AI inherits the conditions around its data

An AI system can search and summarise information, but it cannot make an outdated policy current or resolve a conflicting business definition by itself. If the source material is incomplete or unclear, an answer may sound convincing while missing the context a user needs.

This does not mean every record must be perfect before experimentation. It means the team should know which data matters to the use case and what quality issues could change the result.

Quality is specific to a purpose

For one use case, freshness may matter most. For another, completeness, consistent identifiers or clear ownership could be more important. A broad data-cleaning programme can consume time without improving the task being tested.

Describe the output first, then trace the input information and decide what must be checked. Use a small representative sample to expose missing values, duplicates, stale records and conflicting definitions.

  • Is the source current enough for the decision?
  • Can records be linked consistently across systems?
  • Are important fields complete and valid?
  • Is the source authoritative and owned?
  • Can users see where an answer came from?

Retrieval systems need useful source material

Knowledge assistants commonly retrieve passages from approved documents and use those passages to draft a response. The quality of retrieval depends on content structure, metadata, access permissions and how documents are maintained.

A pilot should test not just whether the system finds an answer, but whether it finds the right source, respects the user's access and recognises when the answer is missing. Outdated or duplicate documents need an owner and a removal process.

Measure quality alongside usefulness

Evaluation should include examples where the system must decline, ask a clarifying question or route work to a person. Track whether information is supported by the source, whether a reviewer can verify it and how often exceptions occur.

For high-impact work, keep human judgement and existing approval controls. The quality of the workflow includes how people catch and resolve errors, not only how often an automated result looks correct.

Improve the smallest important part

If a data issue blocks a promising use case, assign an owner, fix a narrow slice and repeat the evaluation. This links data investment to a concrete business need and makes progress easier to demonstrate.

A quality baseline is a practical foundation for trustworthy reporting and responsible AI. It does not require buying a large governance platform before learning what the business needs.

All insights