Sign Up
Sign Up

AI product investigation

A regenerate, a thumbs-down, or an abandoned answer shows the failure. It does not say what the person needed instead.

When this applies

Use this when the product under study is itself an AI feature or an agent, and product analytics already show a failed or repeated outcome. The connector is https://mcp.usercall.co. The entry tool is create_study. The shared loop is on Ask users why. The agent reading this page can investigate the product. It cannot interview itself.

Signal

Keep the event and the rate. Regenerating, rating down, copying nothing, and leaving mid-answer are separate signals. A support ticket that says "the AI was wrong" is the same kind of clue as a thin poll: it names a failure and skips the cause.

Example: 22% of research-summary generations are regenerated within two minutes (last 14 days, 3,400 generations, 748 regenerations). Sessions that regenerate twice end without a copy or export.

Unknown

The open question is why the first summary was not usable. The task may have been misunderstood, the answer may have been wrong, the person may not have trusted it, or the next step may have been hidden. The regeneration rate has not separated those.

Interview affected users

Interview people who regenerated a research summary or left without copying it. Ask what they expected the first answer to do. Pass the rate in business_context so the conversation does not spend itself reconstructing the metric.

create_study({
  key_research_goal: "Why was the first research summary not usable? Ask what they wanted it to do, what was missing or wrong, and what they did after regenerating.",
  business_context: "AI research-summary feature. Last 14 days: 748 of 3,400 generations were regenerated within two minutes (22%). Sessions that regenerate twice end without a copy or export. Interview people who regenerated or left without copying.",
  target_interviews: 8,
  duration_minutes: 12
})

Share interview_link with that group. Prompt logs and traces stay the record of what the model produced. The study is the record of what the person was trying to do.

Evidence

Poll get_study_status until complete, then read the summary and set it beside the 22% rate.

get_study_results({
  study_id: "<study_id>",
  format: "summary"
})

A theme such as "they wanted citations before the prose" is evidence you can act on. It does not rewrite the regeneration rate. Use format: "full" when you need the words they used.

Stop here

If the trace shows a timeout, an empty retrieval, or a schema error, that defect is the finding. Interview users after the answer is actually being produced and still being rejected.