What information an AI agent needs and what should stay out

More information does not mean better answers

Loading the whole shared folder looks efficient, but it mixes versions, contradictions and data the agent never needs. Quality depends on choosing sources with an owner, a validity date and a purpose. Each document should say which decisions it can support and who maintains it.

Start from the questions the agent will handle and gather only the evidence needed. Turn real conversations into anonymised examples. Mark mandatory policies and also silences: topics where it must ask for help. A small, clean set is easier to test than thousands of ungoverned files.

Four layers of knowledge

The first layer is stable facts: location, hours, services and channels. The second holds operating rules: requirements, zones, notice periods and cancellations. The third is examples of good judgment: how to explain a difference or what to ask before quoting. The fourth is dynamic data that must be looked up in systems, not copied into a document, such as availability or payment status.

Separating layers lets you assign controls. Facts can be updated from an approved source; rules need versions; examples guide but do not replace policies; dynamic data needs authentication and validation. Without that separation, the agent can answer with an old policy when it should check the current state.

Privacy and least privilege

Before connecting a source, ask whether the task can be solved with less data. To qualify a prospect, a name, a need and a date range may be enough; you do not need to expose a full history. To answer a policy question, no access to payments is needed. Design permissions by action and by role.

Never use documents as a place to store credentials. Secrets live in a manager and are handed directly to the authorised system. Review retention, logs and provider access. If conversations are used for testing, remove names, phone numbers, emails, identifiers and details that would identify a person.

• A defined purpose for every source
• An owner and a validity date
• Minimum access
• Anonymised examples
• A process for correction and removal

How to keep the knowledge current

Assign a responsible person and a cadence. Urgent changes need a short path; normal changes can be reviewed in batches. Record who approved the new version and rerun the relevant tests. Updating a price should not change the cancellation rule without anyone noticing.

Measure unanswered questions and sources that cause conflicts. Those signals form the knowledge backlog. Do not train the agent to improvise when information is missing; use it to discover which part of the business still lives only in someone’s head.