Ready for AI, or just adopting it?

What is an AI readiness assessment?

An AI readiness assessment is a structured evaluation of an organisation’s data, infrastructure, governance, talent and strategy that answers one question before budget is committed: can we adopt and sustain AI in this part of the business, or not yet. It is scored against evidence rather than opinion, and it ends in a go, no-go or fix-first decision with a named owner against every gap. Anything that ends in a summary deck is a different document.

The case for running one is not diligence theatre. AI project failure is well documented and concentrated in exactly the causes a readiness assessment is built to surface early: data that was never fit for training, governance that was never written down, and accountability that was never assigned to a person.

How wide is the AI readiness gap?

Wider than most boards assume, and it sits between adoption and readiness rather than between adopters and holdouts. 96% of organisations are implementing AI models while only 2% rank themselves as highly ready to meet AI deployment demands, and only 7% of enterprises say their data is completely ready for AI adoption.¹

The pattern repeats across independent surveys. 88% of organisations report using AI in at least one business function, up from 78% a year earlier, yet only 1% consider their AI strategies mature.² Adoption is now close to universal. Readiness is a minority position, and the distance between the two is what an assessment exists to measure.

Part of the gap is comprehension rather than capability. 65% of leaders do not know when or where to apply AI, and 52% lack a foundational understanding of how AI works.³ That has a direct consequence for how the assessment is run. If the people scoring readiness cannot describe what the technology actually does, the score measures confidence, not capability. Put someone who has taken a model into production, internal or through AI consulting support, in the room to argue with the self-scores.

How do you know your data is actually ready for AI?

Data readiness is not volume. AI-ready data is accurate, complete, properly labelled, lineage-tracked and governed well enough to feed a model without introducing bias, compliance exposure or unreliable output. A warehouse can be large, current and entirely unready on every one of those tests.

The evidence for weighting data heaviest is unusually direct. RAND estimates a historical AI project failure rate above 80%, roughly twice the rate of non-AI IT projects, with poor data quality cited in 85% of those failures.⁴ Gartner expects organisations to abandon 60% of AI projects through 2026 because the underlying data was never AI-ready.⁵

The uncertainty inside that figure is the finding. In a Q3 2024 survey of 248 data leaders, 63% either did not have the right data management practices for AI or were unsure whether they did.⁶ Not knowing scores the same as not having, because both mean nobody can answer the question in the review. Confidence is scarce elsewhere too: only 26% of organisations feel certain their data can support AI-enabled initiatives.⁷

Four diagnostics separate a genuine data readiness check from a data inventory.

• Completeness at field level. Not row counts. What share of the specific fields a model would consume are populated, in range, and current for the period being trained on.
• Lineage. Whether any given value can be traced back to the system that produced it and the transformation that changed it, without asking the one analyst who remembers.
• Labelling. For supervised work: who labelled it, against what definition, and how far that definition has drifted since.
• Permission. Whether the terms under which the data was collected allow it to be used for training or inference at all.

What should an AI readiness assessment score?

Six dimensions cover the ground. The weighting is where most published frameworks go wrong: they score infrastructure and tooling heavily because those are the easiest things to observe, when the failure evidence points at data and governance instead. The weights below are our recommendation, not an industry standard, and they are deliberately lopsided.

Score each dimension per use case on a short scale, with evidence attached to every score. A dimension nobody can evidence scores zero, not average. The output is not a number for the deck. It is a ranked list of what has to change and who owns each item.

An assessment that does not end in a go, no-go or fix-first decision with a named owner against every gap is not an assessment. It is a shelf document nobody reopens before the next pilot.

What an AI maturity model shows, and what it hides

An AI maturity model is a staged framework, commonly four or five levels running from ad hoc experimentation to fully embedded and governed AI operations. It does one job well: it shows a board the distance between where the organisation sits today and the next level, in language that does not need a data engineer to interpret. The benchmark is sobering. Fewer than 25% of large enterprises operate at Level 4 or above.⁸

What a maturity model hides is variance. Maturity is a property of a use case, not of a company. A firm can be genuinely mature in customer service automation and immature in finance forecasting, because those two run on different data, different owners and different regulatory exposure. A single company-wide score averages them together and buries the one place where the risk actually sits. Score by use case, then report the spread rather than the mean.

That number is worth sitting with. The most AI-ready industry measured scored 34 out of 100.¹ Wherever your own benchmark lands, the comparison that matters is against what the use case requires, not against the sector average.

AI audit or AI readiness assessment: which do you need?

They run in opposite directions. An AI audit is retrospective. It examines systems already in production for bias, accuracy, compliance and control, and it produces findings. An AI readiness assessment is prospective. It examines whether the conditions for a system to succeed exist yet, and it produces a decision. Regulated organisations usually need both, on different cycles, with different owners.

The sequencing argument favours readiness first, because governance debt compounds faster than technical debt. Skipping a policy on data use, model monitoring or human review at the pilot stage means retrofitting it later under audit pressure, when the system is live, the training data is already inside it, and the cost of change is at its highest.

What do you fix first when the score comes back low?

Not the tooling. The most common root cause found in follow-up interviews after a low score is an ownership gap rather than a capability gap: no single accountable person for data quality in the domain the use case depends on. Naming that owner costs nothing, takes a week, and changes what every subsequent fix is measured against.

A workable order of operations:

1. Name one accountable owner per data domain. A person, not a committee, with the standing to reject a use case that depends on data they cannot vouch for.
2. Write the governance policy before the pilot. Data use, model monitoring, human review, escalation. Two pages in force beats twenty pages in draft.
3. Fix data at field level, for one use case only. Organisation-wide remediation programmes tend to outlive the sponsors who fund them. Repair the fields one bounded use case actually consumes, and prove the pattern there.
4. Re-score before you re-fund. Run the same rubric at the next gate so movement is visible and comparable rather than asserted.

The organisations that move from experimentation to embedded operation are rarely the ones that bought the most capable model. They are the ones that closed the ownership and data gaps first, then chose a use case narrow enough to prove it. That is the same discipline that separates automation that stays switched on from automation that quietly gets turned off, and it is where most AI transformation programmes are won or lost. If you want a second pair of eyes on your scoring before the next funding gate, talk to us.

Written by the group's editorial team with the practice leads who run these builds. Reviewed before publish. Spotted an error? Tell us and we will fix it.