How do we keep personal information out of an AI pipeline that analyzes calls?

Minimization is an engineering decision, not a policy statement

The default behavior of most integrations is to copy whole objects because that is easier than choosing fields. That single default is responsible for most of the personal data sitting in places nobody intended.

The alternative is a field-level allowlist: the pipeline pulls a named set of fields and drops everything else at the boundary. It takes an extra hour to specify and it means a card number in a free-text note never enters your warehouse at all. Do this at ingest, because anything you exclude later has already been stored.

Redaction helps and it is not complete

Automated redaction handles the structured cases well — a spoken card number, a sequence of digits that looks like an account, an email address read aloud. It is much weaker on the unstructured cases, because people say identifying things in ordinary sentences that no pattern will catch.

So treat redaction as risk reduction rather than a guarantee. Where the analysis does not require the words themselves, store the derived output and not the transcript. A rubric score and a short reason often carry all the operational value with a fraction of the exposure, which is how call analysis can be useful without spreading personal data across systems.

Reference beats copy

Recordings usually already live in a telephony platform with its own access controls and retention. Copying them into a second system doubles your exposure and doubles the number of places a retention policy has to be enforced correctly.

Store a pointer instead — the recording identifier and the system that holds it — and fetch on demand for authorized users. Fewer copies is the single most effective privacy control available in an integration, and it costs nothing.

Retention applies to what you derive, too

People set retention on recordings and forget the artifacts: transcripts, summaries, extracted entities, embeddings, evaluation records. Those are derived from personal data and they persist long after the source is deleted. Define retention for each artifact class, and make deletion of the source cascade to them.

The legal layer varies more than most people expect. Consent requirements for recording calls differ by state, and rules can differ again depending on where the caller is. Disclosure practices, retention obligations and what counts as sensitive data are all jurisdiction-dependent. Design the system so the technical controls are strong and configurable, then have counsel confirm what your specific situation requires. We build the data layer to support whatever policy you land on rather than assuming one.