Finding the Call That Went Wrong Before the Review Gets Posted

To detect unhappy customer calls without listening to everything, you need something that picks the calls worth hearing. Signal classifies every recorded inbound call and flags the ones where the caller had a bad experience, with a stated reason. Reviewing three flagged calls on a Tuesday is a habit. Reviewing four hundred recordings is a project that never happens.

The reason nobody reviews recordings

Almost every small office that turns on call recording reviews the recordings for about two weeks.

The math defeats them. A practice taking 150 inbound calls a week produces roughly 12 hours of audio. Sampling randomly means you mostly listen to calls that went fine, which is both boring and uninformative, and the one call that lost a customer sits in the 97% you skipped. So the tab stops getting opened, and the recordings become an archive you only touch when somebody complains — which is to say, after the damage.

The bad call is not hard to learn from. It is hard to find. That is a retrieval problem, and retrieval problems are the kind software is actually good at.

What is red-flag call detection?

Signal runs each recorded inbound call through transcription and then a single model pass that produces the summary, a sentiment classification, and a red-flag decision with a written reason.

A red flag means the model judged that the caller had a bad experience on that call. The reason is stated in a sentence — the caller was transferred repeatedly, the caller said they had called three times already, the caller ended the call without getting an answer — so you know why the call surfaced before you spend six minutes listening to it.

Flagged calls land in a review queue with two states, needs-review and reviewed, and Signal emails the flags to your account's users as they happen. They also appear in your weekly call report, so a week where nobody logged in still ends with somebody seeing them.

Which calls tend to get flagged

Patterns we expect you to see, in rough order of how often they matter:

1. The repeat caller. Somebody says "this is the third time I've called about this." That sentence is the single most reliable predictor of a review that is about to be written.
2. The transfer maze. A caller who gets handed off twice and re-explains the problem each time is losing patience whether or not they say so.
3. The unanswered question. The call ends politely and the caller never got what they called for. These are invisible in every metric you have — the call was answered, its duration was normal, nobody complained.
4. The billing surprise. A price that differs from what somebody remembers being quoted. Worth catching early because the recording settles it.
5. The long hold. Especially when the caller mentions it.

None of these show up in call volume, answer rate, or average handle time. A flagged call is often a perfectly normal-looking row in your CDR.

What to do with a flagged call

Open it, read the transcript, and decide one of three things.

Call the person back. This is the whole point. A customer who had a bad call on Tuesday and gets a phone call on Wednesday is in a completely different mood than the same customer on Friday. The window where a recovery is cheap is measured in days, and it closes before the review gets posted.

Fix the routing. Several flags naming the same friction usually describe a configuration problem rather than a staffing one. Repeated transfers means the auto attendant is sending calls to the wrong first stop. Long holds clustered at 8am means your morning needs a queue rather than a ring group. Understanding routes covers where a call actually goes and why.

Mark it reviewed and move on. Plenty of flags are a caller who was already having a bad day and a staff member who handled it correctly. Marking it reviewed keeps the queue honest, and a queue you trust is a queue you keep opening.

What red-flag detection is not

It is not a staff scorecard. Sentiment and red flags describe the caller's experience, not an employee's performance rating. We do not produce a per-agent score, and we would argue against building one. The moment staff believe a model is grading them, they optimize for the model instead of for the caller, and you lose both the metric and the behavior.

It is not exhaustive. A model reading a transcript will miss things a human would catch, particularly tone that never becomes words. Treat the queue as the calls most worth your attention, not as a complete inventory of every call that disappointed somebody.

It is not a substitute for the recording. The flag is an accusation until you inspect it. That is why the transcript sits on every flagged call — why we show every call's transcript makes the argument at length.

And it is not real-time. Nothing intervenes during a live call and nobody gets a mid-call alert. Signal works on a recording after the call ends.

What has to be on before any of this works

Three things, in order.

Call recording. Red-flag detection reads a transcript, and a transcript comes from a recording. Recording is inbound-only on Signal, because the feature exists to understand what customers experience.

A consent announcement. Signal requires one. You upload the audio and Signal plays it to the caller before recording begins. Federal law permits one-party consent, but many states require all parties to consent, so this is a legal requirement rather than a courtesy setting. The recordings library covers uploading it; call recording laws by state covers which states change the calculus.

The AI toggle, which is off by default on every account. Transcription and classification send call audio to OpenAI and transcript text to Anthropic, and Signal has no BAA with either provider. A practice handling protected health information should leave AI transcription and summaries off pending a BAA. The rest of the system — routing, queues, voicemail, call reporting — is unaffected by that decision.

Frequently asked questions

How can I find out which customer calls went badly? Record inbound calls with a consent announcement, then enable AI features so Signal transcribes and classifies each one. Calls where the caller had a bad experience are flagged with a written reason and land in a review queue, emailed as they occur and summarized in your weekly call report. You review a handful rather than sampling hundreds.

What is red-flag call detection? A classification Signal produces on each recorded inbound call, in the same model pass as the summary and sentiment. It marks calls where the caller appears to have had a bad experience and states why — repeated transfers, a third call about the same issue, an unanswered question. Flagged calls carry the full transcript for inspection.

Do I have to record calls to use it? Yes. Red-flag detection reads a transcript, and the transcript comes from a recording of the call. Recording on Signal is inbound-only and requires a consent announcement that plays before recording starts. Without recording enabled, there is nothing for the classifier to read.

Does red-flag detection score my employees? No. The classification describes the caller's experience on a specific call, not an employee's performance, and Signal produces no per-agent rating. Many flagged calls involve staff who handled a hard situation well. Use the queue to fix routing and recover customers, not to build a leaderboard.

If reviewing calls has quietly fallen off your list, the fix is usually the filter rather than the discipline. We are happy to talk through what turning recording on would mean for your office, including the parts you should probably leave off.