AI & knowledge

Designing knowledge assistants people can trust

Useful answers begin with clear sources, permissions, and uncertainty handling.

Identify the source of truth

An assistant should know which documents and systems contain approved information. Keep document ownership and update routines clear.

Respect access boundaries

Retrieval should match the information a user is authorized to see. Permissions belong in the data and retrieval design, not only in the interface.

Make sources visible

Where possible, connect answers to the supporting material. References help people verify a response and investigate the original context.

Test the difficult questions

Include ambiguous requests, outdated material, missing answers, and conflicting information in evaluation. An assistant needs a useful way to say that the evidence is insufficient.

Plan ongoing operation

Documents and workflows change. Monitor quality, refresh content, review feedback, and account for model and retrieval changes over time.

Worked example: a supported answer and a useful refusal

An illustrative assistant asked about onboarding should retrieve the current approved guide, summarize the relevant steps and show the source. If two guides conflict, it should identify the uncertainty and route the question to the document owner. If the user lacks access to a source, the answer must not reveal its contents. Test each of these states alongside ordinary questions; source links alone do not prove a response is correct.

A conversation is a good place to start

Have a similar question?

Let’s connect the idea to your business and the next practical step.

Discuss your project