Call script compliance AI: checking every mystery call and every recorded call against the script

A mystery shopping provider running a telephone programme for a bank places a few hundred calls a month to branches and call centres. Each call is scored against a script: was the greeting delivered, did the agent identify themselves, was the mandatory disclosure read, was the savings product offered, how was the call closed. A trained evaluator listens to each recording and fills in the scorecard.

A contact centre’s own QA team does the same thing with its own calls, listening to a handful per agent per month out of thousands.

Both are sampling. Both have an evaluator listening to a whole call to answer a dozen questions. And both produce scores that the people being scored dispute, because the sample was small and the evaluator was tired.

What the AI does with a recording

The recording is transcribed and the transcript is read against the script. Concretely, the script becomes a set of questions, each answered with a yes or no and a confidence score, and each pointing to the place in the transcript where the answer was found or should have been.

Was a greeting delivered in the first fifteen seconds? Did the agent state their name? Was the regulatory disclosure read in full, or paraphrased, or skipped? Was the product offered, and was it offered before or after the customer’s objection? Did the agent confirm the next step? Was the call closed with the required phrase?

Each answer comes back with the excerpt. “Disclosure not found. Closest match at 02:14: ‘so that’s all covered by the standard terms’.” The evaluator does not listen to the call. They read the excerpt and confirm or overturn.

Fields also come out. Call duration, hold time, the products mentioned, the reason given for a decline, whether the customer asked for a callback. These become data rather than notes.

Why coverage is the change, not the model

The value is not that a model can score a call. It is that every call gets scored.

For a mystery shopping provider that means every programme call is evaluated against the same standard with the same consistency, and the evaluator’s time goes to the calls where the model was uncertain. The scorecard delivered to the client is backed by the full recording set, and disputes are settled with an excerpt.

For a contact centre it means quality moves from a monthly sample to the whole month. An agent who skips the disclosure once in fifty calls is found. A team that stopped offering the retention product after a script change is found in the first week. Revenue leakage from offers never made becomes a number rather than a suspicion.

Text channels get the same treatment

Chat and email replies rarely get the scrutiny calls do, mostly because reading them is as slow as listening. The same pipeline reads a chat transcript or an email thread against the service standard. Was the request acknowledged? Was the required information given? Was the next step stated? This is also where the standard can be applied to a bot’s replies as readily as a human’s, which matters as more of the first contact moves to automation.

Where to be careful

Three things, from experience.

Paraphrase versus omission. A disclosure read in slightly different words is compliant in some regulatory contexts and not in others. The question has to be written to match the actual requirement, and the confidence threshold set so that paraphrases go to a person.

Audio quality. Recordings with heavy crosstalk or poor line quality transcribe badly. A quality check on the audio before transcription, and a “could not assess” outcome that is distinct from “not compliant”, keep bad audio from becoming false failures.

Agent trust. Agents accept scoring they can see. Every verdict should come with the excerpt and the timestamp. A scorecard that says “disclosure missing” without showing where is a scorecard that gets argued with.

What a rollout looks like

Providers and QA teams that get value quickly start with one script and one programme, run the automated scoring alongside the human sample for a few weeks, and compare. The comparison does two things. It calibrates the questions and thresholds. And it shows what the sample was missing, which is the number that justifies going wider.

Then recordings sync nightly through the API, every call is scored, and the evaluators work an exception queue with excerpts instead of a sample of full recordings.

If you run telephone mystery shopping or a QA function and want to see a month of calls scored against your script, book a demo. Bring the script and a few recordings.