline is not closed revenue; no conversion or cost figures were disclosed. Salesforce sells the product, so this is a showcase of its own 'customer zero' use.
Salesforce
Over 4 million inquiries handled autonomously in about 15 months; 'now double what human agents are handling'
Vendor self-reported by Marc Benioff on the 27 May 2026 earnings call, covering help.salesforce.com and the 1-800-NO-SOFTWARE line. No resolution quality, escalation or cost-per-case data given.
Salesforce
Engineering about 15,000, roughly flat for two years; Slackbot 3.8M annualised hours saved and about 3% company productivity; code shipped doubled year on year
Vendor self-reported on the Q1 FY27 earnings call: Benioff attributes flat engineering headcount to AI efficiency while hiring concentrates in sales. No dollar savings for internal Agentforce use were disclosed. Illustrates headcount avoidance rather than layoffs.
Gibson Dunn (law-firm alert) on EU Digital Omnibus on AI
Annex III high-risk date moves from 2 Aug 2026 to 2 Dec 2027; Annex I from 2 Aug 2027 to 2 Aug 2028
secondary: Law-firm client alert on the provisional political agreement (reached 6 May 2026, confirmed by Council 13 May; formal adoption pending at publication). Standalone high-risk systems (e.g. HR, credit) get roughly 16 more months; AI embedded in regulated products one more year. Commission tabled the omnibus on 19 Nov 2025.
Gibson Dunn (law-firm alert) on EU Digital Omnibus on AI
Art. 50 applies 2 Aug 2026; watermarking grace to 2 Dec 2026 for systems already on market; AI Office gets inspection and fining powers
secondary: Same alert. Transparency duties are not delayed, but providers already on the market get four extra months for the Art. 50(2) marking obligation. The AI Office gains investigation, on-site inspection, binding-commitment and fining powers (the alert gives no fine amounts). Sandbox deadline moves one year to 2 Aug 2027.
FPT Software (company release)
$256M five-year deal with an Asian energy group; contracts above $10M doubled in 2025; 54,000+ staff; overseas IT services +13.7% in 2025
Vendor-reported figures, not independently verified. Also claims about 30% internal productivity gains in selected activities and an AI Factory that processed over 1,100 billion tokens at about 70% utilisation. Useful as a Vietnam-based offshore vendor's self-description, not as outcome evidence.
US Census Bureau (reported by Princeton Info)
Nearly 20% of US businesses used AI as of early May 2026 (range 17-20% from 14 Dec 2025 to 3 May 2026); 20-23% expect to within six months; 250+ employees 37%, 100-249 employees 32%; use rose at firms with 20+ employees but not significantly below 20; a later Fed analysis put 250+ at 38.5% and 1-4 employees at 19.8% (14 June)
secondary: local-news write-up of a Census report dated 26 May by Grundy, Breaux and Khatiwoda; the Census report itself was not opened. Article dated 2026-09-02. Measured official survey but Nov 2025 question widened to 'any business function', so earlier figures are not comparable.
US Census Bureau (reported by Princeton Info)
Retail about 14% using AI and about 17% expecting to within six months; information 39.7% and finance and insurance 33.9% (as of 3 May 2026)
secondary (same article and caveats as the national BTOS row); underlying Census report not opened. Measured official survey of use, not of results.
HDBank (Vietnam)
Headline claim: 70% processing time saved; 6 processes automated
Vendor-reported customer story on FPT IS site (akaBot is an FPT IS product). Body gives no baseline or accuracy number. RPA rather than generative AI. Page date is a last-updated stamp.
Techcombank (Vietnam)
Agent build time ~8 months to <1 week; BA turnaround ~3 days vs 10; conservative productivity uplift 15-20%; ~1,000 hours saved (~5.6 virtual FTE)
Self-reported by Techcombank CTO Tran Hoang Quan to The Asian Banker. Business-analyst agent accuracy ~70% and 20-30% of its work still needs human input. Quality-engineering saving of 20% is a target, not an achieved result. Article notes cost-to-income fell 32.7% to 30.8% in 2025 but does not attribute it to AI.
Gartner
36% of 101 chief procurement officers are 'very confident' they can redesign roles and processes for AI (fielded Jan-Feb 2026); GenAI gains for individuals shrink at team-output level, which Gartner calls the 'AI productivity paradox'
Self-reported analyst survey, primary Gartner release opened. No size of the individual-to-team drop and no adoption rate are given.
Rockwell Automation / Sapio Research (2026 State of Smart Manufacturing)
34% of operations AI-augmented; 18% in pilot mode; 43% of collected data used effectively; 1,560 respondents in 17 countries
Vendor-sponsored survey (Rockwell, fielded by Sapio Research); 11th annual edition. Manufacturers expect more than half of operations to be AI-supported by 2030. The release gives no quantified ROI or workforce figures, only that outcomes are measurable.
Zurich Insurance / Cytora (reported by Reinsurance News)
Triage time -80% (75 to 15 min); digitisation accuracy 70-80% to 98%; straight-through intake 10% to 95%
secondary: Reinsurance News on a 90-day rollout across five countries, with expansion to 20+ markets planned over 16 months. Figures are company/vendor-reported with no baseline period or method stated. The 'up to 60 minutes' time-to-quote target had no reported result.
Six US health systems, JAMA Network Open (reported by Textbook of Digital Health)
Burnout 51.9% to 38.8% after 30 days; 263 ambulatory clinicians
secondary: news post summarising a quality-improvement study of 263 ambulatory clinicians across six health systems. Cognitive task load and after-hours documentation also fell. Short follow-up and self-reported burnout. Page date is the news post date, journal date not stated.
Kasikornbank (Thailand)
Covers 22 divisions
Enterprise-wide generative AI programme pairing secure tools with training, governance and dedicated support staff; won Best AI Talent and Workforce Enablement Initiative in Asia Pacific (The Asian Banker). No quantified results stated.
SoftBank Group
Net profit 5,002.2 billion yen; investment profit 7,286.5 billion yen, of which 6,730.4 billion yen from OpenAI
secondary: Nippon.com report. Profit is mark-to-market gains on an AI investment, not operating returns from AI use; contrast for executives reading AI headlines.
Aesthetify GmbH / Verbraucherzentrale NRW (OLG Hamm)
Injunction with fines up to EUR 250,000 per breach; judgment of OLG Hamm, case 4 UKl 3/25
A cosmetic-treatment business's website chatbot told visitors its two doctors held specialist titles they did not have. The Higher Regional Court in Hamm treated the bot's statements as the company's own misleading commercial acts under the Unfair Competition Act. Source is a compliance register summary (CompanyScope, 27 Sep 2026) of the judgment; judgment not final and appeal to the Federal Court of Justice was allowed. Only sum fixed is EUR 260 plus costs.
Gartner (marketing practice)
15.3% of marketing budgets to AI; 30% report mature AI readiness; 70% say processes not mature enough to scale
Gartner 2026 CMO Spend Survey of 401 marketing leaders in North America, UK and Europe (mostly firms above $1B revenue), fielded Jan-Mar 2026. 56% say budgets cannot fund their 2026 strategy. Analyst-firm survey; large-company skew.
Gartner (marketing practice)
AI-ready organisations: 21.3% of marketing budget on AI and 8.9% of revenue on marketing, versus 7.8% of revenue average
Same 401-leader Gartner CMO survey (Jan-Mar 2026). Benchmark of spend shape for AI-mature marketing teams; correlation, not proof that higher AI spend causes better results.
Heidi Health (reported by Healthcare Management UK)
Vendor-reported: over 4 million hours freed; 86% cut in paperwork; 95% report reduced burnout
secondary: vendor-reported figures from Heidi's own Impact Report as covered in a vendor-led article; methodology, sample sizes, comparison groups and time periods not described. Site examples include a 2,700 to under 200 clinic-letter backlog cut at one trust (est. 139,000 GBP saving) and a 5.10 GBP return per 1 GBP in one frailty team. Treat as unverified marketing evidence.
Expedia Group
250m+ service interactions a year; over half resolved by self-service; over 30% AI-powered
Company-reported by CEO Ariane Gorin on the Q1 2026 earnings call. secondary: CX Dive, 12 May 2026; exact call day not stated (date given as year-month). No cost or satisfaction outcome given.
Intuit QuickBooks (AI Impact Report, Australia)
Regular use 40% (July 2024) to 69% (Jan 2026); daily use 9% to 28%; 79% of users report productivity gains; 12% say AI is core
Vendor-reported: Intuit report built with University of Chicago economists, 34,364 SMB owners in US, Canada, UK, Australia (3,790 in Australia) plus administrative data on 5.3m Intuit customers. Page dated 12-13 May 2026.
Intuit QuickBooks (AI Impact Report, Australia)
25% report reduced costs vs 21% increased costs; 43% report increased revenue; 19% higher headcount vs 6% lower
Vendor-reported owner survey. Net cost effect is roughly neutral; 28% say workdays are shorter, 14% longer. Top barriers: privacy and security 39%, fear of failure 39%.
US Census Bureau (reported by SelfEmployed.com)
Share of employment at firms rating current performance 'excellent': 23.1% for AI users vs 13.8% for non-users; 17.3% of firms using AI in late 2025; 62.0% vs 41.2% rating expected future performance above average or excellent
secondary: trade-site article citing a BTOS data release it dates 23 April 2026 (page date 30 April); the Census release was not opened. Self-rated performance weighted by employment, correlation not causation; the site's 'pays back in a week' claim is the author's, not Census data.
Meta Platforms
Capex guidance 'as much as $145 billion'; four hyperscalers combined up to $725 billion for 2026
secondary: Yahoo Finance report on Q1 2026 results; combined figure is the article's estimate, not a company disclosure.
Samsung Electronics
Revenue 134 trillion won; operating profit 57 trillion won; HBM sales expected to more than triple in 2026
secondary: third-party transcript with apparent transcription errors; verify against Samsung's release. Shows AI profit accruing to suppliers of AI hardware.
Alphabet; Meta; Microsoft
Alphabet $180B-$190B; Meta $125B-$145B; Microsoft about $190B total this year
secondary: Fortune summary of Q1 2026 earnings. Microsoft's increase partly attributed to higher component pricing (about $25B per Nadella as reported). AWS revenue $37.6B, up 28% year over year; Amazon capex not stated in the article.
TCS, Infosys, Wipro, HCLTech (via The Register)
FY26 revenue: HCLTech +11.2%, Wipro +4%, Infosys +3.1%, TCS -0.5%; HCLTech FY27 guide -3% to -5%
secondary: Press summary of FY26 results. Executives cited 'AI deflation'; TCS's CEO called it 'degrowth'. Headcount at each firm (all above 200,000) moved by only a few percent with no major cuts. Exact margins and TCS guidance not given in the article.
DBS; OCBC; UOB (Singapore)
DBS headcount down 1,624 (3.9%) to 39,721; three banks combined 104,266, down 2.6% from 107,072
secondary: FinTech News Singapore. DBS said in Feb 2025 it expected to reduce about 4,000 temporary and contract roles over three years, mainly via attrition. The article says the banks did not directly attribute lower headcount to AI productivity gains.
Stanford HAI (reported by Siemens Xcelerator Community)
Measured productivity gains of 14% to 26% in software development and customer support; weaker or negative in tasks needing more judgment
secondary: Siemens blog summary of the AI Index; underlying studies not named on that page, so check against the Stanford report before citing. (Also relevant to workflow: software-engineering.)
Stanford HAI (reported by Siemens Xcelerator Community)
Agent success on real computer environments rose from 12% to ~66% in a year; agents fail about 1 in 3 structured tasks; agent deployment in single digits across nearly all business functions
secondary: Siemens blog summary; benchmark names not given on that page.
Infosys
FY26 revenue growth 3.1% cc; FY27 guidance 1.5% to 3.5% cc; large deals $14.9 bn; over 30,000 developers on GitHub Copilot
Earnings call transcript furnished to the SEC (Form 6-K). Management said AI deals generally carry better pricing and margins but productivity gains are largely passed to clients; no AI revenue figure given. vendor-reported: client results such as Hertz COBOL migration at 60% lower cost and BP payment accuracy of 95% are company claims.
FPT Corporation (company release)
Q1 revenue VND 12,480B (+8.7%); Global IT Services VND 9,034B (+10.4%); Japan +18.8%; new bookings VND 13,833B (+22.2%)
Vendor-reported. Eight contracts above $10M signed in Q1 (+14.3% YoY). The prior-year base was restated for equity-method treatment of FPT Telecom. A materials-manufacturer AI deal on FleziPT is claimed to cut the client's costs about 40% over three years (vendor claim, unverified).
UnitedHealth Group (reported by Fierce Healthcare)
$1.5B AI investment; 2:1 return target; 12-18 month paybacks; OptumReal ~500M claims processed
secondary: Fierce Healthcare on the Q1 2026 earnings call. About a third (~$500M) targets OptumInsight becoming AI-first. A digital prior-authorisation tool reportedly has a 96% approval rate in early months. Management-stated targets, not audited results.
Wipro
FY26 IT Services revenue $10.5 billion, down 1.6% YoY; Q1 FY27 guidance -2% to 0% sequential
Earnings call as transcribed by a third-party aggregator. CFO flagged margin volatility from wage increases, lower-margin large deals and AI-related investments; no AI revenue figure given. secondary: transcript site, not the company filing.
Gallup
Frequent use: leaders 67%, managers 52%, project managers 50%, individual contributors 46% (n=23,717 US employees; fielded 2026-02-04 to 2026-02-19)
Gallup survey of US employees in organisations that make AI available; frequent use means a few times a week or more.
Gallup
78% of frequent AI users strongly agree their manager actively supports AI use vs 44% of others; organisation supports experimenting 72% vs 44%; clear policies 68% vs 47%
Association rather than causal result. Same survey as the leader/IC usage cut (n=23,717).
Gallup
Non-users vs infrequent users: prefer current way of working 46% vs 36%; ethically opposed 43% vs 25%; do not believe AI can assist 39% vs 22%
Barrier breakdown from the Gallup Feb 2026 US employee survey (n=23,717).
SEC (FY2025 enforcement results)
1 AI-related matter listed in the release (Nate, Inc.), over $42 million raised
SEC release covering fiscal 2025 (1 Oct 2024 to 30 Sep 2025) lists the Nate founder case as the AI-related item among the matters highlighted. Useful scale check: AI-washing enforcement is real but a small slice of headline cases.
Ambient Clinical Documentation Collaborative (Mass General Brigham, UCSF and others), JAMA
-16 min documentation and -13 min EHR time per 8 clinical hours (~10% and ~3%); +0.5 visits/week; ~$167/month added revenue per clinician
secondary: Healthcare in Europe report on a JAMA study of 1,800+ scribe users vs 6,770 controls across five US academic centres, tracked over two years. Only 32% used the scribe on more than half of visits; those users saw about 2x the EHR and 3x the documentation reduction. After-hours EHR time did not differ significantly. Authors call the savings modest. Observational, not randomised.
VinUniversity
77% of surveyed businesses used AI in specific functions; 56% planned to raise spending
Report of 100+ industry professionals, cited by TechRepublic. Methodology differs from the AWS study, so figures are not directly comparable.
Stanford HAI
Organisational AI use rose from 55% (2023) to 88% (2025); Europe 91%, North America 90%; 32% of organisations expect AI-related workforce decreases next year
Compiled from third-party surveys (largely McKinsey). Exact release day not on the page I opened; month taken from secondary coverage dated April 2026. Primary Stanford chapter opened.
Stanford HAI
2021-2025: headcount of software developers and customer support agents aged 22-25 fell while most older age groups grew
Observational labour data compiled by Stanford; not proof AI is the cause. Month from secondary coverage. Primary chapter opened. (Also relevant to workflow: software-engineering.)
US courts (Charlotin database)
Over 1,353 hallucination cases catalogued globally; Oregon federal court $110,000; Sixth Circuit $30,000; Nebraska lawyer suspended
secondary: compliance blog summarising Damien Charlotin's public AI hallucination database and individual rulings from March-April 2026. Listed amounts sum to about $155,000 against the article's $145,000 Q1 claim, so treat totals as approximate; month is the latest event cited, not the article date.
Cambridge Centre for Alternative Finance (Cambridge Judge Business School)
81% adopting AI; 40% at advanced stages; 14% say transformational; 55% of industry and 63% of regulators find AI value hard to measure; 40% report higher profitability, 43% no change
Global survey with partners including BIS, IMF and WEF; 130 regulators responded, other respondent counts not given on the landing page. Date taken from the PDF path (April 2026). 52% of industry actively adopting agentic AI.
Estonian Ministry of Education and Research / AI Leap Foundation
154 of 156 upper-secondary schools; 4,800 teachers; 20,000 students; EUR 4 million pilot budget (about EUR 200 per student); baseline student AI use 64-90% of grades 6-12
self-reported: government-linked factsheet; reports participation and baseline usage only, no pilot outcome results. Date inferred from filename (1 April 2026). Scale-up adds 38,000 learners in 2026-27.
European Central Bank
38% of euro-area firms at advanced stage and 33% early stage; significant AI users expect 2.8% employment growth vs 0.6% for non-users (SAFE survey Q4 2025)
Expectations not realised outcomes; respondent count not stated. Primary ECB Economic Bulletin box opened.
Mingdeng Du, Wuhan University (arXiv)
600x fall from $60 to $0.10 per million tokens; economy-tier price half-life 1.10 years; reasoning models 31.5x non-reasoning prices on average
Working paper on 2020-2026 pricing. Mid-tier half-life 1.55 years; the flagship tier fits exponential decay poorly (R-squared 0.031). Cost reduction attributed to productivity gains, not GPU hardware. Provider concentration (HHI) fell 54% from 2023 to 2026. Not peer reviewed.
Federal Reserve Bank of Atlanta with Bank of England Bundesbank and Stanford (Yotzov Bloom et al.)
Over 80% of firms report no impact of AI on employment or productivity over the last 3 years; ~70% actively use AI; firms expect +1.4% productivity and -0.7% employment over next 3 years (~6,000 executives in US UK Germany Australia)
Self-reported executive survey in an Atlanta Fed working paper (WP 2026-3); central-bank-affiliated primary page opened. Number of firms not stated.
Federal Reserve Bank of Atlanta with Bank of England Bundesbank and Stanford
Over two-thirds of top executives use AI regularly but average 1.5 hours per week; one quarter report no use (~6,000 executives)
Self-reported; relevant to leadership ownership. Primary working-paper page opened.
Eliciting Insights (reported by Fierce Healthcare)
75% use at least one AI application (59% in 2025); clinical note-taking 68%; more than half of those quantifying ROI report 2x+
secondary: second annual survey of executives at 120 US health systems. Other adoption: AI documentation improvement 43%, AI coding 36%, patient-text drafts 36%, AI denial prediction 25%. ROI claim covers only systems able to quantify ROI. Go-to-market research firm, not independent academic.
Concentrix
Revenue $2,500.4M (+1.9% constant currency); non-GAAP operating margin 11.8% vs 13.6%; FY26 guide $10.035B-$10.18B
Q1 FY26 results exhibit (8-K). Baseline for the later guidance cut: FY26 revenue guidance fell by about $110M at the low end by June. CEO said the company helps clients 'capture measurable value from AI'.
FPT Corporation (Vietnam)
~70% of common inquiries handled by AI; labour productivity per employee 'tripled' (no baseline)
VNA report via Bao Tay Ninh; company claim, no before/after cost or time figures and no stated baseline for the tripling. Useful as an example of an unmeasured headline claim.
Canary Technologies
82% expect growth in AI use within a year; 85% will put 5%+ of IT budget into AI; 71% see significant or transformative industry impact
Vendor-reported: Canary Technologies survey of 400+ hotel technology buyers across North America, EMEA and APAC, run early 2026. Benefits cited (staff time, guest satisfaction, revenue) were self-reported, not quantified. secondary: Hotel Management article.
Protiviti and BoardProspects
26% of boards discuss AI at every meeting; 63% of high-ROI firms vs 13% of low-ROI firms put it on every agenda
Third annual Global Board Governance Survey, fielded Q4 2025 with 772 board members and C-suite executives. Firms reporting high AI ROI were far more likely to treat AI as a standing board item, and 95% of them were confident integrating AI in operations versus 33% of low-ROI firms.
Goldman Sachs 10,000 Small Businesses / Babson College / David Binder Research
76% of small businesses use AI; 93% of users report positive impact; 84% cite efficiency
Survey of 1,256 owners in the Goldman Sachs 10,000 Small Businesses programme (a programme alumni sample, not a random national sample), fielded 27 Jan to 4 Feb 2026. All figures are self-reported by owners.
Goldman Sachs 10,000 Small Businesses / Babson College / David Binder Research
14% fully integrating AI into core operations; 49% lack technical expertise; 73% want training
Among AI users, 50% cite data privacy and security, 49% lack of technical expertise and 48% difficulty choosing tools; 73% say they would benefit from training and resources to implement and evaluate AI. Self-reported, programme-participant sample (n=1,256).
Goldman Sachs 10,000 Small Businesses / Babson College / David Binder Research
87% say AI augments employees; 67% expect increased revenue
Owner perception, not measured workforce data. 67% expect revenue to rise from AI use (expectation, not realised result). Programme-participant sample, n=1,256.
American Medical Association (reported by HCI Innovation Group)
Use 38% (2023) to 81% (2026); average use cases 1.1 to 2.3
secondary: HCI Innovation Group report on the AMA 2026 survey. Adds the 2023 comparison and breadth of use (2.3 use cases per physician); 40% are equally excited and concerned. Self-reported. The AMA's own page gives 81% and ~1,700 respondents.
Zalando
AI product content 0% to 90% of output in a year; content output +70%; campaign creation six weeks to days
Company-reported in the FY2025 results release alongside revenue of EUR 12.3bn (+16.8%) and adjusted EBIT of EUR 591m (+15.6%). Link between AI and the profit line is not quantified.
Zalando
Zalando Assistant 6 million users (4x); delivery-promise precision +22 percentage points; code changes up over 20% with about 3,000 tech staff on AI coding tools
Company-reported, FY2025 results release (12 Mar 2026). Code-change count is an activity measure, not quality or cost.
American Medical Association
81% of ~1,700 surveyed physicians use AI professionally; 88% worry about skill loss; 85% want a say in adoption
AMA Center for Digital Health and AI 2026 Physician Survey on Augmented Intelligence, nearly 1,700 physicians (fieldwork dates not stated on page). Top uses: research summaries 39%, discharge instructions/care plans/notes 30%, billing codes and visit notes 28%, chart summaries 28%, assistive diagnosis 17%. 70% see AI as a way to automate burnout-driving tasks. Self-reported usage, not measured outcomes.
Wolters Kluwer (Future Ready Lawyer 2026)
62% report 6-20% weekly time savings; 32% attribute an 11-20% revenue increase to AI; 92% use at least one AI tool
Vendor-reported: Wolters Kluwer sells legal software. Sample size, geography and fieldwork are not stated on the page. Barriers cited: ethics and data privacy (39%), inadequate training (39%), resistance to change (35%).
Deloitte AI Institute
28% moved 40%+ of AI experiments to production; 18% achieving revenue benefits; 89% have not redesigned jobs; only 15% rate talent preparedness high (LSHC n=378)
Self-reported sector cut of the Aug-Sep 2025 survey; no ROI figure given. Primary Deloitte PDF opened.
Atlanta Fed; Bank of England DMP; Bundesbank; Macquarie University (via CEPR VoxEU)
5,000+ CFOs and executives, same questions Nov 2025-Jan 2026: AI use US 78%, UK 71%, Germany 65%, Australia 59% (about 69% overall); 89% of firms report no productivity impact in the last three years (average +0.29%); expect +1.4% productivity over three years; UK follow-up: about two-thirds of expected job reduction is fewer new hires, not layoffs
Self-reported executive survey, VoxEU column read. Same underlying paper as P3-028/029; added here for the country split, the Bank of England panel detail and the hiring-slowdown finding. Column date not shown, month taken from the working-paper date (24 Mar 2026).
Baslandze et al. (NBER Working Paper 34984)
Mean AI-attributed labour productivity growth 1.8% in 2025, expected 3.0% in 2026; revenue-implied 0.6% and 1.8%; about 750 senior financial executives
Working paper, not peer-reviewed. Manufacturing, construction and low-skill services show smaller positive gains than finance and high-skill services. No AI-driven employment effect in 2025; a 0.1% decline expected in 2026 in the sample. The gap between perceived and revenue-implied gains is framed as a 'productivity paradox'.
AI Hallucination Cases database (Damien Charlotin), reported by JL Ellis
Over 600 cases in December 2025; over 1,000 by March 2026; about half involve self-represented litigants
secondary: blog post summarising the public database; sanctions range from warnings to financial penalties to at least one suspension. Count depends on the database's own inclusion rules.
RIETI / Money Forward Institute of Economic Research (Konishi and Kubo)
About 87,000 SMEs; payment records for generative AI services 2022 to 2025
Behavioural data rather than self-report: adoption rose sharply and in sync across industries in early 2023, and later differences track pre-existing digital infrastructure. The summary page gives no adoption percentages (RIETI Discussion Paper 26-J-018).
Block, Inc.
More than 4,000 of '10,000 plus' employees cut
secondary: Fortune reporting on Block's shareholder letter. Quoted: 'A significantly smaller team, using the tools we're building, can do more and do it better.' Shares rose more than 20% premarket. Resulting headcount not stated in the article.
ServiceNow
Over 90% of targeted Level 1 volume handled autonomously; resolution above 99% for those categories
secondary: The Register. Vendor self-reported in its own environment, tested by employees, with general availability only expected in the second half of 2026. Scope is password resets, unlocks, access and VPN issues; no durations or baseline given.
Gartner
55% expect fewer entry-level hires; 51% expect overall workforce reductions; 86% say new talent-pipeline processes are needed; 509 leaders
Gartner survey of 509 supply chain leaders worldwide, fielded July to October 2025, plus 15 interviews with Supply Chain Top 25 companies. Expectations, not measured headcount change.
Gartner (supply chain practice)
55% expect lower entry-level hiring; 51% expect overall workforce levels to shift downward; 86% say talent development needs new processes
Gartner survey of 509 supply chain leaders (July-Oct 2025) plus interviews with 15 Supply Chain Top 25 leaders. Expectation data, not realised headcount change.
Oxford University Hospitals NHS Foundation Trust (reported by HTN)
37.3% of 398 logged uses needed accuracy edits; 24 hallucination edits; 44.4% of users saw a hallucination at least once
secondary: HTN report of the pilot evaluation of Accurx Scribe, Tortus, Heidi and Lyrebird. 88% of clinicians had measurable time savings and the share spending 5 to 10 own hours on documentation fell from 24.2% to 2.2%, but edits were needed for 126 clinical-information errors, 43 spelling/grammar and 24 hallucinations. Useful counterweight to headline time-saving claims.
European Central Bank (Banking Supervision)
More than 85% of large banks under European supervision use AI in some form
Speech by Pedro Machado, ECB Supervisory Board representative, citing ECB annual data collection on innovative technologies. Says use is spreading into IT operations, legal and document analysis and customer-facing tools, and that governance is uneven.
Oxford University Hospitals NHS Foundation Trust
Nearly 90% report reduced documentation time; over half save 30+ minutes a day; 73% report improved wellbeing
Trust's own announcement of a pilot of four ambient voice technology tools across inpatient and outpatient settings. 99.7% of patients consented. Self-reported time savings; sample size not stated on the page. Dictation quality rated excellent or good by almost 70%.
Telstra (reported by CFOtech Australia)
Software production and release schedules about 20% faster; 380 internal AI use cases; agents deployed for multi-step tasks; over 75% of staff with AI tools use them weekly (half-year to 31 Dec 2025)
secondary: trade article on Telstra's results; company-reported, no baseline or method. Results themselves not opened.
Telstra (reported by CFOtech Australia)
AI Virtual Assistant (launched Nov 2025) nearly tripled the share of customers resolving enquiries themselves; 86% of consumer service interactions completed through digital self-service
secondary and company-reported; no baseline percentage given for the tripling.
Telstra (reported by CFOtech Australia)
2,356 roles cut in calendar 2025, headcount down over 7% to just over 29,000; underlying labour costs -9% (A$181M); underlying operating expenses -A$179M (2.4%); profit +8.1% to A$1.2bn; smaller workforce linked to AI by 2030 with no number; CEO said she could not point to a specific role AI directly replaced; Telstra denies AI drove a 550-role cut announced July 2025
secondary: company results as reported by one trade outlet; cost fall is real but attribution to AI is not shown. One date inconsistency in the article (Data and AI Academy completions).
State Bank of Vietnam (reported by Asia News Network)
Existing AI systems have until September 2027 to comply
secondary: Asia News Network report on a draft SBV circular. Banks would have to tell customers before using chatbots or virtual assistants, disclose AI-generated content, avoid exploiting vulnerable customers and assign humans to review AI-decision complaints. Draft status; final text not checked.
Walmart (Q4 FY2026 earnings call, CFO John David Rainey)
About 60% of stores receive some freight from automated DCs; about half of US e-commerce fulfilment volume automated; AI in associate app reaches 1.5 million workers
secondary: CIO Dive report of the earnings call; self-reported. The article presents this as automation rather than AI and gives no dollar or percentage contribution to results.
Booking Holdings
Customer-service cost per reservation down about 10% year on year while bookings rose about 10%
Company-reported by CFO Ewout Steenbergen on the Q4 2025 earnings call. secondary: CFO Dive (19 Feb 2026); contact volumes and AI share of contacts not reported.
Gartner (as reported by SaaStr)
Software spend $1.4T, +14.7% in 2026; GenAI model spend +80.8%; about 9% of IT budgets absorb price rises on existing software
secondary: Blog restating a Gartner forecast dated 3 Feb 2026 (Gartner original not opened). The 9% price-rise figure and the view that much growth is reallocation are the blog author's reading. Renewals increasingly bundle AI features at higher prices.
McKinsey & Company / Singapore EDB / Tech in Asia
46% past pilots to scaling (35% global); 60% report <5% EBIT impact; a further 18% report no financial impact
Primary-page figures from EDB site and Straits Times coverage: 330 senior leaders, survey July-Aug 2025. Replaces the draft figure of 67% under 5% EBIT and an '8% full-scale' number, neither of which appears in the primary sources checked. Executive survey, self-reported.
BNY (reported by InvestmentNews)
134 AI 'digital employees' alongside ~48,100 staff; nearly all staff completed 10-hour AI training; headcount down from ~53,400 in 2023
secondary: InvestmentNews citing CNBC interview; headcount and spend figures attributed 'reportedly'. Goldman analyst EPS estimate not company data.
Goldman Sachs (reported by InvestmentNews)
~6 months of agent development with Anthropic engineers; no measured result published; CIO called direct job cuts 'premature'
secondary: InvestmentNews citing CNBC interview with CIO; 'reportedly' for the six-month detail; no metric.
DBS Bank (reported by FCC Singapore)
AI initiatives generated S$1bn economic value in 2025, up a third from S$750M in 2024; derived by comparing customers offered AI-driven solutions with a control group; CEO credits contextual nudges and automated engagement for record deposits, wealth and fee income
secondary: trade report of DBS Q4 2025 results briefing; company-reported, partly control-group based. DBS Annual Report 2025 CEO letter (opened, no publication date shown, so not used as the row date) states about S$1bn from 2,000+ models, 430+ use cases and a coding assistant cutting data scientists' coding time by up to 20% on certain tasks.
DBS Bank (reported by FCC Singapore)
About 4,000 contract and temporary roles to be cut over a few years through natural attrition (stated early 2025); over 60% of staff actively use in-house DBS GPT; CEO: aim is to move staff into higher-order roles
secondary and company-stated; no total headcount or cost-income ratio in the source. Annual report says DBS-GPT helps two-thirds of employees.
UK Department for Work and Pensions (reported by Computing)
Six-month evaluation Oct 2024-Mar 2025: 1,716 Copilot users vs 2,535 non-users, net 19 minutes per person per day saved, mainly search, email drafting and summarising; 73% say output quality improved
secondary: Computing article; survey-based comparison, self-reported, no time logs; the government report was not opened. A separate DBT trial reportedly found time saved did not always become productivity.
Nuffield Trust / NIHR RSET
Evidence review; no numerical results reported
Phase 1 of the NHS ambient voice technology evaluation. Finds studies consistently show reduced note-writing time but rarely measure patient experience, safety, retention or system capacity; measurement is inconsistent and system-level costs assessed in only a handful of studies. Flags a possible early rise in documentation time because transcripts must be checked. A multi-site real-world evaluation follows.
Salesforce
87% of sales organisations use AI; 54% of sellers have used AI agents; top performers 1.7x more likely to use prospecting agents
Vendor-reported: Salesforce State of Sales survey of 4,050 sales professionals in 22 countries, fielded Aug-Sept 2025, no third-party fieldwork firm named. Respondents expect agents to cut prospect research time 34% and email drafting time 36% once fully implemented (expectations, not measured results).
Salesforce
130,000 leads contacted; 3,200 opportunities created over four months
Vendor self-reported example inside the State of Sales announcement. No win-rate, revenue or cost comparison with human SDRs is given, so it shows activity and pipeline creation only.
European Central Bank (Survey on the Access to Finance of Enterprises)
35% of SMEs no AI use vs 13% of large firms
Fieldwork 19 Nov to 15 Dec 2025. The share making significant use is similar for SMEs and large firms, so the gap is in experimentation and moderate use.
Deloitte AI Institute
26% moved 40%+ of AI experiments to production; 17% achieving AI-driven revenue growth; only 20% of workers with access use AI daily more than 60% of the time; 84% have not redesigned jobs (ER&I n=591)
Self-reported sector cut of the same Aug-Sep 2025 survey. Report gives no quantified ROI. Primary Deloitte PDF opened.
Careerminds (reported by People Matters)
35.6% of employers had rehired more than half of roles eliminated due to AI; 52% of rehires within six months (survey of 600 HR professionals Feb 2026)
secondary: reported by People Matters; vendor survey of an outplacement firm; Robert Half cited ~32% of US hiring managers refilled AI-cut roles.
NBER (Cruces Fernandez Meijide Galiani Galvez Lombardi)
Education gap 0.548 SD in control fell to 0.139 SD with AI; 1,174 adults aged 25-45
Online workplace-style task experiment, not a firm field study; gap partly reappeared in follow-up task without AI (NBER w34851, revised May 2026).
MASAI trial, Sweden (The Lancet), reported by ecancer
Interval cancers 1.55 vs 1.76 per 1,000 (-12%); aggressive-subtype cancers -27%; reading workload -44% (interim)
secondary: ecancer report of full randomised results: 53,043 women AI-supported vs 52,872 standard double reading, four Swedish sites. The 44% radiologist-workload reduction comes from the earlier interim analysis. Strong RCT evidence from a public screening programme.
City of New York
Mayor said bot cost around half a million dollars; foundations reportedly nearly $600,000; city took it down in February 2026 ('beta test has ended')
secondary: The Markup investigative report. Launched fall 2023 on Microsoft cloud; wrongly suggested landlords could discriminate against Section 8 tenants and that refusing cash payment was fine. Mayor called it 'functionally unusable'. No usage numbers published.
Lloyds Banking Group
About GBP 50 million of GenAI value in 2025; more than GBP 100 million additional expected in 2026; internal search time cut 66%; HR assistant resolves around 90% of queries first contact
secondary: The Asian Banker reporting Lloyds' results release. Over 50 GenAI solutions deployed; Athena used by about 20,000 staff; GitHub Copilot used by about 5,000 engineers. Company-reported figures; method of valuing 'value' not shown.
Ping An Life
93% of automatically reviewed claims accurate within 60 seconds; claim approval rate 99.2%; 41.51 billion yuan paid; fastest claim 8 seconds
secondary: aggregator (Tiger Brokers news feed) relaying Ping An's 2025 claims report; company-reported. Review engine uses the DeepSeek model per the report.
UK DSIT and DBT (IFF Research and Technopolis)
16% of UK private-sector businesses with 5+ employees use AI (large 36%, mid-sized 23%, micro 14%); 3,500 phone interviews, fieldwork 12 Feb-2 May 2025, plus 100 qualitative interviews
Primary gov.uk page opened. Self-reported by the technology lead; excludes shadow AI; fieldwork is 2025 so older than the date suggests.
UK DSIT and DBT (IFF Research and Technopolis)
Among 700 AI-using businesses: 75% report improved workforce productivity, 34% lower operating costs, 10% no impact so far; revenue 77% no change and 12% increase; barriers: no identified need 71%, limited AI skills 60%, tools 48%, integration 29%
Primary page opened. All impact figures are self-reported by businesses (the report itself calls them estimates).
UPS (Q4 2025 results, reported by Supply Chain 24/7)
$3.5bn savings from network reconfiguration and Efficiency Reimagined; 195 operations and 93 buildings closed; up to 30,000 operational positions to go in 2026; 2026 savings target $3bn
secondary: Supply Chain 24/7 report of UPS announcements; figures are company-reported and the workforce cut is described by an analyst as not necessarily mass layoffs. Automation, not AI alone, drives the savings.
Saudi Aramco (CEO Amin Nasser at Davos, reported by Bloomberg)
$3bn to $5bn technology realised value expected for 2025; $6bn combined for 2023 and 2024
secondary: Bloomberg report of CEO remarks; self-reported by Aramco. Figure is 'technology realised value' (headline frames it as AI-linked), a range rather than a booked number. Publication date taken from the article URL and Davos caption, not stated in the text.
PwC
56% of CEOs report no significant financial benefit from AI yet; 33% gains in either cost or revenue; 12% both (n=4,454 CEOs, 95 countries and territories)
Self-reported CEO survey by an advisory firm that sells AI services. Primary PwC release opened.
PwC
CEOs reporting both cost and revenue gains are 2-3x more likely to have embedded AI extensively; those with strong AI foundations (e.g. Responsible AI frameworks) are 3x more likely to report meaningful returns (n=4,454)
Self-reported survey; correlation not causation. Primary PwC release opened.
PwC
Companies applying AI widely across products, services and customer experience had nearly 4 percentage points higher profit margins (PwC separate analysis)
PwC's own analysis; method not given in the release so causality is unproven. Primary PwC release opened.
Lenovo CIO Playbook 2026 with IDC
96% planning to increase AI investment by an average of 15% in 2026; 88% expect positive ROI; two-thirds piloting or systematically adopting AI across functions
Lenovo CIO Playbook 2026 with IDC insights as reported by Context.ph. Sample size and fieldwork dates not stated.
Unilever / World Economic Forum Global Lighthouse Network
Forecasting accuracy +39%; lead times -75%; operating costs -24%; 31 use cases, 65% AI or GenAI-driven
One of the few Lighthouse entries that states the AI share of use cases. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Ford Otosan / World Economic Forum Global Lighthouse Network
Production volume +100%; complexity up twelve-fold; labour productivity +44%; quality +6%
Productivity category. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Faurecia Automotive Systems / World Economic Forum Global Lighthouse Network
Customer complaints -94%; related costs -62.5%; scrap costs -75.8%; OEE +10.2%
Productivity category. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Siemens / World Economic Forum Global Lighthouse Network
Lead time -78% (45 days to 10); time-to-market -33%; field failures -46%; productivity +14%
Productivity category. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Midea Kitchen & Bath / World Economic Forum Global Lighthouse Network
Delivery lead time -39%; inventory days -30%; market defect rate -86%; 113 digital use cases, 35% AI-driven
Supply Chain Resilience category. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
Schneider Electric / World Economic Forum Global Lighthouse Network
Onboarding 75 to 15 days; 56% of employees upskilled; technician turnover -42% over five years; automation +55%
Talent category. Shows the workforce side of an automation programme: upskilling and retention alongside output. Self-reported by the sites to the World Economic Forum, not independently verified; the programme covers digital and AI technologies together, so gains are not attributable to AI alone.
World Economic Forum Global Lighthouse Network (Lumina)
94% of successful transformations combine multiple technology domains; sites pairing technology with workforce and sustainability outperform peers by an average of 16% or more
Network-level synthesis of 220+ Lighthouse sites in 30+ countries and eight years of data. The 16% outperformance is the Forum's characterisation of a self-selected, WEF-recognised sample, not a controlled comparison.
Zurich Insurance / Interloom
Vendor-reported: case triage 42% faster; workflow updates ~3x more per quarter
Vendor customer story on Interloom's own site; no baseline, period, sample or method disclosed and no independent verification found. Include only as an example of how thin vendor claims can be next to the Cytora deployment.
Gartner
Stock fell 47.9% in 2025; revenue up 2.7% last quarter; contract value growth single digit
secondary: Motley Fool commentary; AI concerns are the author's and investors', not a company statement.
LegalOn Technologies and In-House Connect
52% using or evaluating contract AI; 79% report less time on routine tasks; 67% faster business response
Vendor-reported: survey of 452 in-house legal professionals (fieldwork December 2025), published by a contract-review software vendor; as reported by Artificial Lawyer. 87% expect AI to help with pre-signature review.
LegalOn Technologies and In-House Connect
3.1 hours average per contract; 78% comfortable with agent first-pass review under attorney supervision; 80% exploring agents
Vendor-reported survey of 452 in-house legal professionals (December 2025). Only a small minority have fully comprehensive playbooks, which limits how far automation can go. Stated comfort is attitude, not measured results.
Equinor
About $130m value or savings in 2025; over $330m since 2020; 100+ new AI use cases identified
Self-reported by Equinor, no independent verification and no baseline stated; value is for Equinor and its partners. Includes $12m saved by AI-found solution on Johan Sverdrup phase 3 and tenfold seismic interpretation capacity (2 million km2 in 2025).
Equinor
About $120m value since 2020; 700+ rotating machines monitored with 24,000 sensors
Self-reported by Equinor, no baseline or method for the value calculation given. A rare multi-year, per-use-case figure from a large energy operator.
Deloitte AI Institute
25% of respondents have moved 40% or more of AI experiments into production; 66% report efficiency gains, 40% cost reduction, 20% revenue growth; 37% have little or no process change (n=3,235 director to C-suite; 24 countries; fielded Aug-Sep 2025)
Self-reported survey. The PDF states n=3,235 in text but N=3,325 in one figure; efficiency and cost percentages were hard to map to chart labels (extraction caveat). Primary Deloitte PDF opened.
Deloitte AI Institute
84% have not redesigned jobs around AI; 53% respond mainly by educating the workforce; worker access to sanctioned tools ~60% but under 60% of those with access use AI in daily workflow (n=3,235)
Self-reported survey. Primary Deloitte PDF opened.
Deloitte AI Institute
21% report a mature governance model for autonomous agents; 74% plan at least moderate agentic AI use within two years vs 23% today (n=3,235)
Self-reported survey. Primary Deloitte PDF opened.
Citigroup (reported by CeFPro)
9% productivity lift in coding teams (company-stated)
secondary: CeFPro reporting conference remarks by Citi's incoming CFO; baseline and period not given; self-reported.
JPMorgan; Wells Fargo; PNC (reported by CeFPro)
JPMorgan ops specialists could see 40-50% efficiency gains (forecast); Wells Fargo no headcount reduction; PNC headcount flat vs a decade ago while bank ~tripled
secondary: CeFPro on conference remarks, not earnings filings; JPMorgan 40-50% is a projection and 'productivity doubled' claim had unclear units.
OECD
Qualitative: SME AI adoption is low versus other digital technologies and versus large firms; four enablers named: connectivity; data, algorithms and compute; skills; finance; proposes a taxonomy of SME adopters from G7 case studies
Only the abstract page was opened; the full PDF could not be fetched, so no country figures were verified. Treat as a pointer to the primary report, not a statistic source.
Menlo Ventures
About $37B in 2025 vs $11.5B in 2024 (3.2x); 76% of use cases bought (2024: 53% bought, 47% built); coding $4.0B
Vendor-investor report (an investor in AI companies). Buying beat building decisively. Reported enterprise LLM API share: Anthropic about 40%, OpenAI about 27%, Google about 21%. Anthropic is a Menlo portfolio company, so treat share figures with care.
FINRA
GenAI use cases listed rose from 3 in the 2025 report to about a dozen more in 2026
secondary: law-firm summary (Debevoise, 12 Dec 2025) of FINRA's 2026 Annual Regulatory Oversight Report. Lists agent risks including autonomy and scope creep, auditability and sensitive data handling; expects human-in-the-loop review and logs.
Wells Fargo (reported by MarketScreener)
Engineers '30% to 35% more efficient' at writing code; a little over 210,000 employees at 30 September 2025 vs 275,000 in 2019
secondary: MarketScreener summary of CEO Charlie Scharf at a Goldman Sachs conference. Management statement, not an audited measure; he said the bank has not reduced coder numbers and AI will roll out gradually.
Gong Labs
Frequent AI users 77% more revenue than non-users; reps at or above quota fell from 52% to 46%
Vendor-reported: Gong State of Revenue AI survey of 3,048 revenue leaders (US, UK, Germany, Australia) plus analysis of 7.1 million opportunities from 3,613 Gong customers. The report itself says this is a comparison, not proven causation. Publication date taken from secondary coverage (getcoai, 4 Dec 2025); the PDF is undated.
Gong Labs
Core-strategy AI users 31% higher revenue growth and 65% more likely to raise win rates than limited pilots; 59% of teams use personal AI tools
Vendor-reported, same Gong dataset (3,048 leaders; 7.1M opportunities). Self-selected Gong customers; US revenue growth averaged 16% in 2025, down 3 points. 28% of leaders expect job eliminations versus 43% expecting jobs to transform without headcount cuts. Date from secondary coverage.
NEJM AI (Afshar et al.)
Time on notes -0.36 hours/day (95% CI -0.55 to -0.17); work exhaustion -0.44 points on 5-point scale; 66 practitioners; 71,487 notes
Randomised pragmatic trial; date inferred from citation NEJM AI 2025;2(12). Secondary summary by Consultant360; work-outside-work result not robust to outliers.
UCLA Health (NEJM AI)
Nabla arm note-writing time fell 41 seconds (4:30 to 3:49), 9.5% more than control; DAX not significant; burnout scores ~7% better in both arms (238 physicians, 72,000 encounters)
Randomised trial; press release summary opened (not the paper). Gains are modest, effect on burnout needs confirmation per release.
HP Inc.
4,000 to 6,000 jobs by FY2028; $1 billion gross run rate savings over three years
secondary: Reuters-based report of HP's fiscal Q4 announcement. CEO Enrique Lores quoted: 'We expect this initiative will create $1 billion in gross run rate savings over three years.' Affected teams: product development, internal operations, customer support. Restructuring charge amount not stated in the article.
Gartner (finance practice)
59% of finance functions use AI (58% in 2024, 37% in 2023); 91% report low or moderate initial impact
Gartner AI in Finance Survey of 183 CFOs and senior finance leaders, fielded May-June 2025. Growth has stalled after the 2023-24 jump; data quality and data literacy are the biggest obstacles and 25% are unsure how to move from planning to piloting. More mature adopters are nearly three times as likely to report high impact.
Gartner (finance practice)
AP process automation used by 37% of finance AI adopters; knowledge management 49%; error and anomaly detection 34%
Same Gartner survey of 183 finance leaders (May-June 2025). Only three use cases pass one-third adoption among AI-using finance functions; code generation ranks highest for impact despite low adoption. A prioritisation signal for firms choosing a first finance use case.
Korea Chamber of Commerce and Industry (KCCI)
82.3% not using AI; 49.2% of large firms vs 4.2% of small and mid-sized firms
secondary: Korea JoongAng Daily reporting a KCCI survey of 504 manufacturers. 73.6% say AI investment strains budgets.
Korea Chamber of Commerce and Industry (KCCI)
80.7% no AI talent; 60.6% expect minimal effect; 14.5% plan to retrain staff, 3.4% to hire
secondary: Korea JoongAng Daily reporting a KCCI survey of 504 manufacturers. Expectation data, not measured outcomes.
Microsoft Research and Harvard Business School (Dillon Jaffe Immorlica Stanton)
66 firms and 7,137 knowledge workers; two fewer hours on email per week in second half of 6-month trial; no detectable shift in task quantity or composition
vendor-reported: Microsoft product, partly Microsoft authors. Date is arXiv v4.
EY India (AIdea of India survey)
47% multiple use cases live; 10% scaling; 23% early or pilot; 47% say over 10% of proofs of concept reached production
Survey of over 200 Indian enterprise leaders. The page gives no measured ROI; 76% expect significant impact (expectation).
Xero (small business AI research, Australia, UK, US)
1,100 small businesses; about 44% weekly AI use; one in five report revenue increase
secondary: CFOtech Australia reporting Xero research; vendor-reported, methodology and fieldwork dates not disclosed. Daily AI users are more than twice as likely to report growth as non-users (correlation).
Monetary Authority of Singapore
Comments due 31 January 2026
secondary: Mondo Visione republication of the MAS consultation release. Proposes AI inventories, materiality assessment of each AI use, lifecycle controls and board oversight, covering generative AI and AI agents; builds on MAS's 2024 thematic review of major banks.
Rakuten Group
Token usage 17x YoY; 30% of usage for software development; 23,000 merchants monthly use AI tools
vendor-reported: company results coverage. The page quantifies usage but gives no AI-driven cost or revenue gain; useful example of adoption metrics standing in for P&L impact.
Google, Temasek and Bain & Company (e-Conomy SEA 2025)
More than $2.3bn into over 680 AI startups in 12 months; AI above 30% of H1 2025 private funding; Singapore $1.31bn
Regional digital economy GMV on track for over $300bn in 2025. Funding is an input metric, not evidence of realised business value.
Google, Temasek and Bain & Company (e-Conomy SEA 2025)
79% of workers learned to use AI; 43% use it both personally and professionally
Self-reported worker survey; the press release does not give sample size.
McKinsey & Company
39% of respondents attribute any enterprise EBIT impact to AI; most of those say under 5% of EBIT (n=1,993 in 105 nations; fielded Jun 25-Jul 29 2025)
Self-reported survey by McKinsey of its respondent panel; primary page opened.
McKinsey & Company
~6% are AI high performers (5%+ of EBIT from AI); they are nearly 3x as likely to have fundamentally redesigned individual workflows (n=1,993)
Self-reported survey; McKinsey calls workflow redesign one of the strongest contributors to impact among factors tested. Primary page opened.
McKinsey & Company
Nearly two-thirds of respondents say their organisation has not begun scaling AI enterprise-wide; about one-third have begun (n=1,993)
Self-reported survey; page gives no exact percentage for 'scaling'. Primary page opened.
McKinsey & Company
23% scaling an agentic AI system somewhere; 39% experimenting; no more than 10% scaling agents in any single function (n=1,993)
Self-reported survey. Primary page opened.
John Deere
Nearly 50% average cut in non-residual herbicide mix; nearly 31 million gallons saved; 5+ million acres; soybean yield +2 bu/acre on average in third-party trials
secondary: CropLife copy of the Deere release; vendor-reported for the 2025 season, with the yield figure attributed to seven-state third-party trials (Beck's 2024 study). Technology fee $5 per in-crop acre under a savings guarantee. No farm-level profit figure given. Release date taken from the Deere news headline in search results.
OECD
5,000+ SMEs in seven countries; 50% of SMEs in four G7 countries say staff lack GenAI skills; 23.6% of users train staff
Policy survey, no P&L outcome data. Among non-users the top reasons are "not suited to our work" (57%), copyright (54%) and data privacy (52.5%). Source: research report of 14 Oct 2026.
Hong Kong Monetary Authority
Suspicious transaction report preparation time down 30% to 80%; document processing from one day to about five minutes; risk assessment reports 60% faster; 86% of users rated system favourably
secondary: The Standard (Hong Kong) reporting HKMA sandbox results. Sandbox trial figures, not production results.
Amazon
Over $10bn expected annual incremental sales; shoppers who use Rufus 60% more likely to complete a purchase
Company-reported: CEO Andy Jassy on the Q3 2025 earnings call (30 Oct 2025), as reported by Fortune on 2 Nov 2025. Forward-looking estimate, not audited; the 60% figure compares users with non-users, so it is not a controlled effect.
Amazon
250 million users in 2025; monthly active users +140% year on year; interactions +210%
Company-reported via Amazon's Q3 2025 earnings call, as reported by Fortune. Usage figures only; no cost or margin data given.
Wharton Human-AI Research with GBK Collective
82% weekly use; 72% formally measure ROI; about 74-75% report positive ROI (self-reported); mid-sized and smaller firms report faster ROI
About 800 US enterprise decision-makers, third annual wave. ROI is respondent-judged and often productivity-based, so it is not comparable with P&L attribution studies. Source: research report of 14 Oct 2026.
Wharton Human-AI Research with GBK Collective
89% agree GenAI enhances employee skills; 43% see risk of declining skill proficiency as usage climbs
Skills cut of the 2025 Wave 3 report; related rows exist for weekly use and ROI. Sample size not stated on the page.
Chegg
388 roles (about 45% of staff) cut; $15m-$19m restructuring charges; $100m-$110m planned 2026 non-GAAP cost cuts
Chegg cited the 'new realities of AI' and reduced Google referrals. Founder-CEO Dan Rosensweig returned as CEO and the company kept its standalone status. secondary: Fortune report of 28 Oct 2025.
Mondelez International
Marketing content production costs down 30% to 50%
secondary: summary of a Reuters report. Tool built since 2024 with Publicis and Accenture; first short videos live on social, TV ads targeted for the 2026 holiday season. Company-stated figure; the article does not explain how savings were measured.
Menlo Ventures (reported by HIT Consultant)
$1.4B spend (~3x prior year); 22% deployed; health systems 27%, outpatient 18%, payers 14%; 700+ executives
secondary: HIT Consultant coverage of Menlo's State of AI in Healthcare. Claims healthcare adopts AI 2.2x faster than the wider economy. Investor-sponsored survey; spend estimate method not shown on the page.
Amazon (internal documents reported by The New York Times)
160,000 hires avoided by 2027; about 600,000 by 2033 if sales double; about 30 cents saved per item; goal to automate 75% of operations
secondary: Benzinga report (republished on Yahoo Finance) of internal Amazon documents seen by the NYT. The headline says jobs 'vanish' but the documents concern future hires avoided, not existing jobs. Amazon says the documents reflect one team's view and are incomplete.
Amazon (internal documents reported by The New York Times)
25% to 50% fewer human workers at Shreveport, Louisiana; design to be copied at 40 facilities by end of 2027
secondary: Benzinga report of NYT-viewed documents. Figure is from internal planning material, not an audited result; Amazon disputes that the documents represent its hiring strategy.
UK Government (gov.uk)
Over 50,000 responses categorised in around 2 hours; run cost £240; experts needed 22 hours to check results; potentially 75,000 days of manual analysis a year (£20 million)
Government press release on the Independent Water Commission review; savings are potential, not audited.
Deloitte Australia / Department of Employment and Workplace Relations
A$440,000 contract; final instalment repaid; up to 20 fabricated citations found
237-page report on automated welfare penalties contained fake references and a fake court quote; a revised version disclosed Azure OpenAI use. Deloitte said findings were unaffected. secondary: BABL AI report of 13 Oct 2025.
Financial Stability Board
FSB report reviewing monitoring approaches and proposing indicators to track AI adoption and vulnerabilities; includes a case study on AI third-party dependencies and service-provider concentration. Page summary gives no figures; the full PDF was not opened.
The Conference Board
72% of S&P 500 flag AI as a material risk, up from 12% in 2023; reputational risk cited by 38%; 45 companies cite AI implementation and adoption
Analysis of S&P 500 Form 10-K filings available through 15 August 2025. Sector counts rose: financials 14 to 63, health care 5 to 47, industrials 8 to 48 (2023 to 2025). Implementation and adoption failure is a named risk category.
Deloitte Australia
Partial refund of about A$98,000 on a A$440,000 fee after fabricated citations
Incident log entry (OECD.AI) on an AI-assisted report on the welfare compliance system containing fabricated citations and quotes attributed to GPT-4o hallucination; Deloitte said main findings were unchanged. Fee and refund amounts come from headlines in the OECD listing, not the incident body.
Lufthansa Group
About 4,000 jobs worldwide by 2030; adjusted EBIT margin target 8-10%; free cash flow above EUR2.5 billion a year
secondary: ch-aviation report of Lufthansa's capital markets targets. Cuts focus on administrative roles, mostly in Germany; company says 'the increased use of artificial intelligence (AI) will ensure greater efficiency in many areas.' Duplicated work also cited.
Harvard Business School (Otis, Clarke, Delecourt, Holtz, Koning)
640 entrepreneurs; no average effect on revenue or profit; high performers up to about +15%, low performers about -10%
Randomised trial of an AI business mentor on WhatsApp. Working paper first issued Dec 2023, revised Oct 2025, later published in Management Science. The authors attribute the split to which advice owners chose to implement. Source: research report of 14 Oct 2026; page confirmed by the research agent.
Chegg
Q3 2025 net revenue $77.7m vs $136.6m (-43%); subscription services $69.1m vs $119.8m (-42%); nine-month revenue -36%
SEC Form 10-Q for the quarter ended 30 Sep 2025 (date shown is period end; filing day not shown on fetched text). The filing excerpt gives the numbers but the AI cause is stated in company announcements, not in the excerpted tables.
University of Chicago Knowledge Lab (arXiv)
Synthetic DiD: +$89 weekly earnings in high-exposure occupations and +0.2pp unemployment (about zero); US CPS Jan 2010-Aug 2025
Preprint using exposure not observed adoption; not peer reviewed. Primary arXiv page opened.
AWS / Strand Partners (Vietnam)
18% of businesses adopt AI (13% prior year); 74% basic uses, 17% intermediate, 9% transformative; 55% cite skills shortage
Primary AWS release on press.aboutamazon.com for the 2025 'Unlocking Vietnam's AI Potential' study; 1,000 business leaders and 1,000 public; fieldwork dates not stated. Gives the year-on-year baseline for the 2026 study.
Salesforce (reported by NBC News)
CEO Benioff: support reduced from 9,000 to about 5,000 heads
secondary: NBC News quoting Benioff on The Logan Bartlett Show podcast and a Salesforce statement; vendor self-reported and the support-only framing is the reporter's. Publication date not shown on the page I opened.
Boston Consulting Group
5% of companies are 'future-built' and achieving AI value at scale; 60% achieve no material value (1,250 executives, 68 countries, 25+ sectors)
Self-reported survey by a consultancy. Date: PDF is titled October 2025 but its text says September 2025; I used 2025-09. Primary BCG PDF opened.
Boston Consulting Group
BCG's split of effort for winners: 70% people, organisation and processes; 20% technology; 10% algorithms; ~70% of AI value sits in core business functions (n=1,250)
BCG framing from its survey and client work; not an independently measured result. Primary PDF opened.
Boston Consulting Group
Future-built firms vs stagnating/emerging: 1.7x revenue growth and 1.6x EBIT margin (n=1,250); BCG expects 2x revenue uplift and 1.4x cost reduction vs laggards in AI-applied areas
Self-reported survey plus BCG projection; correlation not causation. Primary PDF opened.
McElheran, Yang, Kroff and Brynjolfsson (Rise of Industrial AI in America), AEA 2026 programme
One-SD more AI use: about 1.33% lower TFP initially (OLS); about 2% higher labour productivity growth 2017-2021 in younger plants; 23% of plants used AI by 2021, 8% intensively
Preliminary paper (September 2025) using Census MOPS, ASM and about 55,000 manufacturing firms. Older plants bear larger losses; roughly half the TFP loss at older plants is linked to loss of structured management practices. IV estimates imply much larger losses for marginal adopters.
JAMA Health Forum (reported by TechTarget)
60 of 950 devices; 182 recalls; 109 diagnostic/measurement errors; 91.8% of recalls from public companies
secondary: TechTarget on a research letter matching FDA-cleared AI devices to recall entries (Nov 2024 data). 77.7% of recalled devices from public companies lacked clinical validation vs 40% for private firms. 43.4% of recalls came within the first year, about double the rate for all 510(k) devices.
NACD 2025 Private Company Board Practices and Oversight Survey
24.6% adopted an AI governance framework; 16.9% audited where AI is used; 16.9% approved an AI budget (n=65)
Small sample (n=65 private-company boards). 61.5% set aside full-board agenda time for AI, but only 43.1% assessed AI risks and 13.8% created a director education plan. Shows attention outrunning structured oversight, relevant to founder-led firms with informal boards.
Bloomberg
CBA reversed plan to cut 45 customer service roles; said voice bot cut calls by 2,000 a week but union said volumes rose and staff needed overtime
secondary: Bloomberg page text via fetch; claims contested between bank and Finance Sector Union. Date taken from the URL.
Erasmus School of Economics and Chicago Booth (Henkel and Jabarian)
AI-interviewed candidates 12% more likely to receive an offer; 78% chose the AI interviewer when offered the choice
Natural field experiment at a large recruitment firm (unnamed): candidates randomly assigned to human recruiters, AI voice agents, or a choice. AI-interviewed candidates were also more likely to still be employed after one month; technical problems hit about 7% of AI interviews; the page does not say whether 12% is relative or percentage points. Working paper, not peer reviewed.
Erasmus School of Economics and Chicago Booth (Henkel and Jabarian)
Recruiters rated AI-interviewed applicants higher but weighted standardised test results more heavily
Same field experiment (about 70,000 candidates). A workforce-effect signal: automating the interview did not remove human judgement from the decision, and satisfaction and offer acceptance did not differ much between arms.
AWS / Strand Partners (Unlocking The Philippines' AI Potential)
250,000 businesses (21%); 64% of adopters report higher revenue averaging 18%; 66% expect 22% cost savings; 57% cite skills shortage
Vendor-sponsored survey of 1,000 business leaders; 26% of businesses feel their workforce is AI-ready. Cost savings are expectations.
MIT Sloan and Stanford GSB (Choi and Xie)
Monthly close cut by 7.5 days; AI-using accountants support 55% more clients per week; report detail up 12%
secondary: CFO Dive summary of a working paper using transactions from 79 small and medium companies (supplied with help from an accounting-AI software provider) plus a survey of 277 accountants. 8.5% of accountant time shifted from routine processing to higher-value work; 62% of accountants worry about errors. Observational field data, not randomised.
Gartner
Stock down 27.5% post-results; contract value visibility 'very low'; annual outlook cut
secondary: MarketScreener report of Gartner's Q2 2025 results. Article attributes pressure to AI substituting for research; no Gartner statement quoted.
US Government Accountability Office
Generative AI use cases rose from 32 (2023) to 282 (2024), about nine-fold; total AI use cases at the same agencies 571 to 1,110
GAO-25-107653 reviewed AI use case inventories at 11 of 12 selected agencies (Department of Defense exempt).
US Government Accountability Office
Officials at 10 of 12 agencies said existing federal policy, such as data privacy policy, could present obstacles to adoption
Other challenges named: sufficient technical resources and budget; keeping appropriate use policies up to date; four agencies said rapid technology evolution complicates policy.
METR
Tasks took ~19% longer with AI allowed; developers forecast a 24% speedup and afterwards believed 20% (16 developers, 246 tasks, Feb-Jun 2025)
Randomised controlled trial; small sample and mature open-source repos so may not generalise. arXiv v2 date used; primary paper opened.
MISA (Vietnam)
Customer-care productivity 1.71x; 350 staff handling work that previously needed 600
Stated by MISA's CEO in SGGP coverage, so company self-reported with no independent check. Same article reports FPT.AI helping firms cut costs 60% (vendor claim, no baseline).
Vietnam Ministry of Public Security (C06) with CMC Corporation
Nearly 100,000 user interactions in first week; about 98% of users rated the experience positively; negative feedback 1.8%
secondary: Viet Nam News report of a ministry announcement; satisfaction figures are self-reported by the ministry. Assistant integrated into the VNeID app.
Massachusetts Attorney General; Earnest Operations LLC
$2.5 million settlement
Alleged failure to test models for disparate impact, training on arbitrary human decisions, a 'Cohort Default Rate' variable that allegedly penalised Black and Hispanic applicants, and a 'Knockout Rule' denying applications by immigration status. Allegations from the AG's release.
MIT NANDA (Challapally Pease Raskar Chari)
Report says '95% of organizations are getting zero return' and 'just 5% of integrated AI pilots are extracting millions in value'; sample = 300+ public AI initiatives reviewed + interviews with 52 organisations + survey of 153 senior leaders at four industry conferences; research Jan-Jun 2025
Primary text read, but from a copy hosted by law firm Webber Wentzel; the MIT-hosted original was not opened. Labelled 'Preliminary Findings'. Success = deployment beyond pilot with measurable KPIs; ROI measured six months after pilot. Authors' own limits: figures 'directionally accurate based on individual interviews rather than official company reporting'; possible selection bias; six-month window may understate; causation not proven. Internal inconsistencies (66% vs ~67% buy share; ~50% vs ~70% of budget to sales and marketing). Sample is far smaller than the 150 interviews and 350-employee survey cited in secondary coverage (see P3-001). Self-reported and interview-based, not measured P&L.
MIT NANDA (Challapally Pease Raskar Chari)
60% of organisations evaluated enterprise-grade task tools, 20% reached pilot, 5% reached production; purchased partnerships ~66-67% of deployments vs ~33% internal builds; partnership pilots 'twice as likely' to reach full deployment; workers at 90%+ of firms use personal AI tools but only 40% of firms bought an official LLM subscription
Primary text (law-firm-hosted copy); self-reported interviews and survey. Authors state the partnership-success link 'does not necessarily prove causation'. The ~22% internal-build success rate in P3-002 is not in the primary; the primary gives a 2x relative figure and a deployment-share split instead.
MIT NANDA (Challapally Pease Raskar Chari)
Back office: BPO elimination $2-10M a year in customer service and document processing; creative and content agency spend down 30%; risk checks saved $1M a year; front office: lead qualification 40% faster, customer retention +10%; gains reported without material workforce reduction; roughly 50% of GenAI budget goes to sales and marketing (survey text says ~70%)
Primary text (law-firm-hosted copy). Anonymised, interview-based and self-reported; no company named. Budget share is inconsistent within the report (~50% vs ~70%).
Graduate Management Admission Council
Over 1,100 recruiters and hiring managers; fielded January to March 2025
Abstract page gives no percentages; reports employers value knowledge of how to use AI tools more than last year and larger US MBA starting salaries in 2025. Management-education signal.
Gartner
45% vs 20% keep AI projects operational for 3+ years; 91% of high-maturity firms have dedicated AI leaders
Q4 2024 survey of 432 organisations in six countries. Outcome is project longevity, not profit. Self-published by an advisory firm. Source: research report of 14 Oct 2026.
Thomson Reuters
22% visible AI strategy; those firms twice as likely to see AI-driven revenue growth; 3.5x more likely to see critical benefits
2025 Future of Professionals: 2,275 legal, tax, accounting, risk and compliance professionals, surveyed Feb-Mar 2025. Thomson Reuters sells AI products to these professions, so treat as vendor research.
Thomson Reuters
53% see ROI already; 5 hours a week expected savings within a year (4 hours predicted in 2024); valued at about $19,000 per person a year
Vendor research, same survey (n=2,275). Savings are predictions, not measured. The report also estimates about $32bn combined annual US impact for legal (about $20bn) and CPA (about $12bn) sectors.
Boston Consulting Group
72% regular use overall; 51% of frontline employees; 36% feel adequately trained; n over 10,600 workers in 11 countries
BCG press release for AI at Work 2025.
Boston Consulting Group
Five or more training hours (especially in person with coaching) make regular use significantly likelier; 25% of frontline workers say leaders provide enough AI guidance
BCG press release; the page does not give the exact uplift for the 5-hour threshold.
Gartner
Gartner predicts over 40% of agentic AI projects will be cancelled by end 2027 (costs, unclear value, weak risk controls); poll of 3,412 webinar attendees Jan 2025: 19% significant investment, 42% conservative, 8% none, 31% wait-and-see
Forecast, not an observed outcome; the poll is of webinar attendees (self-selected). Primary Gartner press release opened.
Gartner
Gartner estimates only ~130 of the thousands of agentic AI vendors are real (the rest 'agent washing')
Analyst estimate with no published method in the release; primary Gartner press release opened.
VietinBank (Vietnam)
350,000+ questions answered in 2 months; staff time searching for information reportedly cut 95%
Press-reported (Bao Dau tu) with no stated baseline or method; treat as self-reported by the bank. Genie trained on 2,000+ internal documents. No cost, revenue or headcount figures for any named bank in the article.
UK Government (gov.uk)
Planning records digitised in just three minutes each versus 1-2 hours manually; estimated saving of £527m for the public sector each year
Trial councils: Hillingdon, Nuneaton & Bedworth, Exeter. £527m is an estimate in a government announcement, not a measured outcome.
High Court of Justice (King's Bench Division), Ayinde v Haringey and Al-Haroun v Qatar National Bank
Five fake cases in one matter; 18 of 45 cited authorities fictitious in the other; wasted costs of GBP 2,000 each in Ayinde
Divisional Court judgment [2025] EWHC 1383 (Admin) warned that generative AI cannot do reliable legal research and referred lawyers to the SRA and Bar Standards Board. Duty to supervise delegated work stressed. Summary read from the judgment PDF; the court page itself showed only the case header.
Insilico Medicine (Nature Medicine, reported by Drug Discovery Trends)
71 IPF patients, 12 weeks; FVC +98.4 mL vs -20.3 mL placebo; 7 liver-related discontinuations
secondary: AI-identified target (TNIK) and AI-designed molecule. Small randomised Phase IIa; 12 of 18 completed the high dose vs 88% on placebo. 10 of 22 nominated candidates reached clinic. Earlier analyses put AI-origin Phase II success near 40%. Phase IIb/III still needed.
UK Government Digital Service trial (reported by PublicTechnology)
20,000 employees in 12 departments, Q4 2024: average 26 minutes saved a day (about 13 days a year, an extrapolation); 17% saved nothing, 23% over 30 minutes; project delivery 35 min vs tax 11 min
secondary: trade article on the report; self-reported estimates (article does not label them so); no output measure. 85% 'good value' is belief not measurement.
Harvard University, Scientific Reports
Median post-test 4.5 vs 3.5; median gain more than double; effect size 0.63; median 49 minutes on tutor vs 60 in class
Randomised crossover with 194 eligible students in one Harvard physics course; the tutor was purpose-built with pedagogy prompts. Single course, elite students, short-term test.
PwC
AI-skilled jobs paid an average premium of 56% in 2024, up from 25% the year before; skill demands changing 66% faster in AI-exposed occupations
PwC analysis of close to one billion job ads across six continents; vendor-reported (consultancy).
PwC
Productivity growth in most AI-exposed industries 7% (2018-2022) to 27% (2018-2024); least exposed 10% to 9%; jobs in AI-exposed occupations grew 38% 2019-2024 vs 65% in less exposed
vendor-reported: PwC analysis; correlational.
NFIB Small Business and Technology Survey
24% of small employers use AI; 21% at single-digit headcount vs 48% at 50+ employees
NFIB member survey fielded 6 to 31 March 2025 (521 completed responses reported on NFIB's analysis page). 63% expect AI to matter in their industry within five years; 23% do not know.
NFIB Small Business and Technology Survey
98% of AI-using employers report no change in employee numbers; communications 29%, marketing 27%, customer service 9%, accounting 4%
Planned or current AI uses among NFIB member respondents, March 2025 fieldwork. Owner-reported.
The Permanente Medical Group (Kaiser Permanente), NEJM Catalyst via AMA
15,791 hours saved (~1,794 eight-hour days); 7,260 physicians; 2.58M encounters; 84% say better patient communication
AMA write-up of a TPMG internal evaluation (Oct 2023 to Dec 2024). Time savings appear to come from system metrics comparing users with non-users, but the comparison is observational; satisfaction figures are survey self-report. Top-third users account for 89% of activations. Authors are TPMG staff; vendor not named.
Duolingo
Plan to gradually stop using contractors for work AI can handle; 10% of contractor workforce already cut by January 2024
CEO Luis von Ahn later said he 'didn't do that well' communicating the plan, that AI is not replacing employees and hiring continues, and accepted 'occasional small hits on quality'. secondary: TechRepublic report dated 26 May 2025; no financial or user metrics on the page.
World Bank
Overall assessment +0.31 SD, English +0.23 SD; about 1.5 to 2 years of typical schooling gain
Policy Research Working Paper 11125: randomised trial of Copilot-based after-school English tutoring for first-year senior secondary students; page gives no sample size. Authors describe it as highly cost-effective.
International Labour Organization and NASK
About one in four workers in jobs with some GenAI exposure; 3.3% of global employment in highest exposure category; exposure ~11% of employment in low-income vs 34% in high-income countries
ILO Working Paper 140; occupational exposure index, not measured displacement. Concludes transformation is the most likely impact.
International Labour Organization and NASK
Highest exposure: 4.7% of female vs 2.4% of male employment globally; high-income countries 9.6% of female vs 3.5% of male employment
ILO Working Paper 140; exposure index only.
UK Department for Science, Innovation and Technology
F1 score 0.76 on over 2,000 responses across six questions; projected 75,000 days of analysis saved per year across 500 consultations
Government release; the 75,000-day and £20 million per year figures are projections, not measured results. Officials also reviewed every response manually for comparison. Tool still in trial.
Klarna (reported by Entrepreneur)
Company said AI did the work of 700 agents and handled 75% of chats (~2.3M conversations in first month) across 35+ languages (Feb 2024 claim)
secondary: reported by Entrepreneur citing Klarna's own 2024 statements; vendor self-reported not audited.
Klarna (reported by Entrepreneur and Maginative)
CEO said AI service was cheaper but 'lower quality'; hiring humans again; headcount fell 22% to ~3,500 over a year of hiring freeze
secondary: reported by Entrepreneur (2025-05-09) and Maginative (2025-05-08); based on CEO statements to Bloomberg. No satisfaction or cost figures published in what I opened.
IBM Institute for Business Value
25% of AI initiatives delivered expected ROI over the last few years (2,000 CEOs, 33 countries, 24 industries; fielded Feb-Apr 2025)
Self-reported CEO survey commissioned by a vendor with an AI business. Primary IBM newsroom page opened.
IBM Institute for Business Value
16% of AI initiatives scaled enterprise-wide; 50% of CEOs say rapid investment left disconnected, piecemeal technology; 64% invest ahead of understanding value for fear of falling behind (n=2,000)
Self-reported CEO survey by an AI vendor. Primary page opened.
Recursion Pharmaceuticals (reported by Pharmaceutical Technology)
5 programmes ended or paused; cash $509M; runway into mid-2027; shares -13.4% in a day
secondary: three clinical terminations, one clinical pause and one preclinical discontinuation, leaving six active programmes. 2025 cash burn target under $450M. Shows that platform-based AI discovery still faces conventional clinical attrition.
IBM Institute for Business Value
About one-third of workforce needs retraining or reskilling within three years; 54% hiring for AI roles that did not exist a year ago; n=2,000 CEOs
secondary: IBM IBV report reported by HR Dive (article dated 2025-05-09); IBM primary page not opened.
IBM Institute for Business Value
50% say recent rapid investments left technology and expertise out of sync; 61% adopting agents; 65% plan to use automation to close skill gaps; nearly 7 in 10 say success depends on many leaders who can make critical decisions
secondary: IBM IBV findings via HR Dive; n=2,000 CEOs worldwide.
University of Chicago BFI
No significant effect on earnings or recorded hours; CIs rule out effects above 1% (pooled); users report ~2.8% time saved (~25,000 responses from ~7,000 workplaces; 11 occupations; admin-registry linked)
Study linking surveys to administrative payroll data; time savings are self-reported. 30% of workers had chatbot training. Primary working paper opened.
Dillon Jaffe Immorlica Stanton (NBER 33795)
7,137 knowledge workers at 66 firms, Sep 2023-Oct 2024: email time -1.29 hours a week (ITT) and -3.56 hours (about 31%) for regular users on 11.65 hour baseline; meeting time no significant change; completed document counts unchanged; collaborative document time down significantly
Measured (telemetry plus randomised access); primary PDF opened. Early-rollout licences, firm-selected workers; time reallocated, not output or P&L.
Microsoft
46% of leaders say their organisation uses agents to fully automate workstreams; 67% of leaders vs 40% of employees familiar with agents; 80% of workforce lack time or energy
vendor-reported: Microsoft Work Trend Index 2025. Sample size not stated on the blog page; date from URL path.
Microsoft
78% of leaders considering hiring for new AI roles; 33% considering headcount reductions; 53% of leaders say productivity must increase
vendor-reported: Microsoft Work Trend Index 2025. Date from URL path.
AIA Group
Claim handling from more than two days to under 25 minutes; OCR 97% accuracy on claim documents
secondary: Reinsurance Asia trade article; company-reported. The article itself says specific accuracy figures are hard to determine (AI assessment reportedly about 99.7% on simpler claims).
Manulife
GenAI capabilities available to 100% of global workforce, 75% engagement rate
secondary: Reinsurance Asia trade article; company-reported. ChatMFC launched 2024; no accuracy or savings figures given for the underwriting tools.
SEC v. Alberto Saniger Mantinan (Nate, Inc.)
Over $42 million raised from stock sales
SEC complaint (filed 9 Apr 2025, litigation release 11 Apr 2025) alleges Nate told Seed and Series A investors purchases were completed by AI without human involvement while contract employees largely entered orders manually. Allegations, not findings.
International Energy Agency (Energy and AI)
AI fault detection could cut outage durations 30-50%; AI grid tools could unlock up to 175 GW of transmission capacity; energy-related emissions about 5% lower in 2035
IEA modelled potential from broad use of existing AI solutions, not measured utility results. The fetched page gave no explicit publication date; April 2025 is inferred from the report citation year and an S&P Global article dated 10 April 2025 in search results.
International Energy Agency (Energy and AI)
415 TWh in 2024 (1.5% of global electricity) to about 945 TWh in 2030; growth about 12% a year since 2017; US 45% of 2024 use
Demand side of the energy and AI picture, relevant to utilities and to power-intensive buyers. The fetched page gave no explicit publication date; April 2025 inferred (see other IEA row).
Dell'Acqua et al. (NBER w33641) with Procter & Gamble
776 professionals on real product-innovation problems; individual with AI matched a team without AI (solution quality, not financials)
Field experiment inside an upskilling programme. Disclosed P&G funding ties. Outcome is solution quality, not profit. Source: research report of 14 Oct 2026.
UK Public Accounts Committee (reported by PublicTechnology)
About 28% of all IT systems in use across government are currently considered legacy
secondary: trade-press report of the PAC findings on government AI ambitions; the page does not state the report date (month taken from the URL). Also repeats the NAO finding that 70% of entities had difficulty recruiting AI-skilled staff.
McKinsey & Company
150 leaders: 85% exploring/adopted; 64% of implementers anticipate or quantify positive ROI; 15% not started
Survey of 60 payer, 60 health system and 30 services/technology leaders, fielded in one week of Dec 2024. 'Anticipated' ROI is mixed with quantified ROI, so it overstates realised benefit. Consultancy-run, self-reported.
Peterson Health Technology Institute (reported by Fierce Healthcare)
Mass General Brigham 40% relative burnout reduction (n=223, 6-week pilot); no increase in patients seen per period
secondary: PHTI compiled existing evidence with eight health systems and 10 vendors. Conclusion: scribes reduce burnout and cognitive load but have not proven a financial return. Pajama-time results mixed; MultiCare and Ochsner reported higher billing, others no impact. The Suki claim of $54,000 extra revenue per user per year is a vendor claim cited in the report. Standard metrics are lacking and many systems only ran pilots.
Commonwealth of Pennsylvania (Governor's Office of Administration)
175 employees in 14 agencies; average 95 minutes saved per day; over 85% somewhat or very positive experience; 48% had never used ChatGPT before
self-reported: state administration press release on its own year-long pilot; time savings are participant-reported, no control group stated.
Unilever
About 50% lower imagery cost and about half the production time
Company-reported at Nvidia GTC on 18 Mar 2025, via Campaign. Pilot covered TreSemme, Dove, Vaseline and Clear; image volume not stated.
Unilever
Content production costs -87%; brand lift 23%; content generated twice as fast
Company-reported single-market result (TreSemme Thailand), not independently verified; brand-lift method not described. Source: Campaign report of Unilever's GTC remarks.
Epoch AI
Price for GPT-4-level GPQA fell about 40x per year; range 9x-900x per year
Analysis of price per million tokens to reach fixed benchmark scores. For example an MMLU score of 86 cost $37.50 per million tokens with GPT-4 (March 2023) versus $0.18 with Gemini 2.0 Flash (Feb 2025). Authors say it is unclear the fastest declines will persist. Price per capability falls; spend can still rise.
Deloitte
74% of pioneers vs 44% of followers expect ROI above 10%; 47% vs 17% say ROI exceeds expectations; followers rate themselves ready on talent 7%, risk and governance 16%
Financial-services subset of 542 leaders from a 2,773-leader, 14-country survey fielded July to September 2024. Self-assessed 'pioneer' status; consultancy survey.
Australian National Audit Office
74% of AI models in production did not have completed data ethics assessments (August 2024); ATO had 43 in-house AI models in production at 14 May 2024
Auditor-General Report No. 26 2024-25 on AI governance at the Australian Taxation Office.
Australian National Audit Office
Overall conclusion: 'The ATO has partly effective arrangements in place to support its adoption of AI'; no evidence of structured, regular monitoring of in-house AI models in production; seven recommendations, all agreed
Same audit report; the strategy was not reported on regularly between October 2022 and January 2024 and no evaluation approach had been set.
Allianz
Insurance Copilot launched 2024 for motor claims in Austria; Allianz says only 'preliminary assessments' show 'significant business impact' with no figures
Company article; useful as an example of a scaled deployment with no published P&L metric. Primary Allianz page opened.
DBS Bank (Business Pulse Check Survey)
73% plan to invest in Gen AI; 32% experimenting; 72% plan to train staff
Bank-run survey of more than 110 companies in early 2025; small sample and a lender with an interest in SME tech lending.
Deloitte Global Health Care Executive Outlook (reported by Becker's)
121 executives: 10% significant ROI, 33% moderate, 12% low, 37% too early to measure
secondary: Becker's Hospital Review on the Deloitte report released 29 Jan 2025. About 43% reported moderate or significant return. Earlier snapshot than Deloitte's 2026 life-sciences survey already in the library.
Presto Automation Inc.
Prospectus supplement of 17 Nov 2023: over 70% of Presto Voice orders required human agent intervention
SEC order: from Nov 2021 to May 2023 Presto made false and misleading statements about its AI drive-thru product, did not disclose that third-party speech recognition powered all early units, and misreported orders completed without human intervention. No civil penalty given cooperation; consented without admitting or denying. SEC order shows humans entered orders about 70% of the time in the advanced pilot and nearly all orders at most other locations.
Salesforce (Small and Medium Business Trends Report, Singapore cut)
94% using or experimenting; 87% of adopters report higher revenue; 72% find it hard to keep up
secondary: CFOtech Asia reporting Salesforce research. Vendor-reported; Singapore sample only about 50 respondents of 3,350.
Ni Wang Feng Lu Wang Zhou (Alibaba collaboration; SSRN)
Randomised agent-level rollout of an AI diagnosis and solution-suggestion assistant: speed and quality improved, engagement rose, low performers gained more, high performers' service quality declined
Measured randomised field experiment; SSRN abstract opened, which gives no effect sizes or sample size. Pointer for the 'raises the floor, may lower the ceiling' pattern.
World Economic Forum
About 170 million jobs created (14% of current employment), 92 million displaced, net 78 million; 39% of key skills expected to change by 2030 (44% in 2023 edition)
WEF article summarising the Future of Jobs Report 2025; employer survey based projections, not observed outcomes.
Microsoft Research (Dillon Jaffe Peng Cambon)
Half an hour less reading email per week; documents completed 12% faster; nearly 40% of those given access used it regularly
vendor-reported: authors are Microsoft employees studying Microsoft product. Date from page; arXiv listing 2025-04-15.
Salesforce (6th Small and Medium Business Trends Report, India cut)
78% using or experimenting; 93% of users report revenue increase; 41% fear being left behind
secondary: Channel Drive reporting Salesforce research. Vendor-reported; India sample about 200 of 3,350 leaders.
MIT Center for Information Systems Research (Weill, Woerner, Sebastian)
Stages 1-2 below industry average financial performance; stages 3-4 well above; biggest gain moving from stage 2 to stage 3
Cross-sectional and correlational. Based on the 2022 Future Ready Survey (N=721) plus 2024 interviews, so it predates generative AI. CISR is funded by a corporate member consortium. Source: research report of 14 Oct 2026.
Salesforce (6th Small and Medium Business Trends Report)
91% of AI-enabled SMBs report increased earnings; 83% of growing SMBs use AI vs 55% of those with declining revenue
secondary: SmallBizTrends reporting Salesforce research. Vendor-reported: 3,350 SMB leaders across 26 countries, fielded 3 Aug to 16 Sep 2024. Correlation, not causation; article promotes Salesforce Agentforce.
US Census Bureau (Dinlersoz and Goldschlag)
Firms with 1-4 employees: AI use 4.6% to 5.8%; 250+ employees: 5.2% to 7.8% (26 biweekly observations); at current trends 25% adoption takes about 6 years for 250+ firms and about 15 years for 1-4 employee firms
Measured official survey (BTOS), primary page opened. Broad AI definition incl. generative AI; pre-Nov 2025 question wording so not comparable with 2026 figures; slope differences between size classes not tested for significance by the authors.
Bank of England; FCA
75% of firms use AI (58% in 2022); one third of use cases are third-party implementations; 2% of use cases fully autonomous
Third survey of 118 firms. Median use cases per firm expected to rise from 9 to 21 in three years; foundation models are 17% of use cases; top three providers account for 73% of named cloud providers.
US FinCEN
FIN-2024-Alert004. Says FinCEN saw rising suspicious activity reports describing suspected deepfake use from 2023; criminals combine GenAI images with stolen data to build synthetic identities that launder proceeds. No SAR counts given in the alert.
Andreessen Horowitz (Guido Appenzeller)
10x per year for equivalent performance; GPT-3-class cost from $60 to $0.06 per million tokens in 3 years
Investor essay. Prices for a fixed MMLU level of 42 fell 1,000x in three years; for level 83 about 62x since March 2023. Author calls the methodology 'far from perfect' and notes frontier models (o1 at $60 per million output tokens) still cost as much as GPT-3 did at launch.
Försäkringskassan / Inspectorate for Social Security (reported by Amnesty International)
ISF concluded the algorithm 'in its current design does not meet equal treatment'; investigation found it disproportionately flagged women and people with foreign backgrounds
secondary: advocacy-organisation report (Amnesty) citing Lighthouse Reports and Svenska Dagbladet; the page does not state its date (month inferred from URL) and gives no counts. The agency disputed the ISF analysis. Machine learning risk scoring since at least 2013.
US Department of the Treasury (reported by Nextgov)
Treasury 'prevented and recovered over $4 billion' in fraudulent and improper payments in FY2024 vs $652.7 million prior year; machine learning on check fraud resulted in $1 billion in recovery
secondary: Nextgov report of a Treasury announcement; self-reported by Treasury. The article notes sub-totals do not reconcile with the $4 billion headline, so treat attribution to AI as unverified.
US Senate Permanent Subcommittee on Investigations (reported by Becker's Hospital Review)
UnitedHealthcare post-acute denial rate 10.9% (2020) to 22.7% (2022); 46M prior-auth requests, 7.4% denied in 2022
secondary: report on UnitedHealthcare, Humana and CVS covering 2019 to 2022. AI relevance is indirect: the subcommittee recommended CMS rules so predictive technologies do not override human review. The insurers disputed the characterisation. Government investigation, not an AI audit.
International Monetary Fund
Chapter 3 of the October 2024 GFSR. Says labour-market and patent data point to a significant rise in AI use in capital markets; flags faster markets, higher volatility under stress, reliance on a few third-party AI providers and cyber risk. No adoption percentages on the landing page.
Australian Digital Transformation Agency
Only a third of participants used Copilot daily; 64% of managers perceived efficiency and quality gains in their teams; 40% could redirect time to higher-value work; savings 'up to an hour' on some tasks
Six-month whole-of-government trial evaluation. Date taken from the report file path (October 2024); the page does not state agency or participant counts. Concerns listed include entry-level job effects, bias and vendor lock-in.
Cruise (US NHTSA consent order)
Total $1.5 million monetary penalty; corrective action plan required
secondary: law-firm blog on the NHTSA consent order. October 2023 incident: vehicle 'operating autonomously, dragged a pedestrian several feet before coming to a stop'. Transport sector; no matching industry id.
FTC (DoNotPay)
$193,000 proposed settlement payment
FTC alleged DoNotPay marketed itself as the world's first robot lawyer without testing output against a human lawyer's and without employing attorneys. Proposed order bars unsupported claims that it can replace professional services.
FTC (Ascend Ecom)
At least $25 million in consumer losses alleged
FTC alleged the company falsely claimed cutting edge AI tools would generate passive income from online storefronts; court halted the scheme and appointed a receiver. Allegations.
FTC (Ecommerce Empire Builders; FBA Machine / Passive Scaling)
FBA Machine: more than $15.9 million consumer losses alleged; EEB: up to $35,000 taken from storefront buyers
FTC alleged both sold AI-powered ecommerce income schemes with unsupported income promises ($10,000 monthly claims at EEB; guaranteed income and risk-free claims at FBA Machine). Courts halted schemes and appointed receivers. Allegations.
FTC (Rytr)
Some subscribers allegedly produced hundreds and sometimes tens of thousands of reviews
FTC alleged Rytr's testimonial and review generation tool produced false reviews at scale. Proposed order bars marketing any service dedicated to generating reviews or testimonials. Commission vote 3-2; no amount stated.
Microsoft, Accenture and a Fortune 100 firm (Cui Demirer Jaffe Musolff Peng Salz)
26.08% increase (SE 10.3%) in completed tasks; less experienced developers had higher adoption and larger gains
RCTs across three firms; each experiment noisy and results vary. SSRN record last revised 2025-08-21.
Gartner (reported by MarketScale)
58% of finance functions used AI in 2024, up 21 points on 2023; 42% not using, half of them planning to; 66% more optimistic than a year earlier (n=121 finance leaders)
secondary and dated: MarketScale piece of 2026-09-13 re-reporting a 2024 survey (date taken from CFO Dive per the article). The coverage cites no savings or cycle-time figures, so the finance-close gap is NOT closed.
Bick Blandin Deming (NBER w32966)
Nearly 40% of adults 18-64 use GenAI; 23% of employed used it for work in prior week; 9% every workday; 1.4% of total work hours saved
Nationally representative US survey, late 2024 data per abstract; self-reported time savings.
University of Pennsylvania / Wharton (Bastani et al.)
Practice: GPT Base +48%, GPT Tutor +127% vs control; exam: GPT Base 17% worse than control, GPT Tutor about the same as control
secondary: Wharton Knowledge write-up of the study; nearly 1,000 high school maths students in Turkey. Guardrailed version used hints rather than answers.
Walmart
More than 850 million product data points created or improved; executives say it would have needed nearly 100 times the headcount
Company-reported on the fiscal Q2 earnings call by CEO Doug McMillon and Walmart U.S. CEO John Furner (Modern Retail). Output volume, not a P&L result; no cost or conversion figure disclosed.
RAND Corporation (Ryseff, De Bruhl, Newberry)
65 experienced data scientists and engineers interviewed; 84% cited leadership-driven root causes as primary
Qualitative study. The widely quoted "more than 80% of AI projects fail" is RAND quoting outside work, not its own measurement. Pre-agent era. Source: research report of 14 Oct 2026.
OCBC Bank (Singapore)
Task completion times reduced by up to 50% in a 1,000-employee trial; over 1 million prompts answered by May 2024
secondary: OpenGov Asia interview. Self-reported by trial participants, not measured output. Launched November 2023.
SEC v. Ilit Raz (Joonko)
Investors allegedly defrauded of at least $21 million
SEC alleges the AI recruiting startup's CEO falsely claimed over 100 customers, over $1 million revenue and over 100,000 active candidates, using fabricated testimonials and falsified bank statements. SEC enforcement director described it as old school fraud using new school buzzwords. Allegations, not findings.
Mastercard
Detection rate of compromised cards doubled; false positives reduced 'by up to 200%'; merchant risk identification up 300%
vendor-reported: Mastercard press release. Baselines for the percentages are not defined in the release.
BCG, Drug Discovery Today (reported by Health Tech World)
Phase I success 80-90%; Phase II ~40% (limited sample); end-to-end 5-10% to 9-18% projected
secondary: trade report of a BCG-authored paper. Page does not state the number of molecules analysed. Phase II success is in line with industry averages, so the Phase I advantage may not carry to efficacy. Early, small sample.
Microsoft and LinkedIn
75% of knowledge workers use AI at work; 78% of AI users bring their own AI tools to work
vendor-reported: Microsoft/LinkedIn primary report. Date taken from the blog URL path; page text carries no date.
Microsoft and LinkedIn
39% of AI users received AI training from their employer; 60% of leaders worry company lacks vision and plan for AI; two-thirds of leaders would not hire someone without AI skills
vendor-reported: Microsoft/LinkedIn Work Trend Index 2024. Date taken from URL path.
Stanford RegLab and Stanford HAI
Lexis+ AI and Ask Practical Law AI incorrect over 17% of the time; Westlaw AI-Assisted Research over 34%; general chatbots 58% to 82% in earlier work
Over 200 pre-registered legal queries. Retrieval-augmented tools hallucinate less than general models but not rarely; vendors had published no rigorous evaluations. Date from the Stanford HAI release (May 2024); the fetched page itself did not print the day.
US SEC
$400,000 combined civil penalties ($225,000 Delphia; $175,000 Global Predictions)
SEC press release 2024-36. Delphia (2019-2023) allegedly made false statements about using AI and machine learning with client data and the SEC found it lacked the capabilities claimed; Global Predictions allegedly called itself the 'first regulated AI financial advisor'. Both settled without admitting or denying findings.
SEC (Delphia; Global Predictions)
$225,000 penalty (Delphia) and $175,000 (Global Predictions); $400,000 total
First SEC AI-washing settlements. Delphia (2019-2023) claimed its AI used client data to predict winning companies and trends, which it lacked the capability to do; Global Predictions called itself the first regulated AI financial advisor. Both settled without admitting or denying.
UK National Audit Office
37% of the 87 government bodies that responded had deployed AI; 74 use cases deployed; most bodies with deployed AI had only one or two
NAO survey of government bodies in 'Use of artificial intelligence in government' (HC 612). Report also states 70% were piloting or planning AI, with a median of four use cases explored per body.
UK National Audit Office
70% cite difficulty recruiting or retaining staff with AI skills; 21% had an AI strategy (a further 61% planned one); 30% had risk and quality assurance processes explicitly covering AI risks
Barriers and governance findings from the same NAO survey of 87 government bodies.
European Central Bank (reported by Thomson Reuters Institute)
More than 40 GenAI use cases in banking supervision; 14 applications used by more than 3,500 people
secondary: Thomson Reuters Institute summary of remarks by ECB Supervisory Board member Elizabeth McCaul. Shows how a regulator itself is adopting AI.
The Permanente Medical Group, NEJM Catalyst (reported by Epocrates)
3,442 physicians; 303,266 encounters in about 10 weeks; two notable hallucinations reported
secondary: Epocrates summary of the Tierney et al. NEJM Catalyst study. Primary care users showed larger drops in after-hours EHR time and in-visit note time than non-users (p<0.05), with high users gaining most. The hallucinations: a prostate exam recorded as done when only scheduled, and a diagnosis never made. Not randomised.
Viettel Telecom (Vietnam)
~33 million customer calls handled; ~21 billion VND saved per year
Reported in Thoi bao Tai chinh Viet Nam as a Viettel result (self-reported, no method or baseline). Dated Feb 2024, so older than the 2026 evidence.
Viettel / Supreme People's Court (Vietnam)
~30% saving in case-handling time; 12,000+ assistant accounts issued (2023 report)
Vendor/press-reported, no baseline or method. A July 2025 SGGP article separately says a Viettel GenAI legal assistant on Llama 3.3 cut legal staff working time by about 30%. Account figure from a Sept 2023 VnEconomy piece.
Air Canada / BC Civil Resolution Tribunal
C$812 awarded (fare difference on C$1,630.36 paid)
Tribunal rejected Air Canada's argument that the chatbot was a separate legal entity and held the airline responsible for all information on its site. Small sum, but a clear liability precedent. Date given as year-month because the page says only that the decision came out 'this week'; secondary reporting by CBC.
UK Government Digital Service
1,000 users invited to live private pilot; nearly 70% found responses useful; just under 65% satisfied (157 survey respondents); answers did not reach the highest accuracy level demanded for GOV.UK
GDS blog on its first generative AI experiment; reported a few cases of hallucination, mostly on ambiguous or inappropriate queries.
Gartner (reported by MH&L)
About 50% planning GenAI within 12 months, 14% already implementing; 5.8% of function budget allocated; 65% to hire dedicated staff
secondary: MH&L report of a Gartner survey; sample size not stated. Useful as a 2024 baseline against later measured-outcome evidence.
Humana; Optum/naviHealth (reported by Becker's Payer)
Alleged 1% target band around algorithm estimates; alleged 90% of appealed decisions overturned
secondary: Becker's report of a class-action complaint filed 12 Dec 2023 in W.D. Kentucky. These are allegations, not findings. Optum says the tool is not used to make coverage decisions; Humana says adverse decisions are made by physician medical directors. Illustrates legal and reputational exposure when a prediction tool anchors claim decisions.
Harvard Business School (Otis Clarke Delecourt Holtz Koning)
No average treatment effect detected; high performers up just over 15%, low performers about 8% worse; gap 0.27 SD
HBS Working Paper 24-042; small-business entrepreneurs in Kenya, field experiment via WhatsApp.
University of Minnesota and University of Michigan law schools (Choi, Monahan, Schwarcz)
Large, consistent time reductions; small, inconsistent quality gain; lowest-skilled gained most where quality improved
Randomised trial with law students on realistic tasks, blind-graded; participants reported higher satisfaction. Students are not practising lawyers. Forthcoming in Minnesota Law Review (2024).
Harvard Business School, Wharton, Warwick, MIT Sloan and BCG
758 consultants; 12.2% more tasks, 25.1% faster, quality over 40% higher inside the frontier
Pre-registered field experiment at Boston Consulting Group (about 7% of its individual-contributor consultants) with random assignment to no AI, GPT-4, or GPT-4 plus prompt guidance. Working Paper 24-013. Lower-skilled consultants gained about 43%, higher-skilled about 17%.
Harvard Business School, Wharton, Warwick, MIT Sloan and BCG
19 percentage points less likely to produce correct solutions on a task built so GPT-4 would err
Same BCG experiment (Working Paper 24-013). On a task designed to sit outside AI capability, consultants using GPT-4 were less often correct than the no-AI group, which is the core risk of over-reliance.
Gartner (HR practice)
5% implemented generative AI; 9% piloting; 84% expect productivity gains
Dated baseline: Gartner HR survey (samples of 105-177 HR leaders depending on question), as reported by FutureCIO; the article does not state the year of fieldwork and the date is taken from page metadata. Useful only as a starting point against which later adoption can be compared.
Royal Commission into the Robodebt Scheme (Australia)
Scheme 'neither produced accurate results' and did not comply with the income calculation provisions of the Social Security Act 1991; kept going until November 2019
secondary: Australian Policy Online catalogue entry for the report; the page does not state debt counts or amounts. Report first published 7 July 2023, updated 11 July 2023.
MIT (Noy and Zhang)
453 college-educated professionals; time down 40%; output quality up 18%; 2x as likely to use ChatGPT in real job two weeks later
Online experiment on writing tasks; abstract figures via Stanford SCALE listing of the Science paper.
US Consumer Financial Protection Bureau
98 million users (about 37% of US population) engaged with a bank chatbot in 2022; all top 10 US commercial banks use chatbots
CFPB issue spotlight. Cites industry estimates of about $8 billion annual savings (about $0.70 per interaction) and one survey where 80% of consumers left more frustrated; warns of 'doom loops', inaccurate information and legal liability. Cost and survey figures are third-party numbers quoted by the CFPB.
Panasonic Connect (Japan)
About 12,500 employees with access; 55,380 questions in first month; legal review cut from about 1 hour to under 10 minutes (employee report)
Published by Microsoft, a vendor with an interest. Rolled out 17 Feb 2023; Microsoft's article itself notes productivity impact is so far anecdotal.
Walmart (with Pactum)
About 3% savings on contracts; three in four suppliers prefer negotiating with the AI
secondary: Engadget relaying Walmart statements to Bloomberg (26 April 2023). Dated but still the best-known procurement deployment; vendor and customer self-reported, deal count and payment terms not stated, and the 3% figure is not explicitly attributed on the page.
NBER (Stanford and MIT authors)
~14% more issues resolved per hour on average and ~34% for novice or low-skilled agents; 5,179 customer support agents, staggered rollout
Field study of a real deployment at one Fortune 500 software firm. Older than 2025 but the most-cited measured before/after; NBER page opened (revised Nov 2023).
OECD (Broecke)
secondary: abstract only, no figures retrieved. Surveys of employers and workers in manufacturing and finance across seven countries; training and consultation linked to better worker outcomes.
Australian Government (Services Australia / Department of Social Services)
Unlawful debts issued to 416,000 Australians; at least 381,000 people repaid $751 million; debts worth $1.76bn cancelled; scheme ran July 2015 to November 2019
Automated income-averaging debt raising. Figures from a Minister for Social Services statement; page credits the royal commission with nine weeks of hearings and more than 100 witnesses.
Netherlands Court of Audit (Algemene Rekenkamer)
3 of 9 audited algorithms met all basic requirements, 6 did not; 6 of 9 organisations did not know which staff had access to their algorithm and data
Audit using the Court's framework; issues included bias checks, outsourcing responsibilities and IT general controls. Police Criminality Anticipation System did not check for bias.
Michigan Medicine (JAMA Internal Medicine, reported by Fierce Healthcare)
AUC 0.63 vs 0.76-0.83 claimed; sensitivity 33%; PPV 12%; alerts on 18% of patients
secondary: external validation on about 38,500 hospitalizations (2,552 with sepsis), Dec 2018 to Oct 2019. The vendor model produced alerts on 18% of all hospitalised patients while missing 67% of sepsis patients. Often cited as the case for local validation of vendor clinical AI before go-live. Older evidence; Epic has since revised the model.
IBM Watson Health (STAT)
No figures on open portion of page
secondary: STAT analysis of why IBM's Watson Health was sold off. Argues IBM chose overly ambitious targets (cancer treatment, drug discovery) without deep scientific engagement and 'let the marketing get ahead of itself'. Remainder of article paywalled; sale price and investment not confirmed here.
Netherlands Court of Audit (Algemene Rekenkamer)
Ministries supplied 'many dozens of algorithms'; no fully self-learning algorithms found; government interests central rather than citizens' interests
First Court of Audit review of central-government algorithms; recommends common definitions, quality requirements and continuous monitoring. Less quantitative than the 2022 audit.
Michigan Unemployment Insurance Agency (MiDAS)
State review found the automated system was wrong 93% of the time; system flagged 53,633 fraud cases between late 2013 and mid-2015
secondary: Michigan Public report on a state review; number of cases reviewed is not stated on the page.
No use cases match these filters yet. or ask the library.