There is a moment in almost every AI project when the excitement fades. It usually arrives about three weeks in, when someone opens the actual data.
The contracts are scanned PDFs with handwriting in the margins. The customer records have four spellings of the same company. The product catalogue lives in three systems that disagree. The policy documents exist in seven versions, and nobody is sure which one is current.
The model was never the problem. The data was. That is what AI data readiness is really about.
The numbers behind the feeling
Two recent studies put figures on what many teams already suspect.
Analyst firm BARC surveyed 225 data, analytics and AI leaders across North America and Europe for its 2026 study on unstructured data. 70% said less than half of their unstructured data is discoverable and usable for analytics or AI. About a third have inconsistent or missing controls for data bias and lineage, and about a third cannot consistently enforce data access and usage policies (IDM, reporting BARC).
Meanwhile, Info-Tech Research Group found that 40.9% of IT leaders rank improving data governance among their top data priorities for 2026, above AI-specific initiatives. As its analyst Pooja Khandelwal put it: "Without these data foundations, even the most advanced AI and IT initiatives will stall or fail to deliver value" (Info-Tech).
In other words, the people closest to the problem have reached the same conclusion: data first.
Why AI makes data problems louder
Bad data has always been expensive. A dashboard built on poor data gives you wrong numbers, but at least a human reads them with some scepticism.
AI removes that buffer. A language model reading out-of-date policy documents will confidently give out-of-date answers. An agent working from duplicated customer records will act on the wrong one. A model trained on biased history will repeat that bias at scale. AI does not fix data quality problems. It amplifies them and puts them in front of customers.
There is also a confidence gap. BARC found four in five respondents believe they can extract value without compromising governance, even though around a third lack consistent controls. That is a risky combination.
A readiness check you can run this month
You do not need a two-year data strategy to know whether you are ready for a specific AI use case. Pick the use case, then ask five questions about the data it depends on.
Can we find it? Is the data catalogued, or does it live in shared drives, inboxes and people's heads? If a new team member could not locate it in a day, an AI system will struggle too.
Can we trust it? Is it accurate, complete and current enough for the decision being made? BARC found accuracy, consistency and business relevance are the most common ways organisations judge readiness. Measure them for your data, not in general.
Do we know where it came from? Lineage matters most when something goes wrong. If an AI output is challenged, can you trace it back to source?
Who is allowed to use it, and for what? Access controls, consent and retention rules need to apply to AI systems as strictly as to people. If you cannot enforce them, you cannot safely let an agent near the data.
Who owns it? Every dataset an AI system depends on needs a named owner who is accountable for its quality. No owner, no production.
If the answers are mostly yes, build. If they are mostly no, the data work is the project, and it is a far better investment than a pilot that will stall.
Skills are the hidden bottleneck
Interestingly, BARC found the top barrier to using unstructured data for AI is not technology at all. Skills and training (36%) rank ahead of data privacy (30%) and data quality (29%). Info-Tech's priorities point the same way, calling for data champions and wider data and AI literacy.
Tools can classify, validate and index data. People still need to decide what "good" looks like, own it and keep it that way.
Foundations are not a delay
It can feel like data work slows AI down. In practice, it is what makes the second, third and tenth AI use case fast. The organisations BARC classes as mature, cross-functional AI adopters, around 23% of the market, are not the ones with the most models. They are the ones whose data is ready for whatever model comes next.
If you want an honest view of whether your data can support the AI you are planning, our AI & Data practice runs focused data readiness assessments. For the step after, see our guide to evaluating AI systems before they go live.
Frequently asked questions
What does AI-ready data mean?
AI-ready data is discoverable, accurate, current, well-governed, traceable to source and owned by someone accountable for its quality, so it can be safely used by AI systems.
How much unstructured data is ready for AI?
In BARC's 2026 study, 70% of organisations said less than half of their unstructured data is discoverable and usable for analytics or AI.
Should we fix data before starting AI projects?
Fix the data a specific use case depends on before building it. Broad data programmes are useful, but targeted readiness for the first use cases delivers value faster.