The public method · Version monthly-2026-09-8a896c86
We publish every question, its rationale and the evidence behind it. You see the scoring model, the weights, the thresholds and every change. Trust in a benchmark comes from showing how the number is made. We do.
Scoring model
Every scored question maps to 0 to 4 points. Likert items run from strongly disagree (0) to strongly agree (4). Behavioural items run from anchor (a), 0, to anchor (e), 4. We store your raw answers, never just the computed scores. That is what lets us recalculate your score when the assessment changes, without asking you to sit it again.
Readiness: 18 questions across six dimensions, maximum 72 points, normalised to 0 to 100. Each dimension produces a 0 to 100 sub-score from its three questions. Realisation: 6 questions, maximum 24 points, normalised to 0 to 100.
The headline AiR Bench is a weighted blend: 60% readiness, 40% realisation. The split reflects today's evidence. Realisation is the goal. But for the mid-market majority, the score has to stay diagnostic of what to fix, and readiness is where the fixes live. As the population matures, we expect the weighting to shift towards realisation. That shift will be a published finding in its own right.
The AiR Grid splits at 50 on both axes. That is the midpoint of the scale, not a band boundary, so the grid and the bands stay independent readings.
Foundations need to be laid before a meaningful pilot is viable. Strategic clarity and leadership alignment come first.
Some foundations in place with one or two critical gaps. The dimension detail shows exactly which.
Strong foundations across most dimensions. Positioned to move into structured production work and start generating measured return.
The conditions for AI to compound value, and evidence that it is doing so. The question is sequencing, not whether to act.
We calibrate the bands so the top band stays genuinely rare. The research is consistent: real high performers are 5 to 13% of the population. A benchmark where a third of respondents are Leading is flattery, not measurement.
Low readiness, low realisation
Early. The work is to define one quantified opportunity and build the conditions around it. Honest, common, fixable.
High readiness, low realisation
The GenAI Divide quadrant, and the most populated in the research: conditions exist, value does not. The gap is execution design: workflow redesign, production discipline, measurement.
Low readiness, high realisation
Value is being created ahead of foundations. Often a few brilliant individuals or one heroic team. The value is real and fragile: governance, data and process debt will tax it.
High readiness, high realisation
Conditions and evidence. The work is sequencing and protecting focus. This is where compounding competitive advantage lives.
The questions · 24 scored · Why we ask each one
Behavioural anchors give us five clear states to score against, so answers are hard to game and easy to verify. We use Likert agreement scales only when the question is about attitude, not behaviour. Questions marked ✳ form the short-form subset.
Whether AI ambition has been converted into named, owned, quantified opportunities.
D1.1 ✳ · Agreement (Likert 0 to 4)
We have named specific AI opportunities, each tied to a measurable business outcome with a single accountable owner.
Why we ask · Named, owned, quantified opportunities are the single clearest marker of strategy converted into work. McKinsey's high performers are distinguished by outcome-based objectives, not by ambition statements.
Evidence: McKinsey, The State of AI (2025 edition)
D1.2 · Behavioural anchors (0 to 4)
Which best describes your AI priorities today?
Why we ask · The ladder from general ambition to board-reviewed portfolio is observable and hard to flatter: each anchor describes a state you are either in or not. The register offers limited direct evidence on strategic clarity in mid-sized firms. The McKinsey trust entry records only about one third of organisations reaching maturity level 3+ in strategy, alongside governance and agentic-AI governance, which is consistent with quantified portfolios remaining a minority practice. The Anthropic Econ Scenario Explorer is logged as a weak signal on scenario planning, but it is an interested party's model that by its own disclaimer omits policy responses, business cycles and demand effects, so it evidences nothing about firm-level strategy quality.
Evidence: Anthropic (Anthropic Economic Institute), Econ Scenario Explorer v1.0 / 'Scenarios for our Economic Future' · McKinsey, 'State of AI trust in 2026: Shifting to the agentic era'
D1.3 · Agreement (Likert 0 to 4)
Senior leadership can explain why AI matters to this organisation in the language of our strategy, not in general terms.
Why we ask · When leaders can only describe AI in general terms, prioritisation defaults to whoever shouts loudest. Strategy-specific language is a proxy for genuine strategic integration.
Evidence: McKinsey, The State of AI (2025 edition)
Leadership engagement, frontline involvement, and actual weekly use across the organisation.
D2.1 ✳ · Agreement (Likert 0 to 4)
Our leadership team actively champions AI adoption and dedicates meaningful time to it.
Why we ask · High performers show roughly three times the leadership engagement of everyone else. It is the strongest single differentiator in the McKinsey data.
Evidence: McKinsey, The State of AI (2025 edition)
D2.2 · Behavioural anchors (0 to 4)
What share of your people use AI tools in their actual work at least weekly?
Why we ask · Weekly use in real work is the adoption measure that survives scrutiny, where licence counts and pilot enrolment do not. The Stanford HAI AI Index entry records 53% population adoption inside three years, faster than the PC or the internet, which sets the diffusion backdrop against which any low internal figure should be read. The Microsoft Work Trend Index, drawing on over 100,000 anonymised Copilot chats and a survey of 20,000 AI-using workers across ten countries, records 49% of Copilot conversations supporting cognitive work, indicating that use is already substantive rather than peripheral where it occurs. Both are population or user-base measures, not organisation-level UK mid-market figures.
Evidence:
D2.3 ✳ · Agreement (Likert 0 to 4)
The teams who will use AI day to day have been involved in shaping how it is deployed.
Why we ask · The MIT research locates the failure of most pilots in a learning gap rather than a technology gap, and frontline involvement in deployment design is the observable counter. The Klarus UK and Ireland mid-market survey adds a sharper edge: 91% of senior decision makers expressed confidence in internal expertise across all areas of AI deployment, while lack of AI expertise was the joint top reason, cited by 48% of those that had explored AI, for projects failing to move beyond pilot. Since this item is answered from a leadership vantage point, it is exposed to exactly that confidence gap, and results should be read with the source caveats in mind: vendor-sponsored, self-reported, out of window.
Whether the work AI will touch is understood, documented and being redesigned, not just augmented.
D3.1 · Agreement (Likert 0 to 4)
The processes we want AI to improve are documented, well understood and currently measurable.
Why we ask · You cannot redesign what you cannot describe, and you cannot prove improvement without a baseline. Documented, measurable processes are the precondition for measured value.
Evidence: MIT NANDA, The GenAI Divide: State of AI in Business 2025
D3.2 ✳ · Behavioural anchors (0 to 4)
Has any workflow been redesigned end to end around AI, rather than AI being added to the existing way of working?
Why we ask · End-to-end workflow redesign is the behaviour that separates absorbing a tool from changing how work is done. The McKinsey State of AI entry supplies the reason it matters now: the share of organisations scaling AI is rising while the share reporting positive EBIT contribution sits at 37%, essentially unchanged from 2025, with cost reduction reported most often at function level in supply chain and service operations. Deployment without redesign is the most plausible shape of that decoupling. The evidence is a global, cross-size, self-reported panel with no UK or mid-market cut, so it motivates the question rather than setting an expected distribution.
Evidence: McKinsey, The State of AI: Global Survey 2026 (QuantumBlack)
D3.3 · Agreement (Likert 0 to 4)
For our target use cases, we know which steps AI will replace, which it will augment, and which it will leave unchanged.
Why we ask · Replace, augment, leave alone: the three-way split is the working vocabulary of process redesign. Knowing it per use case marks the difference between a plan and a hope.
Evidence: Anthropic Economic Index (January 2026)
Whether the data the use cases need is accessible, usable and governed.
D4.1 ✳ · Agreement (Likert 0 to 4)
We have access to the data our priority use cases require, and it is in a usable, organised state.
Why we ask · Anchoring to priority use cases is deliberate: data readiness in the abstract is unanswerable, whereas readiness for the three things you intend to do is not. Two register entries raise the stakes on this item. Microsoft AI's 'Frontier Tuning' applies reinforcement learning to a customer's own workflow traces inside their environment, which makes proprietary data an input to model quality rather than only to prompts. The Klarus UK and Ireland mid-market survey names data quality among the blockers to scaling. The Microsoft entry is a vendor announcement that is partly product marketing, and the Klarus release is vendor-sponsored, so both inform the framing without evidencing outcomes.
D4.2 · Behavioural anchors (0 to 4)
If you adopted a new AI tool tomorrow, how quickly could it be connected to the business data it needs, with appropriate controls?
Why we ask · Time-to-connect is the practical measure of data readiness, and the repeated, governed pattern at the top anchor is what Cisco calls the antidote to AI Infrastructure Debt.
Evidence: Cisco AI Readiness Index 2025
D4.3 · Agreement (Likert 0 to 4)
Security, privacy and quality constraints on our data are understood and managed; they shape our AI work rather than blocking it.
Why we ask · Constraints that are understood become design inputs; constraints that stay vague become vetoes, and the difference shows up in pilot-to-production conversion. The Klarus UK and Ireland mid-market survey names data quality and governance directly among the blockers to scaling, in the only UK-specific source the register holds. The MIT NANDA briefing records conversion improving in vendor-partnership deployments only, which suggests that where constraints are navigated it is often with external help rather than internal capability. Both sources are self-reported or vendor-sponsored and neither offers an audited view.
Whether use is governed by policy people actually know, with defined human review and explainability.
D5.1 ✳ · Agreement (Likert 0 to 4)
We have a clear policy on acceptable AI use, and the people doing the work actually know what it says.
Why we ask · The second clause is the question. Most organisations have a policy; far fewer have a policy anyone could recite. Governance that exists only as a document governs nothing.
Evidence: Cisco AI Readiness Index 2025
D5.2 · Behavioural anchors (0 to 4)
Where AI contributes to work that matters, is there a defined human review point?
Why we ask · As deployment becomes agentic, the defined human review point is the control that matters most, and the top anchor describes review as a living system because increasing autonomy demands one. The register records agentic capability shipping as product, with ChatGPT Work running GPT-5.6 Sol as its operating agent, while the McKinsey trust entry places only about one third of organisations at maturity level 3+ in agentic-AI governance. Multi-vendor incident analyses are logged as watching precisely because early production incident patterns are shaping what adequate oversight looks like. The governance figures come from a roughly 500-organisation panel scored on a vendor maturity model whose scoring detail was not retrieved.
Evidence: McKinsey, 'State of AI trust in 2026: Shifting to the agentic era'
D5.3 · Agreement (Likert 0 to 4)
We could explain to a customer or a regulator, today, where and how AI is used in our business.
Why we ask · Explainability on demand is the practical test of governance maturity, and 'today' is in the question because a capability requiring six weeks of preparation is not a capability. The register flags this item as due for re-anchoring: the second tranche of EU AI Act implementation guidance may give explainability obligations a legal definition, which would change what the question is actually testing. Until that guidance is read, the item measures self-assessed readiness against no external standard. The McKinsey trust entry, though still watching and based on fieldwork predating the window, records only about one third of organisations reaching maturity level 3+ in governance.
Evidence: McKinsey, 'State of AI trust in 2026: Shifting to the agentic era'
Whether the organisation ships digital change, and how fast.
D6.1 · Agreement (Likert 0 to 4)
We have launched new digital tools or ways of working successfully in the past 24 months.
Why we ask · Organisations that ship digital change ship AI change. A 24-month window keeps the evidence recent enough to mean something.
Evidence: DORA, State of DevOps (2014 to present)
D6.2 ✳ · Behavioural anchors (0 to 4)
From decision to working in production, how long does digital change typically take here?
Why we ask · Lead time from decision to production is the most durable delivery metric there is, carried here into the AI context. Speed is itself a readiness asset because the frontier keeps moving: the register records Gemini 3.5 in May, MAI models and Kimi K2.7 Code in June, and the GPT-5.6 family in July, with ChatGPT Work shipping agentic capability as product alongside it. An organisation whose typical lead time exceeds the interval between frontier releases will always be integrating the previous generation. The Klarus UK and Ireland mid-market survey names lack of AI expertise as the joint top reason projects fail to move beyond pilot, which is one mechanism by which lead time stretches.
D6.3 · Agreement (Likert 0 to 4)
When an initiative is stalling, we find out quickly and act, rather than letting it drift.
Why we ask · The 95% pilot failure figure is mostly a story of drift. Fast detection and decisive action on stalling work is the organisational muscle that prevents it.
Evidence: MIT NANDA, The GenAI Divide: State of AI in Business 2025
Production deployments, pilot-to-production conversion, measured value, where value concentrates, frontier responsiveness.
R1 ✳ · Behavioural anchors (0 to 4)
How many AI use cases are in production use today, meaning relied upon in real work, not piloted?
Why we ask · Production use, relied upon in real work, is the bar, and the definition sits inside the question because the adoption-to-value gap lives in the space between a pilot and a dependency. The St. Louis Fed and Brookings BPEA working paper is the calibration warning here: a single-question framing about 'production processes' materially undercounts adoption relative to an eight-technology framing, with 10% on the narrow wording implying roughly 45% on 'any purpose'. Counting discrete use cases rather than asking a yes-or-no adoption question is the deliberate response to that framing sensitivity, and it means AiR Bench figures are not comparable with headline adoption percentages from either framing.
Evidence:
R2 · Behavioural anchors (0 to 4)
Of the AI initiatives you have started in the last 18 months, what share reached production and are still running today?
Why we ask · Pilot-to-production conversion is the rate that separates sustained deployment from abandoned pilots. Adding 'and are still running today' closes the gap the McKinsey State of AI entry describes, where a rising share of organisations scale AI while the share reporting positive EBIT contribution stays essentially unchanged at 37%. A workflow that reached production and was later switched off should not count as conversion. The MIT NANDA briefing records conversion improving in vendor-partnership deployments only, so the question also needs to survive the case where a partner delivered the launch but nothing persisted after handover.
Evidence: McKinsey, The State of AI: Global Survey 2026 (QuantumBlack)
R3 ✳ · Behavioural anchors (0 to 4)
Has any AI deployment delivered value you have measured against a baseline?
Why we ask · Measured against a baseline is the only claim of value this instrument accepts, and anchor (b), 'we believe so, but have not measured it', is named without judgement because it is where much of the market sits. Two entries support keeping the bar high. The McKinsey State of AI entry records 37% reporting positive EBIT contribution, essentially unchanged from 2025 despite rising scaling, with several benefits reported in non-financial terms. The WRITER survey of 1,200 executives and 1,200 AI-using employees records 97% deploying agents in the past year and 59% investing over $1m annually, against only 29% seeing significant organisational ROI. Both are self-reported perception rather than audited financials, and neither carries a UK mid-market cut.
Evidence: McKinsey, The State of AI: Global Survey 2026 (QuantumBlack)
R4 · Behavioural anchors (0 to 4)
Where is the value you have seen concentrated?
Why we ask · Value climbs a ladder from individual to team to function to business model, and the location of value is a first-class measurement rather than a footnote. Two entries sharpen the middle of the ladder. The McKinsey State of AI entry records function-level cost reduction most often in supply chain and service operations, alongside non-EBIT gains in innovation, differentiation and customer and employee satisfaction, while enterprise-level EBIT impact stays with a minority. The Microsoft Work Trend Index entry records 58% of AI users, and 80% of its 'Frontier Professionals' segment, saying they produce work they could not a year ago, which is value concentrated at the individual rung. Both are self-reported and neither carries a UK mid-market cut.
Evidence: McKinsey, The State of AI: Global Survey 2026 (QuantumBlack)
R5 · Behavioural anchors (0 to 4)
When a significant new model capability is released, how long before it reaches your workflows?
Why we ask · The frontier moved repeatedly inside this register's span: Gemini 3.5 Flash and Pro in May, Microsoft AI's MAI family and Frontier Tuning in June, Moonshot's Kimi K2.7 Code in June, and the GPT-5.6 family in July following a restricted trusted-partner preview. Frontier responsiveness measures whether an organisation compounds those moves or pays for them later. Two register entries qualify the top anchor. The US Executive Order and the reported Anthropic access clash show that availability is now partly gated by government security posture, so an established evaluation route cannot guarantee access timing. The Anthropic Economic Index update records productivity contribution revised upward for integrated deployments, which is why the anchor rewards reaching workflows rather than reaching a sandbox.
Evidence:
R6 ✳ · Behavioural anchors (0 to 4)
Have you deployed any agentic systems, meaning AI carrying out multi-step work with defined autonomy, oversight and a named owner for when it goes wrong?
Why we ask · Agentic capability is now shipping as product rather than as model: ChatGPT Work with GPT-5.6 Sol as its operating agent, and Gemini 3.5 positioned as Google's strongest agentic and coding family. Governance is recorded as lagging that shipment, with the McKinsey trust entry placing only about one third of organisations at maturity level 3+ in agentic-AI governance specifically. Multi-vendor incident analyses are logged as shaping what 'defined autonomy and oversight' should mean in practice. Adding a named owner for failure to the definition prevents an unsupervised script, or an agent nobody is accountable for, from scoring as an agentic deployment.
Evidence: McKinsey, 'State of AI trust in 2026: Shifting to the agentic era'
Benchmarks and percentiles
Every readout shows your percentile against the full UK dataset and against your sector, on the headline score and each dimension. We publish sector percentiles only when the sector cell reaches the minimum sample of 25 respondents. Below that threshold, the readout says so plainly.
Until the dataset reaches critical mass, we seed benchmarks from the published research distributions (McKinsey, Cisco) and label them research-calibrated wherever they appear. The calibration targets are clear: high performers held to 5 to 13% of the population, the median in the Developing band, and realisation trailing readiness (the GenAI Divide). We will announce the switch to live data in the changelog. The assessment never pretends seeded data is its own.
Behavioural anchors are the backbone of scoring, verified advisory scores act as a calibration reference, and we flag repeated answer patterns internally. That is how we discourage gaming. We never call anyone out in public.
The evidence base · Founding sources
95% of enterprise AI pilots delivered no measurable P&L impact. The barrier is a learning gap, not the technology. Vendor partnerships succeed roughly twice as often as internal builds. Budgets concentrate in sales and marketing while ROI concentrates in operations.
88% of organisations use AI in at least one function, but only around 21% have redesigned any workflow end to end, and high performers are roughly 5% of the population. High performers show 3x leadership engagement and outcome-based objectives.
Around 13% of organisations qualify as Pacesetters and the share is static year on year. 83% plan agentic deployment while foundations lag. Introduces the concept of AI Infrastructure Debt.
Five economic primitives for measuring AI's workplace impact. Reliability-adjusted productivity contribution falls from 1.8 to roughly 1.0 percentage points: real but more modest than claimed, and dependent on integration quality.
Generative AI reached 53% population adoption within three years, faster than the PC or the internet. Adoption pace varies sharply by geography and correlates with GDP.
Four outcome metrics plus a capabilities model, published methodology, annual re-benchmarking, and tier classification turned a research programme into the industry's shared vocabulary.
The evidence base is updated every month. A research review decides what findings are strong enough to change the assessment, and revisions follow from that, not from the calendar alone. Every citation carries the study's scope, so claims cannot overreach.
Changelog · Semantic versioning
Patch versions clarify wording, with no scoring change. Minor versions adjust anchors or weights, with a published note. Major versions change questions: a new phase for the benchmark, with a bridge analysis. When the assessment is updated, your stored answers can score differently under the new version. Your answers have not changed; the criteria have. We show you that movement directly.
v1.1 · expected Q4 2026
Agentic deployment anchors (R6, D5.2) and frontier-responsiveness recalibration (R5). An agentic pilot that scores near the top of R6 today is expected to score lower under v1.1, not because the answer changed, but because the frontier did. Next research review: Research review in 9 days.
v1.0.1 · 2026-08-13
Patch fork, no measurement change. On 2026-07-28 a copy pass (commit 5957f82) reworded explanatory prose inside the definition that is hashed and frozen: three dimension introductions, the rationales behind R1, R3 and R6, and one quadrant diagnosis. No question, anchor, tier, weight or scoring rule changed, so an identical set of answers scores identically under 1.0.1 and 1.0.0. The version forks because the frozen definition moved, not because the measurement did.
v1.0.0 · 2026-06-09
Founding version. Six readiness dimensions, the realisation block, behavioural anchors as the scoring backbone. Weights set at 60/40 readiness to realisation, reflecting the current evidence: realisation is the goal, but for the mid-market majority the score must remain diagnostic of what to fix.
The covenant
Part of the product, not the small print
Individual results are private. Always. We never publish, share or sell results that identify you or your organisation.
Public
Aggregated, anonymised data is public: sector comparisons, score distributions, dimension patterns and the relationship between readiness and real value.
Minimum cell
We only publish a cut when it reaches 25 respondents. Below that, the cut does not publish, full stop.
Deletion
You can delete your data at any time. Deletion flows through to the aggregates at the next publication cycle.
No trackers
No third-party advertising trackers anywhere in the product. It would be incoherent next to this covenant.
Statement of interest
Projject builds this assessment, and Projject sells services. Here is how we keep the two honest. The assessment is independent work, not a sales pitch. The methodology is public, the scoring is deterministic and published, and recommendations come from an evidence-mapped library. They cite research, not Projject offerings. The readout has one call to action: a no-cost conversation, clearly labelled.
The benchmark only has value if you trust it. And you will only trust it if it survives scrutiny from someone who never intends to buy anything. That is the standard we hold every release to.
AiR Advisory · Protocol summary · Versioned with the assessment
Our consultants run a structured 60-minute evidence session with you. We check a defined sample of answers, weighted toward the behavioural anchors and the realisation questions, against real artefacts: production systems shown live, measurement baselines, governance documents and deployment records.
Where the evidence does not support an answer, we correct it and recalculate the score. You keep the corrected result, a badge valid for 12 months, and a one-page note recording what was evidenced. Advisory scores sit in a separate verified set in the public dataset. The gap between self-reported and verified results becomes one of the most interesting findings in the annual report.
The public method · Version monthly-2026-09-8a896c86
The method is public. You can inspect every step.
Benchmark your business →