Metis

A field guide and an assessment

Readiness is a measurement, not a feeling.

Metis scores ten dimensions in four pillars against five maturity levels and shows where the gaps are. Then it turns them into a sequenced 12-month plan.

Directionwhy, and where
Peoplewho, and how well
Foundationswith what, and how safely
Executionhow it runs, and what it returns

Take the assessment Read the guide

Scroll

Fig. 1 — An organisation as a skyline: ten dimensions, each as tall as its maturity.

01 The paradox

Nearly nine in ten respondents say their firm uses AI. Only 37% report any effect on earnings.

Most respondents say AI has made them personally more productive. Far fewer can trace it to the bottom line. The gap is not adoption; it is readiness.

0Say AI has improved their individual productivity (McKinsey 2026)

0Attribute any EBIT impact to AI; 39% in 2025 (McKinsey)

0Are high performers, attributing 5% or more of EBIT to AI; unchanged since 2025

BCG, using a different method, finds 5% of companies achieving AI value at scale and 60% reporting minimal revenue and cost gains despite substantial investment.

McKinsey, The state of AI in 2026, 25 Aug 2026 (1,719 respondents; self-reported) · BCG, The Widening AI Value Gap, 30 Sep 2025 (1,250 executives).

Fig. 2 — A hundred respondents. Nearly nine in ten report AI use; 37 report any effect on earnings; about six attribute 5% or more of EBIT to it.

02 Pilot purgatory

AI stalls between the demo and the P&L.

Many pilots never reach production. In S&P Global's 2025 survey, 42% of companies had abandoned most of their AI initiatives before production, up from 17% the year before. A pilot proves the technology; production proves the organisation.

0Of proofs of concept scrapped before broad adoption, on average (S&P Global)

0Of their AI initiatives delivered the expected ROI, CEOs say; 16% scaled enterprise-wide (IBM 2025)

0Of organisations have scaled AI across several business units or gone AI-first (Gartner 2026)

S&P Global Market Intelligence, Voice of the Enterprise: AI & ML Use Cases 2025, Mar 2025 (1,006 respondents, North America and Europe) · IBM Institute for Business Value, 2025 CEO Study, 6 May 2025 (2,000 CEOs) · Gartner survey, 1 Sep 2026 (1,303 organisations with US$50m+ revenue).

Fig. 3 — The valley between pilot and production: of every hundred proofs of concept, 46 are scrapped on the way (S&P Global average).

03 Not a technology problem

BCG's rule of thumb: 70% of the effort is people and process.

Algorithms take 10%, technology and data 20%. The split comes from BCG's casework, not from measurement, but the surveys below point the same way. Readiness is mostly people and process.

0Of CEOs say AI success depends more on people's adoption than on technology (IBM 2026)

0Of reported AI impact is accounted for by organisational factors; 32% by individual ones (Microsoft 2026)

0Reported business impact from a clear AI strategy; better tools alone add about 5 (BCG 2026)

BCG, Where's the Value in AI?, Oct 2024 (10-20-70 heuristic) · IBM Institute for Business Value, 2026 CEO Study, 4 May 2026 · Microsoft, 2026 Work Trend Index, 5 May 2026 (20,000 AI users, 10 markets; driver analysis) · BCG, AI at Work 2026, 3 Jun 2026 (11,749 workers; modelled comparison).

Fig. 4 — Where the effort goes: BCG's rule of thumb above, Microsoft's analysis of what drives reported impact below.

04 Shadow AI

Your people adopted AI before you did.

Many staff did not wait for an approved tool: in KPMG's 2025 study, 70% of employees who regularly use AI at work used free public tools. Shadow AI is demand your organisation has not yet met.

0Of AI users brought their own AI tools to work (Microsoft & LinkedIn 2024)

0Of Singapore knowledge workers used generative AI at work; 84% of users brought their own (2024)

0Of employees would use AI tools even if not authorised (BCG 2025, stated intent)

A small, preliminary MIT NANDA study (2025) found that only 40% of companies said they had bought an official LLM subscription, while workers at over 90% of the companies surveyed reported regular use of personal AI tools.

Microsoft & LinkedIn, 2024 Work Trend Index, 8 May 2024 (31,000 knowledge workers, 31 markets) and Singapore release, 9 May 2024 · KPMG & University of Melbourne, Trust, attitudes and use of AI, Apr 2025 · BCG, AI at Work 2025, 26 Jun 2025 · MIT NANDA, The GenAI Divide, Jul 2025 (v0.1, not peer reviewed).

Fig. 5 — After hours: of every hundred people using AI at work in 2024, 78 brought their own tool (84 in Singapore).

05 The perception gap

Many staff hide their AI use. Leaders underestimate it.

Using AI can feel like cheating, or make people look replaceable, so they keep quiet. Leaders then plan for the use they can see. You cannot govern use you do not know about.

0Of employees surveyed say they hide their AI use and present AI work as their own (KPMG 2025)

0Of desk workers would be uncomfortable telling their manager they use AI (Slack 2024)

4 vs 13Heavy AI users per hundred employees: executives' estimate vs employees' own (McKinsey, US)

KPMG & University of Melbourne, Apr 2025 · Slack Workforce Index, Fall 2024 (17,372 desk workers, 15 countries) · Microsoft & LinkedIn, 2024 Work Trend Index · McKinsey, Superagency in the workplace, 28 Jan 2025 (US surveys, Oct–Nov 2024; heavy use means generative AI for at least 30% of daily work).

Fig. 6 — What the executive suite sees and what employees report: four heavy users in a hundred estimated, 13 self-reported.

06 The framework

Readiness has ten parts. They have to move together.

Metis scores ten dimensions in four pillars. Strong technology does not make up for untrained people, and a bold strategy does not make up for missing data. In the scoring, your weakest foundation caps how far you can scale.

Fig. 7 — The ten dimensions as a skyline, grouped into four districts. Select a pillar to light it.

07 Five levels

From Dormant to AI-native, one level at a time.

Every question describes what each level looks like in practice. Your ambition sets the target level; the gap sets the plan.

Five levels, as in SEI's Capability Maturity Model (1993) and Microsoft's Responsible AI Maturity Model (2023).

Fig. 8 — The same organisation at each level. Select a level to see it.

08 People & culture

Where managers use AI, teams report more value and trust.

The link is correlational, but it points to where to start. Fear of job loss is real, and higher among leaders and managers than on the front line. Start with the managers: help them use AI openly first.

0Trust in AI agents where managers model AI use; value +17, critical thinking +22 (Microsoft-led study of 1,800 workers, in WTI 2026)

0Feel positive about generative AI with strong leadership support, against 15% otherwise (BCG 2025, correlational)

0Of AI users are rewarded for reinventing work, even when results fall short (Microsoft 2026)

Microsoft, 2026 Work Trend Index, 5 May 2026 (17/22/30: separate Microsoft-led study of 1,800 workers; 13%: 20,000 AI users in 10 markets; self-reported, correlational) · BCG, AI at Work 2025, 26 Jun 2025 (10,635 employees, 11 countries; 41% worry their job will probably or certainly disappear within ten years: leaders and managers 43%, frontline 36%).

Fig. 9 — The manager effect: where the manager models AI use, the team's reported value, critical thinking and trust in agents are 17, 22 and 30 points higher (self-reported, correlational).

09 Education & learning

Training is a neglected lever. More than half of employees have had none.

Regular AI use is far more common among employees with enough training plus in-person coaching, BCG found. Yet most do not feel adequately trained. Budget hours of practice, not just licences.

0Of training plus coaching: the dose linked to much higher regular use (BCG 2025, correlational)

0Of employees had received any AI training (KPMG 2025, 47 countries)

0Of employees felt adequately trained on AI (BCG 2025)

BCG, AI at Work 2025, 26 Jun 2025 (10,635 employees, 11 countries; correlational) · KPMG & University of Melbourne, Trust, attitudes and use of AI, Apr 2025 (48,000+ people, 47 countries).

Fig. 10 — The dose: regular use is far more common among employees with five hours of training and coaching. Illustrative and correlational: BCG reports the direction, not a curve.

10 AI fluency & skills

Using AI is easy. Using it well is a skill you can measure.

Most employees use AI in their daily work, EY found, but few in ways that change how they work. The 4D framework names four skills fluent use rests on: delegation, description, discernment and diligence. Fluency, not access, is the KPI.

0Of employees use AI in advanced ways; 88% use it in their daily work (EY 2025)

0Have relied on AI output without evaluating its accuracy (KPMG 2025)

0Have made mistakes in their work because of AI (KPMG 2025)

Delegation: Deciding what to hand to AI, what to keep and how to split the work.
Description: Saying clearly what you want, how to approach it and how to behave.
Discernment: Judging what comes back, then accepting, fixing or rejecting it.
Diligence: Owning the result: careful with data, open about AI's role, verified before it is shared.

Take the AI fluency check (12 scenarios) →

EY, Work Reimagined Survey, Nov 2025 (15,000 employees, 29 countries) · KPMG & University of Melbourne, Apr 2025 · 4D framework: Dakan and Feller, with Anthropic.

Fig. 11 — The fluency loop: decide what to hand over, describe it, judge what comes back, own the result.

11 Productivity

Gains are real, uneven and easy to misread.

Inside AI's 'jagged frontier', studies find big gains, often biggest for the least experienced. Just outside it, people who trust the output do worse, and users misjudge their own speed. Measure before you scale; perception is not evidence.

0Time on professional writing tasks, with quality up 18% (Noy & Zhang)

0Support issues resolved per hour; gains concentrated among less-experienced agents (Brynjolfsson, Li & Raymond, QJE 2025)

0Correct answers on a task outside AI's frontier (758 BCG consultants)

0Time taken by 16 experienced developers on their own code; they believed they were 20% faster (METR)

Noy & Zhang, Science, Jul 2023 (453 professionals) · Brynjolfsson, Li & Raymond, QJE 140(2), May 2025 (5,172 agents at one firm; staggered rollout, not an RCT) · Dell'Acqua et al., HBS Working Paper 24-013, Sep 2023 · METR, 10 Jul 2025 (246 tasks, early-2025 tools).

Fig. 12 — A jagged frontier: inside it, AI speeds the work up; one step outside, accuracy falls.

12 Capturing value

Saved time leaks away unless the work is redesigned.

Users say AI saves them hours, but most get little help redirecting the time. A Danish study found no detectable effect on workers' earnings or recorded hours in 2023–24. Value appears where the workflow changes, not where the tool lands.

0Of regular frontline (non-manager) AI users say AI saves them at least a full workday a week (BCG 2026)

0Get limited or no guidance on what to do with the time saved (BCG 2026)

¾ vs ¼Nearly three-quarters of AI high performers have redesigned workflows; a quarter of others have (McKinsey 2026)

BCG, AI at Work 2026, 3 Jun 2026 (11,749 workers, 14 markets; self-reported) · Humlum & Vestergaard, Large Language Models, Small Labor Market Effects, NBER w33777, 2025 (Denmark, 2023–24 data) · McKinsey, The state of AI in 2026, 25 Aug 2026 (association, not cause).

Fig. 13 — A reported day a week saved, and where it drains when nobody redirects it.

13 Data foundations

Data is ready for AI one use case at a time.

Gartner calls traditional data management 'too slow, too structured, and too rigid for AI teams'. Metis tests each use case at three gates: representative, accurate enough, governed. Certify data for each use case, not once for the estate.

0Lack, or are unsure they have, the right data practices for AI (Gartner survey, 2024)

0Of AI projects without AI-ready data will be abandoned through 2026: a Gartner prediction

0More invested in data, governance, people and change, as a share of revenue, where AI succeeds than where outcomes are poor (Gartner 2026; self-reported)

Gartner, Lack of AI-Ready Data Puts AI Projects at Risk, 26 Feb 2025 (survey and prediction) · Gartner, press release, 16 Apr 2026 (353 D&A and AI leaders, Nov–Dec 2025; self-reported). The three gates are Metis's own.

Fig. 14 — Metis's three gates for each use case: representative, accurate enough, governed. The 60% is a prediction and is drawn dashed.

14 Governance, risk & security

In IBM's 2026 study, 92% of AI-breached firms lacked proper AI access controls.

Prompt injection tops OWASP's 2026 list of LLM risks, and OWASP says no reliable prevention exists yet. Access controls limit the damage when a filter fails. Contain what AI can reach; do not count on filters.

0Of breached organisations had an AI-related breach, up from 13% in 2025 (IBM 2026)

0Reported a security incident involving shadow AI, up from 20% in 2025 (IBM 2026)

0Lacked AI governance policies to manage AI or detect shadow AI; 63% in 2025 (IBM 2026)

0More per breach, on average, where shadow AI was high (IBM 2025; correlational)

IBM & Ponemon Institute, Cost of a Data Breach Report 2026: The AI tipping point (602 organisations breached Mar 2025–Feb 2026) and 2025 report, 30 Jul 2025 (600, Mar 2024–Feb 2025); shares of breached organisations, not of all firms · OWASP, LLM Top 10 2026, 4 Aug 2026.

Fig. 15 — Doors into the AI estate: where AI was breached, fewer than one in ten had proper access controls.

15 The rules · Sep 2026

The deadlines moved. The direction did not.

The EU legislates; Singapore guides. The Digital Omnibus deferred EU high-risk duties by 12 to 16 months; the bans, literacy duty and general-purpose AI rules still apply. Deferred is not cancelled: build your AI inventory and human oversight now.

EU AI ActIn force Aug 2024. Bans and AI literacy from Feb 2025, general-purpose AI rules from Aug 2025, most other rules from Aug 2026.
Digital OmnibusRegulation (EU) 2026/1744, in force 27 Jul 2026. High-risk duties move to 2 Dec 2027, and to 2 Aug 2028 for AI in regulated products.
AI literacySince the Omnibus, providers and deployers must support their staff's AI literacy rather than ensure a sufficient level.
SingaporeVoluntary Model AI Governance Frameworks: 2020 (2nd ed.), generative AI (2024) and agentic AI (Jan 2026; v1.5 May 2026).
ISO/IEC 42001A certifiable AI management system standard (2023). It does not by itself show AI Act compliance.
NIST AI RMFVoluntary US framework, 2023: Govern, Map, Measure, Manage.

Regulations (EU) 2024/1689, 2026/1744 · IMDA & PDPC (2020); IMDA & AI Verify Foundation (2024); IMDA (2026) · ISO/IEC 42001 · NIST AI 100-1. Not legal advice.

Fig. 16 — The EU timeline, with the high-risk deadlines the Omnibus moved; Singapore's voluntary frameworks in a lane of their own.

16 Agents raise the bar

When software acts for you, every weak foundation shows.

Agents do not just answer; they act, with your data, tools and credentials. On METR's software benchmark, the tasks they can finish half the time doubled in length about every seven months to early 2025. Give every agent an identity, least privilege and a human checkpoint.

0Of respondents at US$1bn+ firms are scaling agents in at least one function, up from 27%; smaller firms flat at 22% (McKinsey 2026)

0Report a mature governance model for autonomous agents (Deloitte 2026)

0Of agentic AI projects will be cancelled by the end of 2027: a Gartner prediction (2025)

McKinsey, The state of AI in 2026, 25 Aug 2026 · Deloitte, The State of AI in the Enterprise, Jan 2026 (3,235 leaders) · Gartner, 25 Jun 2025 (prediction) · METR, Measuring AI Ability to Complete Long Tasks, Mar 2025 (benchmark; 50% success; software tasks).

Fig. 17 — An agent's reach, and the three controls that bound it: identity, permissions, checkpoint.

17 Eight archetypes

Ten scores make a pattern. Metis names eight.

The assessment matches your profile to one of them, with the risk it usually carries; the report adds the first move out of it. The pattern tells you what to fix first.

The BystanderStaff use AI; the organisation does not.Risk · Invisible shadow use
The EnthusiastGrassroots energy running ahead of the guardrails.Risk · Shadow AI and data leakage
The BuilderPlatforms ready, people not yet.Risk · Shelfware
The GuardianSafe, compliant and mostly unused.Risk · Policy becomes the product
The StrategistA strong story that has not reached the work.Risk · Strategy on slides
The ExperimenterPilots everywhere, value nowhere yet.Risk · Pilot purgatory
The ScalerBalanced foundations, ready to industrialise.Risk · Spreading tools without redesigning work
The NativeAI is how you work, not a project you run.Risk · Complacency

18 If you are in Singapore

The ambition is there. The plan often is not.

Government schemes can pay for part of building that plan: advice, approved solutions and training. Terms change with each Budget, so we link to the official pages rather than quote amounts. Check eligibility before you commit spend.

0Of Singapore leaders said in 2024 their company must adopt AI to stay competitive (79% globally)

0Worried their leadership lacked a plan and vision to implement it (60% globally)

Enterprise Development GrantCo-funds upgrading and transformation projects; AI-led change can qualify. Check current support levels.
SMEs Go DigitalIMDA's Industry Digital Plans, pre-approved solutions under the Productivity Solutions Grant, and CTO-as-a-Service advice.
Enterprise Innovation SchemeEnhanced tax deductions on qualifying innovation and training spend. IRAS sets what counts.
SkillsFuture and TeSATraining support; IMDA's TechSkills Accelerator is being expanded to build 'AI bilingual' professionals.
AI Singapore100 Experiments co-develops AI projects with firms that its AI Readiness Index rates as ready.
PDPC advisory guidelinesHow the PDPA applies to personal data used in AI recommendation and decision systems (March 2024).

Microsoft & LinkedIn, 2024 Work Trend Index: Singapore, 9 May 2024 (Singapore leader subsample; indicative). Programme links are under Sources; terms are as each agency publishes them.

19 The engagement

From opinion to an agreed plan in six weeks.

We measure first and debate second. Every score traces back to an answer, an interview or a document.

Six weeks is our engagement design for a mid-sized organisation, not an industry benchmark.

Fig. 18 — Six weeks, five steps, each feeding the next. Select a step to follow it.

20 What you get

A number, a picture and a plan.

Everything the engagement produces is yours to keep, and the report saves as PDF or CSV. Rerun the assessment in a year to see what moved.

Readiness index and maturity level
Assessment
Ten-dimension profile against your target
Assessment · Workshop
Archetype and red flags
Assessment
Leader–staff gap in how AI is seen and used
Pulse · Deep dive
AI fluency baseline by role
Fluency check
12-month plan in four stages: mobilise, prove, embed, scale
Roadmap
KPI scorecard with starting targets
Roadmap
Interview and evidence pack
Evidence

The next step

Most large organisations use AI.
Few are ready for it.

Start with the ten-question quick scan. Then run the full assessment with your leaders and specialists, and the fluency check with every role. Measure it, then move it.

Take the quick scan Start the full assessment

— Notes & sources

Where the numbers come from.

The paradox. McKinsey, The state of AI in 2026: On the road to ROI, 25 Aug 2026 (1,719 respondents in 97 countries, May–Jun 2026; GDP-weighted and self-reported; 'nearly nine in ten' is McKinsey's wording, about respondents' organisations) · BCG, The Widening AI Value Gap, 30 Sep 2025. McKinsey's 6% and BCG's 5% use different methods and are not one statistic.
Pilot purgatory. S&P Global Market Intelligence, Voice of the Enterprise: AI & ML, Mar 2025 (fielded Oct–Nov 2024) · IBM, 2025 CEO Study, 6 May 2025 (shares of CEOs' initiatives, not of companies) · Gartner, only 22% have scaled AI across business units, 1 Sep 2026.
Not a technology problem. BCG, Where's the Value in AI?, Oct 2024 (10-20-70 is a rule of thumb) · IBM, 2026 CEO Study, 4 May 2026 · Microsoft, 2026 Work Trend Index, 5 May 2026 (AI users only; Singapore not surveyed) · BCG, AI at Work 2026, 3 Jun 2026.
Shadow AI. Microsoft & LinkedIn, 2024 Work Trend Index, 8 May 2024 · Microsoft Singapore release, 9 May 2024 · KPMG & University of Melbourne, Trust, attitudes and use of AI, Apr 2025 (48,000+ people, 47 countries) · BCG, AI at Work 2025, 26 Jun 2025 · MIT NANDA, The GenAI Divide, Jul 2025 (preliminary; 52 interviews, 153 survey responses).
The perception gap. KPMG global release, Apr 2025 · Slack, Fall 2024 Workforce Index · McKinsey, Superagency in the workplace, 28 Jan 2025 · Microsoft & LinkedIn, 2024 Work Trend Index (above).
The framework and five levels. SEI, Capability Maturity Model for Software v1.1, Feb 1993 (the five-step scale it popularised) · Microsoft Research, Responsible AI Maturity Model, May 2023. Metis also draws on MIT CISR, Cisco's AI Readiness Index and AI Singapore's AIRI; none of them endorses it.
People and culture. Microsoft, 2026 Work Trend Index (above; the manager-modelling figures come from a separate 1,800-worker Microsoft study cited in it) · BCG, AI at Work 2025 (above). Both are correlational.
Education and learning. BCG, AI at Work 2025 (above; 36% felt adequately trained) · KPMG & University of Melbourne (above; 47% had received AI training).
AI fluency and skills. EY, 2025 Work Reimagined Survey, Nov 2025 · KPMG global release (above) · AI Fluency: Framework & Foundations (framework by Rick Dakan and Joseph Feller, with Anthropic; course materials CC BY-NC-SA 4.0; definitions here are paraphrased).
Productivity. Noy & Zhang, Science, Jul 2023 · Brynjolfsson, Li & Raymond, Generative AI at Work, Quarterly Journal of Economics, May 2025 (first released as NBER w31161, Apr 2023) · Dell'Acqua et al., Navigating the Jagged Technological Frontier, HBS, Sep 2023 · METR, early-2025 AI and experienced developers, 10 Jul 2025 (METR warns against generalising it).
Capturing value. BCG, AI at Work 2026 (above) · Humlum & Vestergaard, Large Language Models, Small Labor Market Effects, NBER, 2025 · McKinsey, The state of AI in 2026 (above).
Data foundations. Gartner, Lack of AI-Ready Data Puts AI Projects at Risk, 26 Feb 2025 · Gartner, successful AI and data foundations, 16 Apr 2026.
Governance, risk and security. IBM & Ponemon, Cost of a Data Breach Report 2026: The AI tipping point, 2026 (release day not confirmed) · IBM & Ponemon, Cost of a Data Breach Report 2025, 30 Jul 2025 (the 2025 baseline and the US$670k). Both are shares of breached organisations, not of all firms · OWASP GenAI LLM Top 10 2026, 4 Aug 2026 (an awareness ranking, not a frequency count).
The rules. Regulation (EU) 2024/1689 (AI Act) · Regulation (EU) 2026/1744 (Digital Omnibus on AI) · IMDA & PDPC, Model AI Governance Framework, 2nd ed., Jan 2020 · IMDA & AI Verify Foundation, Model AI Governance Framework for Generative AI, 30 May 2024 · Model AI Governance Framework for Agentic AI, 22 Jan 2026, and v1.5, May 2026 · AI Verify · ISO/IEC 42001:2023 · NIST AI RMF 1.0, Jan 2023.
Agents raise the bar. McKinsey, The state of AI in 2026 (above; 'scaling' means in at least one function) · Deloitte, The State of AI in the Enterprise, Jan 2026 · Gartner prediction on agentic AI projects, 25 Jun 2025 · METR, Measuring AI Ability to Complete Long Tasks, Mar 2025.
If you are in Singapore. Microsoft Singapore release (above) · Enterprise Development Grant · IMDA, SMEs Go Digital · IRAS (Enterprise Innovation Scheme) · SkillsFuture · MDDI, Committee of Supply speech, 2 Mar 2026 (TeSA, 'AI bilingual') · AI Singapore · PDPC, Advisory Guidelines on Use of Personal Data in AI Recommendation and Decision Systems, Mar 2024.
How to read these numbers. Survey figures are what respondents said, not measurements, and correlational figures show association, not cause. Predictions are labelled as predictions, and the 10-20-70 split as a rule of thumb. The level descriptions, the three data gates and the archetypes are Metis's own. Many publisher sites could not be opened directly when this edition was checked, so figures were matched against publisher excerpts and full-text copies of the reports. The engagement's six weeks is our design, not a benchmark.

Metis is a Helikos Labs guide. Numbers are from the sources above, as of September 2026. The figures and the assessment run entirely in your browser; nothing you enter is sent anywhere.