BlogSupportSystem statusEspañolHosted in the USA

Accessibility & trust

AI call summaries and sentiment: what they are and aren't

After every call, SentraOperator writes a short summary and notes whether the caller seemed positive, neutral or negative. These are useful for review, and easy to misunderstand. This post explains what they are, how they are made and what they are never used for.

What gets written after a call

When a call ends, an AI model reads the transcript and produces four things:

  • A summary of two to three sentences, such as "Caller reported a streetlight out at 200 Elm St. Request SR-1185 filed. Caller asked for the number by text."
  • A sentiment label: positive, neutral or negative.
  • A reason for that label, such as "caller was frustrated that an earlier report had not been fixed."
  • Any follow-up the call suggests, such as a promised callback.

These appear next to the call in the supervisor console's call log, alongside the transcript and recording.

How the transcript is cleaned first

Before the model sees anything, phone numbers, card numbers, account numbers and email addresses are removed from the copy of the transcript it reads. The summary is written from that cleaned copy.

The agent never asks for card or bank numbers in the first place. If a caller reads one out anyway, it is redacted from the transcript. The provider that writes summaries is listed on our sub-processors page and does not use the data to train its models. See our security page for details.

What summaries and sentiment are good for

Summaries save time. A supervisor scanning yesterday's calls can read a line or two for each one instead of playing recordings from the start. A few common uses:

  • Morning review of after-hours messages. Read the summary, then decide who calls back.
  • Finding a call. Summaries are searchable along with transcripts.
  • Spotting patterns. If several summaries mention the same street, that is worth a look.

When a call is transferred, staff also hear a one-line spoken summary, the "whisper", before the caller is connected. That is a separate, live feature. See talk to a person.

Sentiment for review

Sentiment helps supervisors decide where to look first. The reports in the supervisor console show sentiment over time, top request types with the share marked negative, and a list of the unhappiest calls to review.

That list is often the most useful page in the console. Negative calls tend to point at real problems: an FAQ answer that is wrong, a transfer number nobody answers, or a request type with too many questions. Fixing those helps every caller after.

Sentiment describes how a call went. It does not describe the caller.

What they are not

This is the part that matters most for a public agency.

  1. They can be wrong. A summary may miss a detail. A sarcastic "great, thanks" may be labeled positive. Treat both as a starting point, and check the transcript or recording before acting on anything important.
  2. They are not the record. The transcript and recording are the record of what was said. A summary is an aid for reading it.
  3. They never decide anything about a caller. Sentiment does not change how a request is prioritized, whether a call is transferred, or how a resident is treated. It is not a score on the person.
  4. They are not used to rate individual staff. The labels describe the overall call, which is mostly the AI part, and are meant to improve settings and content.

If your city adopts written guidance for staff, say these things plainly. It helps to have it on paper before anyone asks.

Records, and what is coming

Summaries and sentiment are stored with the call, so they follow the same retention schedule as the rest of the call log. Whether they are part of a public record is a question for your records manager and city attorney, since public records laws vary by state. Our post on public records requests and transcripts covers the export tools.

Coming later: quality review

A quality review feature is on the roadmap for 2027. It is planned to score calls nightly and draft suggested fixes, such as a new FAQ entry, for a supervisor to approve. Nothing would change without that approval. It is not available today.

The short version

AI summaries and sentiment are reading aids for supervisors. They are written from a transcript with sensitive numbers removed, they can be wrong, and they are never used to decide anything about a caller. Use them to find calls worth reviewing, then check the transcript before you act.

Read about our safety rules