AI-Ready Data Needs an Owner, a Meaning and an Access Boundary

A customer count can mean active contracts, billed accounts or unique legal entities. Retrieval may find a number for each definition. A useful data product identifies which meaning applies, the source of authority and the person who can resolve disagreement.

For a hypothetical sales assistant, define whether pipeline excludes closed deals, how currency is converted and when the data was refreshed. Carry those definitions and the effective period with the retrieved evidence. A plausible total without its basis can mislead a decision-maker.

Set an owner and a correction route for the underlying product. When users report an answer problem, classify whether the source was wrong, the transformation changed its meaning, retrieval missed the right record or generation misrepresented it. Each failure needs a different fix.

Keep document identity, access metadata and version information connected through ingestion, transformation, indexing and retrieval. Enforce the caller's permissions before restricted content enters model context. A prompt that asks the model to hide confidential information cannot substitute for authorization.

Microsoft's Azure AI Search documentation describes document-level access approaches, including security filters and native permission features, some marked preview. This illustrates the architectural need; it does not mean every connector preserves source permissions automatically or every feature is production-ready for a given deployment.

Test permission changes as well as initial access. If someone leaves a group or a source document is withdrawn, verify the behavior of the index, answer cache, conversation history and generated summaries. Define the maximum delay for updates and revocation, and withhold affected responses when that expectation cannot be met.

Define a small service contract: intended users, supported questions, source owners, freshness targets, access rules and escalation contact. Include questions the product cannot support. Clear limits help users distinguish a documented answer from a request that still needs a person.

Build evaluation examples for a stale policy, conflicting definitions, an absent source and a user who lacks access. Check citations against the exact retrieved version and confirm that an unsupported question produces an understandable abstention or escalation. Keep test data synthetic or approved for the environment.

Review answer failures with the data owner and the application team. Prioritize the sources that affect real decisions instead of connecting every repository at once. A narrower product with known meaning, permission behavior and maintenance responsibility gives a team a stronger basis for expansion.