Fundamentals

Problem 3: Your data isn't ready for AI (and it doesn't need to be perfect)

26 October 2026 · 6 min read · Hidden Node

Ask any data team about AI and you'll hear a sigh before you hear an answer. They know where the bodies are buried: the customer table with four spellings of "Hyderabad", the spreadsheet that's somehow the system of record, the PDFs nobody has opened since 2019.

Gartner puts a number on the risk. It predicts that through 2026, organizations will abandon 60% of AI projects unsupported by AI-ready data. In the survey behind that forecast, 63% of organizations either didn't have, or weren't sure they had, the right data management practices for AI.

That sounds like a reason to spend two years cleaning data before touching AI. It isn't.

"AI-ready" is about the use case, not the whole company

Here's the part that gets lost. Data doesn't need to be perfect everywhere. It needs to be good enough for the specific job you're asking AI to do. In our view, "AI-ready" is a property of a use case, not of a whole company.

A claims-triage assistant needs claim forms, policy wordings and historical decisions. It doesn't care that the marketing team's CRM is a mess. Scope the data to the use case and the problem shrinks from "fix everything" to "fix these four sources".

The questions that actually matter

For your chosen use case, answer these honestly:

1. Can we get to it? Where does the information live, and can a system read it without a six-month integration?
2. Is it representative? Does it include the messy, unusual cases, or only the tidy ones?
3. Do we know what "correct" looks like? Do you have past cases with known right answers you can test against?
4. Are we allowed to use it? Is there personal data involved, and do you have a lawful basis and consent to process it this way?
5. Will it stay current? If the policy changes next month, how does the AI find out?

If you can answer these for one workflow, you're more "AI-ready" than most companies, whatever your data maturity score says.

The good news about messy data

Modern language models are surprisingly good at reading the unstructured stuff that older systems choked on: scanned forms, email threads, handwritten notes. A lot of valuable early projects start precisely there, pulling structured information out of messy documents.

The catch is that you have to check the output. Extraction accuracy should be measured on a sample of real documents before anyone relies on it, and low-confidence results should go to a person.

A practical way to start

• Pick one workflow and list every data source it touches.
• Pull a random sample of a few hundred real cases. Not the best ones, random ones.
• Have someone who does the work mark the correct answers. That becomes your test set.
• Fix only what blocks this use case. Write down everything else for later.

Data quality is a journey with no finish line. You don't need to reach the end before you start. You need to know which stretch of the road your first project runs on.

Sources

• Gartner, Lack of AI-Ready Data Puts AI Projects at Risk, press release, 26 February 2025