Research · Frameworks

Which AI playbook can you trust?

The short answer

The frameworks barely differ in what they say. They differ in how good their evidence is, and none has independent proof that adopting it causes P&L results.

15frameworks checked
3with independent evidence
7self-published or vendor
How to read this
  • Each card is one named framework or study from a consultancy, business school, cloud vendor or standards body: what it says, how good its evidence is, and where it stops.
  • Independent evidence comes from someone without a stake in the framework. Self-published and vendor evidence comes from the firm selling it.
  • Open “Evidence and limits” on any card for the detail and the source.
Updated 25 Aug 2026Download the data ↓
What the best evidence supports

The best-supported pattern is a correlation: firms that redesign whole workflows, put a named owner and KPIs on every use case, and put most of the effort into people and process report much higher returns. Those firms are a small minority (about 6% in McKinsey's 2026 survey).

For founder-led firms

The evidence points to a light, value-first operating model on two or three core workflows, not a purchased framework. No framework has been validated for small firms or for Vietnam, and no AI executive programme has independently measured P&L outcomes. The only randomised trial with small firms (Kenya) found no average effect on revenue or profit from an AI mentor: high performers gained and low performers lost.

Six elements recur across the frameworks
  1. A CEO-owned ambitiontied to the core P&L
  2. Redesign workflowsinstead of inserting tools
  3. Track value rigorouslywith a named owner and KPIs per use case
  4. AI-ready dataand reusable platforms
  5. People and changeBCG's 70% of the effort
  6. Governancein proportion to risk

The 15 frameworks, strongest evidence first

IndependentMixedSelf-published or vendor
Independent survey1 Dec 2025

AI adoption by small and medium-sized enterprises

OECD · Institution

Segments SMEs into adoption stages with stage-specific barriers and policy supports.

Small and founder-led firms adopt later and see lessSkills and training are the binding constraint

Evidence and limits

Evidence. 5,000+ SMEs in seven countries. Skills are the top barrier and cost is not. Only 23.6% of GenAI-using SMEs train staff on AI.

Limits. Policy-oriented, not a firm-level playbook. No P&L outcome data.

Source ↗ Confirmed from search results

Independent (corporate-funded)19 Dec 2024

Enterprise AI Maturity Model

MIT Center for Information Systems Research · Academic

Four stages: experiment and prepare; build pilots and capabilities; develop AI ways of working; become AI future-ready. Moving from stage 2 to 3 needs strategic alignment, modular platforms and data, AI-ready people and teams, and stewardship.

The pilot-to-production funnel narrows hardWinners redesign the work; they do not just add a tool

Evidence and limits

Evidence. Survey of 721 firms plus 16 executive interviews. Stages 1-2 below industry-average financial performance, stages 3-4 well above. The 2025 update says the biggest gain comes from moving from stage 2 to 3, which is the pilot-to-P&L step.

Limits. Cross-sectional and correlational. Survey predates generative AI (2022). Large-enterprise sample.

Source ↗ Page opened

Independent13 Aug 2024

Root Causes of Failure for AI Projects

RAND Corporation · Independent research

Five anti-patterns: misunderstood problem, missing data, chasing technology over the problem, inadequate infrastructure, applying AI to problems too hard for it.

The pilot-to-production funnel narrows hard

Evidence and limits

Evidence. 65 experienced practitioners interviewed. 84% cited leadership-driven root causes as primary.

Limits. Practitioner perceptions only. The "80% fail" figure is quoted from outside work. Pre-agent era.

Source ↗ Page opened

Self-reported survey28 Oct 2025

Accountable Acceleration (annual enterprise adoption study)

Wharton Human-AI Research with GBK Collective · Business school

Three-stage arc: exploration (2023), experimentation (2024), accountable acceleration (2025). Levers are people, process and ROI measurement.

Self-reported benefit outruns measured P&L impactSmall and founder-led firms adopt later and see less

Evidence and limits

Evidence. About 800 US enterprise decision-makers. 82% use GenAI weekly, 72% formally measure ROI, about three quarters report positive returns. Mid-sized and smaller firms report faster ROI.

Limits. A benchmark, not a framework tested for outcomes. "Positive ROI" is respondent-judged and often productivity-based.

Source ↗ Page opened

Case-based22 Sept 2025

AI Governance Alliance playbooks

World Economic Forum · Institution

Nine plays for responsible AI innovation (Sep 2025); a four-action framework for transformational AI in financial services (Jun 2026): board-accountable vision, systems, near-term vs long-term balance, workforce and culture.

Governance and liability catch up with deploymentThe pilot-to-production funnel narrows hard

Evidence and limits

Evidence. Built from showcase cases (financial-services playbook draws on 150+ executives from 100+ organisations).

Limits. Consensus documents with selection bias. Checklists, not evidence.

Source ↗ Confirmed from search results

Widely referenced1 Jan 2023

AI Risk Management Framework 1.0

US National Institute of Standards and Technology · Governance standard

Four functions: govern, map, measure, manage, plus a playbook and a generative AI profile (2024). Voluntary.

Governance and liability catch up with deployment

Evidence and limits

Evidence. No study found linking adoption to financial results.

Limits. A risk framework, not a value framework. Useful to size controls in proportion to risk.

Source ↗ Page opened

Preliminary1 Jul 2025

GenAI Divide: State of AI in Business 2025

MIT Media Lab Project NANDA · Academic

Diagnostic rather than a framework. Pilots fail because tools do not learn or adapt to workflows; partnering with learning-capable tools reaches deployment more often than building internally; budgets skew to visible sales and marketing pilots.

The pilot-to-production funnel narrows hard

Evidence and limits

Evidence. Review of 300+ public initiatives, interviews with 52 organisations and 153 leader surveys. Headline: 95% of organisations see no measurable P&L return.

Limits. Small, non-random sample; narrow P&L definition; methodology described inconsistently. Treat the 95% as a signal, not a base rate.

Source ↗ Confirmed from search results

Certification exists1 Dec 2023

ISO/IEC 42001 AI management system

ISO/IEC JTC 1/SC 42 · Governance standard

Certifiable plan-do-check-act management system: policy, roles, risk and impact assessment, lifecycle, data, transparency and third-party controls.

Governance and liability catch up with deployment

Evidence and limits

Evidence. First major cloud-provider accredited certification reported Nov 2024. A 2026 arXiv review notes adoption evidence remains thin.

Limits. Risk of governance theatre. Value is mainly customer and procurement trust, not P&L. Worth considering only if enterprise customers require it.

Source ↗ Confirmed from search results

VendorLiving document

ADKAR applied to AI adoption

Prosci · Change management

Individual change through awareness, desire, knowledge, ability and reinforcement, plus an AI adoption diagnostic.

Skills and training are the binding constraint

Evidence and limits

Evidence. 1,107 professionals surveyed: 63% cite human factors as a primary challenge; user proficiency is the top issue (38%).

Limits. Diagnoses adoption, not value. Prosci sells the method and certification. Publication date not confirmed.

Source ↗ Confirmed from search results

Self-published30 Jun 2025

AI Maturity Model (five levels) and AI-ready data

Gartner · Consultancy

Five levels from planning to leadership. Differentiators: use-case selection on value and feasibility, governance, engineering practice, dedicated AI leaders, centralised strategy and data, multiple value metrics.

The pilot-to-production funnel narrows hardSelf-reported benefit outruns measured P&L impactAgents can do more than firms have deployed

Evidence and limits

Evidence. 432 organisations. 45% of high-maturity firms keep AI projects running 3+ years against 20% of low-maturity firms. 63% of high-maturity firms run financial or ROI analyses.

Limits. Outcome is project longevity, not profit. Full model is paywalled. "60% of AI projects abandoned" is a forecast tied to AI-ready data, not an observed rate.

Source ↗ Page opened

VendorLiving document

Cloud Adoption Framework for AI and agents

Microsoft · Cloud vendor

Six phases: strategy, plan, ready, govern, secure, manage. Separate startup and enterprise checklists. Governance aligned to the NIST AI RMF. Agent adoption guidance added Dec 2025.

Governance and liability catch up with deployment

Evidence and limits

Evidence. No independent outcome evidence found.

Limits. Platform-centric. Strong on technical readiness, weak on P&L value capture. The startup checklist is one of few size-tiered tracks in any framework.

Source ↗ Confirmed from search results

VendorLiving document

Cloud Adoption Framework for AI, ML and generative AI

Amazon Web Services · Cloud vendor

Six perspectives: business, people, governance, platform, security, operations, with AI-specific capabilities ordered by maturity and a responsible-use capability.

Governance and liability catch up with deployment

Evidence and limits

Evidence. No independent outcome evidence found.

Limits. A capability checklist rather than a value-realisation method; same platform bias as other cloud frameworks. Current version date not confirmed.

Source ↗ Confirmed from search results

Self-published30 Sept 2025

Future-Built and the 10-20-70 rule

Boston Consulting Group · Consultancy

Maturity groups (future-built, scalers, laggards). Five strategies: CEO-led multiyear ambition, AI-first operating model, fit-for-purpose tech and data, workforce involvement and upskilling, rigorous value tracking. Effort split 10% algorithms, 20% technology and data, 70% people and process.

Winners redesign the work; they do not just add a toolThe pilot-to-production funnel narrows hardSelf-reported benefit outruns measured P&L impact

Evidence and limits

Evidence. 1,250 senior executives. Future-built firms show 1.7x revenue growth, 3.6x three-year shareholder return and 1.6x EBIT margin versus laggards. The 2026 index reports 48.5% of firms creating measurable value (secondary report).

Limits. Survey-defined maturity correlates with market outcomes but causality can run the other way. BCG sells the transformation it describes. The 2026 jump partly reflects a reframed segmentation.

Source ↗ Page opened

Self-published25 Aug 2026

State of AI / AI high-performer practices

McKinsey & Company (QuantumBlack) · Consultancy

Practices across strategy, talent, operating model, technology, data and scaling. High performers pursue growth as well as efficiency, redesign workflows, show visible leadership ownership and define human-in-the-loop processes.

Winners redesign the work; they do not just add a toolThe pilot-to-production funnel narrows hardSmall and founder-led firms adopt later and see less

Evidence and limits

Evidence. Annual survey. 37% of respondents attribute any EBIT impact to AI and high performers stay flat at about 6%. Large firms (1bn+ revenue) report scaling AI across the enterprise at 54% against about a third of smaller ones.

Limits. Self-reported attribution; heavy investment is a marker of high performance, not proof of cause. The 6% has not moved in two years, so the practices are not spreading quickly.

Source ↗ Page opened

Self-published2026

State of AI in the Enterprise (2026 edition)

Deloitte AI Institute · Consultancy

Progression from surface use to strategic reinvention, covering agentic AI, governance and workforce redesign.

Skills and training are the binding constraintAgents can do more than firms have deployedWinners redesign the work; they do not just add a tool

Evidence and limits

Evidence. 3,235 leaders in 24 countries (fielded Aug-Sep 2025). Secondary summaries report 84% have not redesigned jobs around AI and 21% have mature agent governance (not checked against the PDF).

Limits. Sample skewed to firms on the leading edge of AI (Deloitte's own note). High ROI-attainment figures conflict with other studies mainly because definitions differ.

Source ↗ Confirmed from search results

Read headline failure rates with care.

Widely quoted failure rates are weaker than their circulation suggests. MIT NANDA's "95%" rests on a small preliminary sample; RAND's "80%" is quoted from outside work, not measured by RAND; Gartner's "60% abandoned" is a forecast. Use them as directional signals. Not assessed in this pass: Google Cloud, Kotter, Bain, INSEAD, IMD, Stanford HAI and HBR frameworks. Vietnam data on adoption come mostly from vendor-sponsored surveys with inconsistent definitions.

Thank you.

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