- Data is ready or not for a single task, so the test starts by listing the fields that task reads and writes, then scoring a sample of real records on five checks.
- Gartner says AI-ready data must be representative of the use case, of every pattern, errors, outliers and unexpected emergence, which is why the awkward records stay in the test.
- Under UK GDPR you must identify the minimum amount of personal data you need to fulfil your purpose, so some fields fail the test because they shouldn't be used at all.
Data readiness for AI means one thing in practice: can the records a specific task depends on be found, trusted, matched across systems and lawfully used? You find out by pulling 50 recent records for that task and checking every field it touches, not by auditing the whole business. Most SMEs pass on some fields and fail on one or two. The failures decide whether you fix first or build around them.
The question is common. Gartner found 63% of organisations either lacked the right data management practices for AI or weren't sure, and predicted that firms would abandon 60% of AI projects unsupported by AI-ready data. That was a February 2025 forecast for large organisations, so read it as a direction rather than a measurement of small firms. The UK picture is closer to home. In the government's survey of 3,500 businesses with at least 5 employees, 23% of firms already using AI said data complexity held back wider use. Among large businesses it was 39%.
Ready for which task?
"Is our data ready?" has no answer on its own. A sales ledger can be perfect for chasing overdue invoices and hopeless for forecasting demand. Same records, different fields.
So write the task in one sentence first: "Match each supplier invoice to its purchase order and flag the differences." Then list every field it reads (supplier name, PO number, line amounts, VAT) and writes (match status, query note). That list is the scope of the test. No task chosen yet? The AI readiness checklist helps pick one, and an AI audit ranks several.
The 50-record test
Export 50 recent records for the task. Fifty isn't a statistical sample, just enough to show recurring problems while one person can still check it by hand. Include the awkward ones: the refund, the duplicate customer, the phoned-in order.
Then score each field on five checks.
| Check | What to look at | A typical failure | Fix first or build around? |
|---|---|---|---|
| Present | Is the field filled in on every record? | Delivery postcode missing on phone orders | Build around if a reviewer can fill the gap; fix first if the task can't run without it |
| Consistent | Same format and same labels each time | "Paid", "paid in full" and "PIF" in one status column | Build around: a model reads variants, but write the mapping down and test it |
| Current | Does the record describe today? | Account owner is someone who left last year | Fix first: the system will act on stale facts with confidence |
| Joinable | Can it be matched to the other system by a shared reference? | Invoices carry a customer name, the CRM keys on an account number | Fix first: matching on names creates errors nobody sees |
| Allowed | Is there a lawful basis, a contract that permits this use, and a real need for the field? | Full date of birth copied across when the task needs none | Fix first: drop the field or document why it's needed |
Two extra steps catch what the table can't. Try the join by hand on ten records: a link a person struggles to find will be hard for the build too. Then ask who corrects a wrong value today. If the answer is "the customer, by complaining", that field needs a person checking it at go-live.
Keep the mess in the sample
Don't tidy the export first. A cleaned sample passes a test the live system will never sit.
Gartner's definition makes the same point: the data has to be representative of the use case, of every pattern, errors, outliers and unexpected emergence. Those awkward records become the test set. They show where the system must stop and hand over. Our guide to intelligent document processing shows how to score accuracy against a test set like this, and shadow mode is how the same records get run alongside the people already doing the work.
Messy data also has a familiar shape. The government's interviews with 100 firms put it plainly: Fragmented, unstructured, or siloed data was a technical barrier to deploying AI effectively. Fragmented and siloed are the joinable check failing. Unstructured is a smaller problem than it sounds, since current models read emails and PDFs well.
What UK data protection asks of the data
The fifth check is the one firms skip, and it's the one a regulator asks about. UK GDPR's data minimisation principle requires you to identify the minimum amount of personal data you need to fulfil your purpose. The ICO's AI guidance is blunt about the temptation to connect everything: You must not collect personal data on the off-chance that it might be useful in the future.
In practice the field list from the first step doubles as your minimisation record. Anything the task doesn't read stays away from the model. The same guidance says that In the vast majority of cases, the use of AI will involve a type of processing likely to result in a high risk, which is the trigger for a data protection impact assessment. The ICO notes this guidance is under review following the Data (Use and Access) Act 2025, so check the current version before you sign anything off.
Fix, build around, or wait
Each failing field then lands in one of three piles.
- Fix first. Stale records, broken references between systems and fields you aren't allowed to use. No model can cover for them.
- Build around. Gaps a reviewer can fill, inconsistent labels and unstructured text. Mappings, extraction and an approval step absorb these.
- Wait. If the task needs history you never recorded, such as reasons for lost sales, no build can invent it. Start capturing the field now and come back when there's enough to test.
A practical rule for the first two piles: compare the failure rate in your sample with the error rate you'd accept from a new member of staff on the same job. Above it, fix first. Below it, build around and let the review step catch the rest. The guide to implementing AI in a business covers what comes after this decision, and why AI pilots fail shows what happens when it's skipped.
Common questions
What does AI-ready data mean?
Data that a specific AI task can read, trust, match and lawfully use. It isn't a property of a whole database. Invoices can be ready for payment matching while the same customer records are useless for churn prediction, so readiness is always judged against one task and the fields that task touches.
How much data do you need for an AI project?
Most SME projects run on an existing model, so they need no training set at all. What they do need is enough real records to test against: the sample used to check the fields, then a few weeks of live items handled in parallel before anything acts alone.
Do you need a data warehouse before using AI?
Rarely, for a first project. If the task reads from one or two systems with working exports or an API, the build connects to them directly. A warehouse starts to earn its cost when several AI tasks need the same joined-up records, or when reporting already suffers from scattered data.
Should you clean your data before an AI project?
Clean the fields that block the task: stale records, broken references and data you aren't allowed to use. Leave the rest of the mess alone, and keep it in the test set. The live system will meet those awkward records, so a test that never sees them proves nothing.
Is a data protection impact assessment needed for AI?
Usually, if the task touches personal data. The ICO's guidance says most uses of AI involve processing likely to result in a high risk, which is the legal trigger for a DPIA under UK GDPR. Starting it during the data test costs less than retrofitting it.
If you run the test, send us the result: the task in one sentence, the systems its records sit in, and which of the five checks failed on your 50 records. Use the contact form and pick Business automation. We'll have an informal scoping chat about the failing fields and then send a price range for the build.
Sources
Every figure in this article links back to the source below it was checked against.