Every answer shows its source
An agent that answers questions is only as useful as people's confidence in its answers. The most reliable way to earn that confidence is to make every answer traceable. Cite, then show. Each statement in an answer sh
An agent that answers questions is only as useful as people's confidence in its answers. The most reliable way to earn that confidence is to make every answer traceable.
Cite, then show. Each statement in an answer should link to a passage in an approved document. People should be able to open that passage in one step.
Answer from approved sources only. Decide which documents the assistant may use, and keep that list under the same control as the documents themselves. When a policy is updated, the assistant should use the new version.
Treat "no answer" as a feature. When no approved source covers a question, the assistant should say so plainly and route the question to the team that owns it. Test this behaviour as carefully as the answers.
Agree the boundaries with the people you serve. Whether the users are clients or colleagues, set out what the agent may and may not do for them, and review it together over time.
Report openly. Share regular summaries of:
- what agents handled;
- what people reviewed;
- what changed as a result.
This turns trust from an assumption into something you can show.
A short checklist:
- Can every answer be traced to a passage in an approved document?
- Do you test the questions the assistant should decline?
- Do the people you serve know what your agents may and may not do?
From the 2SD Technologies team (2sdtechnologies.com). Part 6, the final part, of "Don't just build agents".
Originally published by Dev.to Security. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.