Give your AI assistant better source material.

A company can have a current process, an older PDF describing it, and a message explaining the exception. Someone who has been there a while may know which one to follow. A new person has to ask around.

Putting an AI assistant in front of that information gives the team another way to search. It also creates a useful reason to settle what the information actually means. Before choosing the assistant's personality or refining its welcome message, decide what it should be able to rely on.

Decide what counts as current

Start with the documents people already use to answer recurring questions. Look for competing versions, missing context, and instructions that depend on someone's memory. An outdated document does not become more useful because an assistant can retrieve it quickly.

Give each important source a clear owner, a review date, and an explanation of where it applies. A date alone is not enough. A recently edited file may contain an old process, while an older policy may still be the approved one. Someone needs to make that distinction.

Keep historical material when it serves a purpose, but label it and separate it from the sources intended to guide current work. The aim is to make the answer to “Which version should we use?” easy for a person to establish before expecting software to work with it.

Choose a useful slice of the business

You do not need to organize every document in the company before starting. Choose a collection small enough for someone to understand and maintain, tied to a set of questions the team actually asks.

For a hypothetical service business, that could be an assistant helping employees explain the onboarding process to new customers. Its sources might include the current setup guide, the preparation checklist, the service description, and the contact responsible for exceptions. An old sales proposal with a special arrangement would need to be clearly separated from those general instructions.

This boundary makes missing information easier to see. If the setup guide assumes that everyone knows who sends the welcome email, write that down. If two documents disagree about what the customer needs to provide, resolve the disagreement with the person responsible for the process.

Keep access tied to the person asking

Decide who should be able to use each source and what the assistant should be allowed to reveal. Access to the chat interface should not automatically mean access to every document behind it.

The details depend on the tool and connection. Microsoft documents that its Copilot connectors respect permissions on the original sources, while also requiring the correct authentication settings for agents that use them. That is a specific capability to configure and verify, not a reason to assume every upload or connection behaves the same way. Microsoft's connector documentation explains that distinction.

Test with the kinds of accounts that will actually use the assistant. A successful answer for an administrator does not establish what an employee or external collaborator can see. Include questions that should return no information because the person asking lacks access.

Make the answer checkable

A useful answer should make it easy to inspect the material behind it. Show the relevant source, link to the document when possible, and preserve enough context for the reader to understand which process or situation the answer describes.

A citation is a starting point for checking. It is not proof that the answer is correct. NIST's generative AI risk profile describes how these systems can produce inaccurate content and fabricated citations. A confident response can still misrepresent the material it appears to rely on. NIST's Generative AI Profile discusses this risk.

For the onboarding example, the employee should be able to open the setup guide and confirm the requirement before telling a customer what to do. That becomes much harder when the assistant gives a polished paragraph with no way to trace it.

Give missing knowledge an owner

Define what should happen when a source is missing, contradictory, or too vague to support an answer. The assistant can identify the gap and direct the question to the appropriate person. It should not fill the space with a plausible company policy.

Keep a record of those unanswered questions and give someone responsibility for reviewing them. Some will need a better document. Others will reveal a decision the business has not made yet. Updating the source can then help both the assistant and the people who would otherwise have to ask the same question again.

If you want help turning your team's knowledge into something people can use, Work with me.

Sources

Microsoft's knowledge-connector documentation explains source permissions and authentication. NIST's Generative AI Profile, section 2.2, describes inaccurate outputs and fabricated citations.