Insights / Checklist / Data foundations and advanced analytics
ChecklistData readiness for AI: the checklist
A practical checklist to test whether your data is ready for AI, covering ownership, quality, access, documents and testing, with a simple way to score the result.
Every AI system is only as good as the data it learns from, reads or acts on. Yet data readiness is often assessed last, after a tool has been chosen and a pilot has started. By then, problems with quality, access or definitions show up as disappointing results, and the AI gets the blame.
Gartner reported in 2025 that 63% of organisations either do not have, or are unsure whether they have, the right data management practices for AI, and predicted that through 2026 organisations will abandon 60% of AI projects that are not supported by AI-ready data. Gartner has also reported that 59% of organisations do not measure data quality at all.
Use the checklist below before committing to an AI use case. Work through it with the people who own and use the data, not only the technology team. Answer each question honestly for the specific data your first use case will depend on, rather than for the organisation as a whole.
Ownership and definitions
Ownership comes first because every other fix depends on it. Without an owner, nobody has the authority to agree a definition or the responsibility to correct errors.
- Every important data set, such as customers, products, suppliers and employees, has a named owner in the business.
- There is a clear place to report data errors, and someone acts on the reports.
- Definitions are agreed and written down for key terms such as customer, revenue, order and active.
- Different systems use the same definitions, or the differences are documented.
A quick way to test this area is to ask three managers from different functions how many active customers the business has. If you receive three different numbers, definitions are the first thing to fix.
Quality
Quality problems that people work around every day become serious when an AI system relies on the data without the same local knowledge.
- You measure completeness, accuracy, duplication and timeliness for the data sets that matter.
- You know the error rate in the specific data an AI use case would rely on.
- There is a process to fix errors at the source, not only in reports or extracts.
- Quality measures are reviewed regularly by the data owners.
If you have never measured quality, start small. Take a random sample of a hundred records from the data your use case needs and check them by hand against the source. The share of records with a missing, wrong or duplicated field is a rough but honest baseline.
Access and privacy
AI projects often stall while teams wait for access, or rush ahead by copying data into spreadsheets and unmanaged tools. Neither outcome is acceptable.
- The right people can reach the data they need without copying it into spreadsheets or personal storage.
- Personal and confidential data is classified, so it is clear what may be used for which purpose.
- Retention periods and consent rules are documented and applied.
- Access is granted by role and reviewed periodically.
Content for assistants
Assistants that answer questions from your policies, product information or procedures depend on documents rather than databases. Outdated or contradictory documents produce confident, wrong answers.
- The documents an assistant will read are current, approved and well organised.
- Each document has an owner and a review date.
- There is a way to retire outdated versions so they are no longer used.
- Conflicting documents on the same topic have been identified and resolved.
Testing
Without a test set, you cannot tell whether an AI system is accurate, or whether a change has made it better or worse.
- You have a test set of real examples with known correct answers, including difficult and unusual cases.
- You can measure results before and after changes to the model, prompts or data.
- The people who will use the system have helped define what a correct answer looks like.
If you cannot measure data quality today, your first AI project is a data project.
What to do with the score
Score each of the five areas from one to five, where one means the practice is absent and five means it is established and reviewed. Do this for the data behind your first use case, not for every data set in the organisation.
Then fix the lowest-scoring area that matters to that use case before building. A low score in content for assistants matters a great deal for a policy assistant and very little for a demand forecast. A low score in ownership matters for almost everything. Re-score after the fix, and only start the build when no relevant area sits below three.
Share the scores with the sponsor of the use case. They make the case for data investment far more clearly than a general appeal, because they link a specific weakness to a specific outcome the business wants.
Source: Gartner, February 2025; Gartner data quality research.
