A promising AI use case can fail because the organisation cannot reliably find, interpret or use the data it needs. Buying a model does not resolve unclear ownership, inconsistent definitions or missing permissions.
A data readiness assessment should answer a practical question: can this specific workflow use these specific data sources safely and reliably? You do not need to modernise every dataset before you can make a useful decision.
1. Start with the decision, not the data lake
Describe the outcome you want to improve and how it is measured today. For a service team, that might be finding an approved answer faster while maintaining accuracy, rather than simply introducing a chatbot.
List the information needed for that outcome and the minimum acceptable freshness. This bounds the assessment and prevents a small operational problem from becoming an open-ended platform procurement.
2. Trace ownership and permissions
For each source, identify who owns its content, who can authorise access and whether the proposed use is permitted. The ability to view a document does not automatically grant permission to train a model on it or send it to an external service.
Record sensitive fields, retention constraints, residency requirements and supplier access. Resolve restrictions before building a prototype that depends on data you may not be allowed to use.
- Who maintains the source of truth?
- Can access be restricted at the record or document level?
- Can information be corrected or removed from downstream systems?
3. Test quality against the intended task
Inspect a representative sample for missing values, duplication, conflicting definitions and outdated records. Include different teams, time periods and languages where they matter to the workflow.
Quality is task-specific. A document can be readable but still lack the version date or approval status needed for a reliable answer. Track defects by their effect on the decision, not only by the number of empty cells.
4. Prove a repeatable path to the data
A one-off spreadsheet export can demonstrate an idea, but it does not prove that a production service can remain current. Test how the data is retrieved, validated, updated and logged.
For document-based AI, check extraction quality, permissions and whether answers can point to an authoritative source. Test an update or deletion end to end so you know how long stale information remains available.
5. Produce a bounded improvement plan
End the assessment with a clear decision: proceed with controls, repair named gaps first, narrow the use case, or stop. Attach an owner and an acceptance test to each priority improvement.
The most useful output is not a maturity score on its own. It is an agreed route from today's data to a workflow that can be operated and trusted, with explicit gates before further investment.
- Separate essential fixes from desirable improvements.
- Agree who pays for maintenance after launch.
- Reassess readiness when the use case or data source changes.
General practical guidance, not legal advice. Adapt these steps to your organisation and seek appropriate specialist advice.