An internal assistant has a thumbs-up and thumbs-down button under every answer. After two months, there are 300 thumbs-downs and nobody knows what to do with them. Some are wrong answers, some are slow answers, and some are correct answers the person did not like.

Feedback is only useful if someone can turn it into a fix. That depends on what you collect and what happens to it next.

Ask for the reason, in one tap

When someone marks an answer as unhelpful, offer a few short options:

  • Wrong or outdated
  • Missing information
  • Couldn't find the source
  • Too long or unclear
  • Should have asked a person

Add an optional comment box, but do not require it. Most people will tap a reason; few will write a paragraph.

Keep the context with the feedback

A thumbs-down without the question and answer is almost impossible to act on. Store the feedback alongside the request, the answer, the sources used, and the version of the assistant's instructions. Our guide to keeping a change log for AI instructions explains why the version matters.

Agree on how long that record is kept and who can see it, especially if questions may contain customer or employee details.

Review a sample every week

Set aside thirty minutes a week for one owner to read recent negative feedback and a few positive examples. Sort each item into one of four next steps:

  1. Fix the content: a document is missing, wrong, or out of date. Send it to the team that owns it.
  2. Fix the assistant: instructions, retrieval, or formatting need to change. Add the case to your test set first.
  3. Route to a person: the question should not be answered by the assistant at all.
  4. No change: the answer was correct. Note it and move on.

Count which category grows. A rising number of "fix the content" items is a documentation problem, not a model problem.

Close the loop with staff

People stop giving feedback when it seems to disappear. Share a short monthly note: the most common problems, what was fixed, and what is still open. When someone reports a specific issue, tell them when it is resolved.

Watch for silence too

Low feedback does not always mean high satisfaction. Check usage alongside it. If a team stops using the assistant, ask them why directly. Our AI rollout guide covers how to run those check-ins.

To set up a feedback process for an assistant you already run, email Storm Nova Labs.