Skip to content

Observability

Insights reports

Have SentientOne read up to 200 recent conversations and report what people ask about, where the answers fall short and what to change, for 15 credits a report. Weekly reports, and trying a suggested fix in an experiment.

Guide3 min readUpdated 31 October 2026

An Insights report reads your recent conversations for you. It groups what people ask into topics, finds where the answers fall short, and suggests fixes, sometimes with a ready-made change to an agent's instructions. Insights is under Evaluate in the left menu.

Before you start

  • Conversations in the period you want to look at.
  • 15 credits on your balance.

Steps

  1. 1

    Open Insights

    Click Insights under Evaluate in the left menu.

  2. 2

    Choose what to read

    Under Generate a report, pick Agents (All agents or one; the agent in the top bar is picked for you) and the Period: Last 7 days or Last 30 days.

  3. 3

    Generate it

    Choose Generate report. The report opens straight away and fills in, usually within a minute or two. You can leave the page; the list shows Writing… until it's Ready.

  4. 4

    Read it

    Start with the Summary, then Issues: they're the most serious first, each with a suggested fix and example conversations.

  5. 5

    Try a fix

    When an issue comes with a change to an agent's instructions, Try it in an experiment opens a new experiment with two variants, Current and Suggested change. Pick the dataset and evaluators and start it. See Run an experiment.

You're done when

The report shows Ready with its topics and issues, and its Cost tile says 15 credits.

What a report contains

SectionWhat's in it
SummaryTwo to four sentences on what people asked and how the answers did
TopicsWhat people asked about, largest first: up to 10, each with its sentiment (Positive, Neutral, Negative or Mixed), how many conversations and what share, a description and example conversations
IssuesWhere the answers fell short, most serious first: up to 8, each with a severity (High, Medium or Low), how many conversations, a description, a suggested fix and examples
Next stepsUp to 6 things to do, in order

How it reads your conversations

  • It looks at the newest 2,000 conversations in the period and reads up to 200: those with a thumbs down or an average score under 50 first (up to 100), the rest spread across the period.
  • Each one goes in as its first question, final reply, thumbs and average scores. All agents includes workflow conversations.
  • It reads Live traffic, and falls back to Playground chats only when the period has no Live conversations; the report then says so.
  • The reading is done by SentientOne's own Claude model, so your LLM key isn't used or billed.

Credits

  • A report costs 15 credits, charged only when it's ready. A report that fails, can't be finished or finds no conversations costs nothing, and says "You weren't charged."
  • With fewer than 15 credits, Generate report is switched off. Deleting a report doesn't refund it.
  • One report is written at a time per account, and up to 6 an hour.

Weekly reports

Switch on Weekly report for … and choose the day (Monday by default) and the hour (08:00 by default). The time zone is your browser's when you first save it. Each weekly report covers the 7 days before it and costs 15 credits; it's skipped, with no charge, when your balance is short or the week had no conversations. There's one weekly report for all agents and one for each agent. No email is sent: weekly reports appear in the list as Made by "Weekly schedule".

If something goes wrong

"There are no conversations in the last 7 days to analyse."

Choose Last 30 days, or All agents. You weren't charged.

"A report is already being written."

One report is written at a time. It opens that report; wait a minute or two.

"The report came out too long to finish."

Try one agent or 7 days. You weren't charged.

Still stuck?

Email us with what you tried and any request id. A person replies.

support@sentientone.ai