How the AI Capability Index works

The AI Capability Index measures the human side of AI adoption: whether an organisation's leadership, governance, culture and people are ready to use AI well, not just whether the tools are switched on.

Methodology version 1.0 · October 2026

Where the Index is today

The framework and questions are complete and draw on published research. Validation is planned but not yet done, and there is no benchmark dataset yet. Organisations taking part now form our founding cohort, and benchmarks are published only once 30 organisations in a group have taken part. Scores are a structured, honest starting point, not a validated measurement. This page is updated as each validation step completes.

Why this Index exists

AI use is now widespread, but most organisations are not yet getting organisation-level value from it. The latest research points to organisational and human factors as the main reason: individuals are adopting AI faster than their organisations are adapting around them.

Three points follow. First, value depends on leadership, culture, governance and management practice, which is the territory HR and L&D leaders influence most. Second, the gap between employee use and organisational adoption means much AI use happens outside sanctioned tools and policies. Third, most existing readiness frameworks are built for technology and strategy audiences. The AI Capability Index is built for the people responsible for the human side.

Two cautions apply to all of these sources. They are self-reported surveys, and they show associations, not proof that one practice causes value. The ONS survey also excludes finance and insurance firms, so UK financial services evidence comes from other sources.

The six dimensions

The Index measures six dimensions, each with five questions and equal weight. Data and technology are combined into one dimension, because HR and L&D leaders rarely have reliable visibility of either. Manager enablement is its own dimension, because recent evidence identifies managers as one of the strongest levers on whether AI use creates value.

The dimensions are a practical, research-informed structure for a conversation about AI capability. Whether they behave as six distinct, measurable factors will be tested statistically once enough data exists (see Validation).

How the Index works

The Index has two levels. The free Leader Index gives one senior person a structured view of their organisation in about ten minutes. The Team Diagnostic tests that view against what other leaders and staff actually report and do.

Why two levels. A single respondent's view is useful for starting a conversation, but it is not evidence. Leaders tend to rate their organisations more favourably than staff do. In Microsoft's 2026 survey, leaders were more likely than employees to say it feels safe to suggest new ways of working with AI (81% vs 67%). The Team Diagnostic exists to close that gap.

Staff scenario questions. Staff are given short, realistic situations and choose what they would do. Examples include being asked to summarise a client file using an unapproved AI tool, or receiving an AI-generated figure that looks wrong. Answers are scored against good practice agreed with the client's own risk or compliance team. This measures judgement and behaviour, which self-rating questions cannot. Scenarios are tailored for regulated sectors, where data handling, customer outcomes and record-keeping carry specific obligations.

Staff anonymity. Staff results are reported only for groups of 10 or more people, and individual answers are never shared with the client.

Question design

Every scored question asks about one thing, describes each answer level in concrete terms, and offers “Don't know”. This means answers reflect what is actually in place, rather than optimism or guesswork. The questions follow six rules:

1. One idea per question. No question asks about two conditions at once, such as whether a policy exists and whether staff understand it.
2. Described levels, not agreement. Each question has five levels, each describing an observable state, from “not in place” to “embedded and reviewed”.
3. “Don't know” is always available. Guessing adds noise. “Don't know” is reported back as a visibility gap, which is useful in itself.
4. Neutral wording. Questions do not signal the “right” answer, and no statistics or explanations are shown while answering.
5. Observable evidence. Levels refer to things a respondent could point to, such as a document, a named role or a regular meeting, rather than feelings.
6. Short enough to finish. 30 scored questions plus a few profile and outcome questions, in about ten minutes. Completion time is measured and published.

Example: leadership ownership

Outcome questions. Four further questions ask about results rather than conditions. They cover the share of staff using AI at least weekly, whether the organisation can point to measured benefits, whether AI has caused any incidents or near misses, and the respondent's confidence that AI is creating value. They are not scored. They exist so the Index can later be tested against real outcomes.

Profile questions. We also ask for sector, organisation size, whether the organisation is regulated (and by whom), the respondent's role, and which AI tools are formally provided. These allow fair comparisons once benchmarks exist.

Scoring

Each dimension is scored from 0 to 100 using only the questions the respondent could answer. The overall score is the simple average of the six dimensions, all weighted equally.

Dimension score. Each answered question gives a value from 1 to 5, and “Don't know” answers are left out. A dimension is scored only if at least three of its five questions were answered. Otherwise it is shown as “not enough visibility to score”.

Overall score. The mean of the scored dimensions, shown only if at least five of the six dimensions could be scored.

Visibility score. The share of the 30 questions answered with something other than “Don't know”. It is reported alongside the overall score, because a high score with low visibility means something different from a high score with high visibility.

Why equal weights. The current evidence does not support weighting some dimensions more than others with any precision, so all six count equally. Weights may be revisited once outcome data shows which dimensions relate most strongly to results.

Maturity stages

The overall score places an organisation in one of five stages. The cut-points are provisional and will be reviewed against data.

Report bands. For each dimension, the report gives guidance for one of three bands: Foundations (0–39), Developing (40–69) or Established (70–100). These bands are also provisional.

Reading the numbers. Scores are shown as whole numbers. For a single respondent, differences of fewer than 10 points between dimensions should not be treated as meaningful.

Benchmarking

There are no benchmarks yet. Comparisons are published only when at least 30 organisations in a group have completed the Index. Until then, no comparison of any kind is shown.

No borrowed benchmarks. Published figures from McKinsey, Microsoft, Cisco, the ONS and others come from different questions, different samples and different definitions. They cannot be converted into a comparison with an Index score, so we never present them as one. Where they appear, they are labelled as context from that source.

Founding cohort. Organisations completing the Index before benchmarks exist are told so. They receive their benchmark position automatically once their group reaches 30 completions.

How benchmarks will be shown. Each comparison will state its group, sample size and date range. It will show the median and the middle 50% range rather than a single average. The first benchmark will cover the whole UK cohort, and sector and size groups will follow as each reaches 30. Financial services is the first focus sector.

One response per organisation. Benchmarks count one Leader Index per organisation (the most recent), so organisations with several respondents do not count twice.

Validation

Validation has not yet started. The plan below sets out what will be done, in what order, and what will be published at each step. Results are published whatever they show.

Limitations

The Index is a structured self-assessment at an early stage. Its scores should be read as a starting point for discussion, not as a validated measurement.

• One person's view. The Leader Index reflects one respondent, and leaders tend to see their organisations more positively than staff do. The Team Diagnostic is designed to address this.
• Self-report. Answers depend on the respondent's honesty and knowledge. Described levels and the “Don't know” option reduce optimism and guesswork, but do not remove them.
• Not yet validated. The dimensions, questions and stage cut-points have not yet been tested statistically or against outcomes.
• No benchmarks yet. Comparisons will not exist until groups reach 30 organisations, and early samples will not represent all UK organisations.
• Associations, not causes. The research behind the framework shows that certain practices go together with better results. It does not prove that adopting them will cause better results.
• A fast-moving field. Tools, regulation and practice are changing quickly, particularly with AI agents. The framework is reviewed at least once a year.
• Not legal or regulatory advice. The Index does not assess compliance with UK GDPR, FCA rules, the EU AI Act or any other law. Organisations should take their own legal and compliance advice.

Independence and data protection

The Performance Project receives no funding from AI vendors, technology providers or consultancies. It earns its income from diagnostic services and publishes only aggregated, anonymised data.

Independence. No vendor, sponsor or partner sees results before publication or influences what is published. If that ever changes, we will disclose it on this page.

What we collect. Respondent name, work email, role and organisation details, answers, and consent choices. Staff taking part in a Team Diagnostic answer anonymously.

How we use it. To produce your results and report, to contact you if you have agreed to it, and, in anonymised form, to build benchmarks and research findings. We do not sell respondent data, and we do not use it to train AI models.

What we publish. Only aggregated results. We never report a group of fewer than 30 organisations, and no individual organisation can be identified.

Retention and your rights. We keep contact details for up to 24 months after your last interaction, or until you withdraw consent. Anonymised answers are kept for benchmarking. You can ask to see or delete your data at any time.

References

Figures were checked against the source pages on 7 October 2026.

Version history

Questions about the methodology?

For methodology questions, research partnerships or due diligence, contact contact@theperformanceproject.tech.