← All insights
Insight

AI headlines shrink when you read the source

We re-read 22 AI headline figures against their primary sources and several changed. What held up, and what to ask before you repeat a number.

12 Nov 2026

The Insight in 7 slides

Scroll through, then download the deck to share with your team.

Civico Insight

AI headlines shrink when you read the source

We re-read 22 AI headline figures against their primary sources and several changed. What held up, and what to ask before you repeat a number.

Civicocivicoventures.com/insights/2026-11-12-ai-headlines-shrink-at-the-source/
02 / 07
+21%
UPS adjusted US domestic operating profit, against the "-98%" headline

The 98% fall is the accounting figure after $1.17bn of restructuring charges. Read the source before you repeat the number.

CivicoUPS Q2 2026 earnings release, company-reported
03 / 07
What changed at the source
Klarna chats handled byAI, as repeated in presscoverage75%Klarna chats handled byAI, in its own pressrelease67%Firms with no AI-linkedheadcount change, asfirst read second-hand96%AI-using firms with noAI-linked headcountchange, in the Censuspaper95%
Klarna chats handled by AI, as repeated in presscoverage75%Klarna chats handled by AI, in its own pressrelease67%Firms with no AI-linked headcount change, asfirst read second-hand96%AI-using firms with no AI-linked headcountchange, in the Census paper95%

Klarna press release, 27 Feb 2024; Census Bureau working paper 26-25. The Census base is AI-using firms, not all firms.

Civicocivicoventures.com/insights/2026-11-12-ai-headlines-shrink-at-the-source/
04 / 07
How to check a number before you use it
01Find thesourceThe release,filing orpaper, not thearticle02Read thebaseAll firms, oronly theusers?03Find thebaselineCompared withwhat?04Name theownerWhose numberis it?05Repeat ordropSay what itdoes not show
01Find the sourceThe release, filing or paper, not the article02Read the baseAll firms, or only the users?03Find the baselineCompared with what?04Name the ownerWhose number is it?05Repeat or dropSay what it does not show
Civicocivicoventures.com/insights/2026-11-12-ai-headlines-shrink-at-the-source/
05 / 07

A headline number is a lead. The source is the evidence.

Civico Insight, 12 Nov 2026
Civicocivicoventures.com/insights/2026-11-12-ai-headlines-shrink-at-the-source/
06 / 07
Three questions before you repeat an AI number
  1. Have I read the original, or only the article about it?
  2. Is it a share of all firms, or only of those using AI?
  3. Compared with what, and who signed the baseline?
Civicocivicoventures.com/insights/2026-11-12-ai-headlines-shrink-at-the-source/
Where Civico fits

Bring a real problem.

The AI Transformation Studio takes a leadership team’s own AI initiatives to a fund-or-stop decision in 12 to 16 weeks. Vietnam Speed, our AI Venture Residential, puts executive cohorts on a real venture in four days.

civicoventures.com
Civicocivicoventures.com/insights/2026-11-12-ai-headlines-shrink-at-the-source/
1 / 7Share on LinkedIn

The question this answers: before you act on an AI number you read this month, how do you know it says what it seems to say?

Bottom line: when we replaced second-hand AI figures with the primary source, several changed: a rate, a base, a cause. The numbers that survive are narrower, and they come from firms that measured against a baseline and named an owner.

This week’s three findings

1. The loudest figures change when you read the source

The number: two-thirds of customer-service chats, not 75%, is what Klarna’s own press release says about its AI assistant’s first month. The 75% came later from press coverage.

What we saw: the UPS release shows US domestic operating profit down 98% on the accounting measure after $1.17bn of restructuring charges, and up 21% on the adjusted measure. The Census working paper says about 95% of AI-using firms report no AI-driven change in employment, a share of firms that use AI, not of all firms. Telstra credits its more than 20% software-development gain to a package of changes (vendors cut from 400 to two, interfaces rebuilt, AI), and DBS’s chief executive says the value of generative AI is harder to isolate than that of its earlier models. All company-reported or firm-reported.

2/3of Klarna's chatshandled in monthone, not 75%+21%UPS adjusted USdomestic profit,against -98% onthe accountingmeasure~95%of AI-using USfirms report noAI-linkedheadcount change>20%Telstra softwareefficiency,credited to apackage thatincludes AI
2/3of Klarna's chatshandled in month one,not 75%+21%UPS adjusted USdomestic profit,against -98% on theaccounting measure~95%of AI-using US firmsreport no AI-linkedheadcount change>20%Telstra softwareefficiency, creditedto a package thatincludes AI
Company releases and the US Census Bureau working paper, as linked above. All company-reported or firm-reported.

So what: none of these figures was invented. Each drifted in retelling: a rate rounded up, a base widened, a cause narrowed to AI. The drift runs one way, towards a cleaner story, which is why a number that has travelled through three articles is a lead and not a result. The firms behind the original numbers are, in fairness, careful: Klarna’s figure is its own and unaudited, Telstra lists its other enablers, the DBS chief executive says what he cannot isolate. For a CEO the cost of the drift is practical. If a board paper quotes “75% automation” and the source says two-thirds, the sponsor is judged against a target nobody set.

A headline number is a lead. The source is the evidence.
Civico Insight

Do this:

  • For every AI figure in your next board or investor paper, link the original release, filing or paper, not the article.
  • Write the base next to the number: all firms, or only those using AI; all chats, or the first month.

2. Training is a system, not a course, and almost nobody has built the system

The number: 15% of AI-using US firms trained their staff for it, and 64% made no organisational change (Census Bureau working paper, firm-reported).

What we saw: an MIT Sloan field study of two organisations that gave staff the same sandboxes and training found 141 organisation-wide AI solutions in use at a medical centre and 3 at a law firm. The medical centre kept training continuous, used shared evaluation rubrics and counted AI work in reviews and promotions; the law firm trained once. On the buyer side of the classroom, a Return on AI Institute survey of 1,006 executives, reported second-hand, links a 23-point edge in high AI value to training both staff and leaders. GMAC’s recruiter survey has 36% of MBA recruiters rating graduates adequately prepared to use AI, against 17% of programme leaders. The UNICON and Nexed survey finds business schools hold 17.2% of AI capability building, with 15% of buyers saying their learning needs are met very well.

Training is rare, and few are satisfied with itUS AI-using firms thattrained staff for AI15%US Census Bureau working paperBuyers who say their AIlearning needs are metvery well15%UNICON and Nexed, 180 organisationsProgramme leaders whorate MBA graduatesprepared for AI17%GMAC pulse survey, reported second-handRecruiters who rate MBAgraduates prepared forAI36%GMAC pulse survey, reported second-hand
Training is rare, and few aresatisfied with itUS AI-using firms that trained staff for AI15%US Census Bureau working paperBuyers who say their AI learning needs are metvery well15%UNICON and Nexed, 180 organisationsProgramme leaders who rate MBA graduatesprepared for AI17%GMAC pulse survey, reported second-handRecruiters who rate MBA graduates prepared forAI36%GMAC pulse survey, reported second-hand
Four different surveys with different bases, so compare the pattern, not the bars with each other.

So what: the Sloan comparison is the most useful result of the month because the inputs were equal. Same tools, same initial training, same readiness: 141 against 3. What differed was everything after the first workshop: continued sessions, shared quality checks, recognition in reviews. That reads as training being a management system, not a course, which fits the Census finding that most firms using AI changed nothing about how they work. It also explains the recruiter gap. Recruiters and programme leaders agree on the skills that matter (judging AI output, knowing its limits and risks); they disagree on whether graduates have them. The Sloan study has two sites, so treat it as a strong pointer, not a rule.

Do this:

  • If you run a company, ask what happens after the first AI workshop: who reviews the work, and is AI use counted in appraisals?
  • If you run a programme, have participants bring a live problem from their own firm and be judged on the decision they take home, not on a completed module.

3. In Vietnam, banks lead with a number and the regulator asks who executed

The number: 25-30% higher productivity from AI and Copilot with headcount roughly flat since 2020-21, said a Techcombank executive at the Smart Banking 2026 conference. It is company-reported, with no baseline or method given.

What we saw: Cake by VPBank reports income up 225% against operating costs up 83%, with more than 100 in-house AI models; the article gives no way to separate AI from growth. The State Bank of Vietnam’s new circular, as reported, caps AI-executed payments at VND 100m a day per individual customer and requires a human able to intervene and a record of whether AI or a person executed each transaction. Decree 142 implementing the AI Law sets three risk tiers and took effect on 1 May 2026. In the UK, the financial regulator’s new review of AI practice names a senior manager accountable for each use case as good practice.

So what: the two halves of this finding pull in opposite directions. The banks quote results with no baseline, so the number cannot be checked. The regulator is writing the opposite habit into rules: record who or what acted, set a limit, keep a human able to intervene. For a founder-led firm the useful reading is not about banking. A firm that logs which decisions the AI made, and who could override them, can later show what the AI did. A firm that does not will be left with a percentage nobody can defend, as in finding 1. The Vietnamese rules are reported through law-firm and legal-news summaries, so read the circular itself before you rely on its figures.

Do this:

  • When a vendor or peer quotes a productivity gain, ask for the baseline, the period and who measured it.
  • For any AI that acts (sends, pays, approves), log whether the AI or a person did it, and name the person who can override.

The surprise of the week

You would expect: a structured, prescribed protocol for using AI in a team to beat people using it freely.

We found: in a Microsoft experiment with 388 Gap Inc. employees in 194 pairs, the structured joint-AI protocol scored about 5 points lower on a 22-point scale.

YOU WOULD EXPECTA prescribed AI workflow beatsad-hoc useTraining everyone in the sameprotocol lifts qualityWE FOUNDThe structured protocol scored about5 points lower on a 22-point scaleIndividual training made nosignificant difference on the mainmeasure
YOU WOULD EXPECTA prescribed AI workflow beats ad-hoc useTraining everyone in the same protocollifts qualityWE FOUNDThe structured protocol scored about 5points lower on a 22-point scaleIndividual training made no significantdifference on the main measure
Microsoft and Gap Inc. field experiment, an arXiv preprint (April 2026); 194 pairs using the same Copilot tool.

What it means: this is a preprint with confounds its authors note (session timing, a grader that favoured longer documents), so it does not prove that structure never helps. It does show that a prescribed AI workflow can backfire. Before rolling out a standard way of working with AI, test it against free use on your own work.

Your next 30 days

Check a number before you use it01Find thesourceThe release orpaper02Read thebaseAll firms orusers only?03Find thebaselineCompared withwhat?04Name theownerWhose number?05Repeat ordropSay what itdoes not show
Check a number before you use it01Find the sourceThe release or paper02Read the baseAll firms or users only?03Find the baselineCompared with what?04Name the ownerWhose number?05Repeat or dropSay what it does not show

If you run a company:

  1. List the AI numbers in your last board or investor paper and link the original source for each.
  2. Replace any figure you cannot trace with one you can, or drop it.
  3. For the AI project you care most about, write the baseline and name the owner this month.
  4. Ask what happens after the first AI workshop: who reviews the output and is it counted in reviews?

If you run an executive programme: your buyers are asking for proof, and so are your recruiters. Only 17% of programme leaders rate graduates as prepared, against 36% of recruiters (GMAC, reported second-hand).

  1. Teach participants to trace a headline to its source and read its base; it is the skill recruiters rate lowest, measuring AI return.
  2. Test any prescribed AI workflow you teach against free use before you teach it.

Where Civico fits

  • For CEOs and COOs: the AI Transformation Studio is a 12 to 16 week programme in which your leadership team builds its own AI initiatives with AI and the work and decisions are redesigned around them; your sponsors decide fund or stop at the end. Companies start with a half-day scoping workshop. Tell us the initiative.
  • For programme directors: Vietnam Speed is a four-day AI venture residential in which EMBA and executive cohorts build, test and pitch a real venture or AI initiative, judged on the decisions they take home. Talk to us about an edition.
  • For anyone checking the numbers: every figure above links to its source, and the rows are in the Use cases view.

Method and limits

This week 22 existing rows were rebuilt on primary sources (the report, release, transcript or filing) and 24 new rows and 4 funding rounds were added; four rows were held back for weak sourcing. Company figures are company statements, not audited AI effects. Second-hand in this note: the Return on AI Institute, GMAC and AACSB survey figures, the Vietnamese circular and decree, and the Techcombank and VPBank figures. The Census paper is dated April 2026 with no day. Several earnings-call transcripts come from a transcript site, not the company. The Gap experiment is a preprint. This is mostly large-firm evidence.

Quick poll

The last time you quoted an AI statistic, had you read the original source?

Anonymous. One tap.

Get the weekly Insight

One short email a week on AI and leadership, with the evidence linked. Unsubscribe any time.

The data behind this Insight

Bring a real problem.

Civico runs experiential programmes for business leaders transforming their companies with AI. Leave with a decision that holds up.

Thank you.

Follow Civico on LinkedIn for the weekly Insight and news from the residentials.