Can I record and score patient phone calls for marketing?

I am a marketer, not a lawyer, and nothing here is legal advice. Recording law varies by state and the consent question is genuinely a legal one. Take it to counsel before you switch anything on.

Why a practice would want this at all

Because the phone is where the booking happens, and the phone is a black box.

A practice can see that a keyword produced forty calls. It can’t see that eleven were records requests, nine were existing patients, six were the wrong specialty, and only fourteen were people trying to become patients. Two accounts with the same call count look identical in the reporting and are worth completely different amounts.

Scoring closes that. Each call gets read against a rubric the practice wrote, and the ones that were genuine new-patient enquiries get counted as such. The bidding then learns which searches produce those and which produce noise. It is the same job as importing booked patients, one step earlier in the funnel and available much faster, because a call can be scored the next morning while a booking may take weeks.

Two permissions, not one

These get conflated and they’re unrelated.

The vendor agreement. Whoever stores the recording holds patient audio. That needs a business associate agreement, on the plan you’re actually on, as covered in the BAA question.

Recording consent. A separate body of state law, unrelated to health privacy. Some states require one party to consent, others require all parties. A multi-office practice can straddle two regimes, and a patient calling from out of state adds another. The workable answer is an announcement on every call in both directions, and that’s a practice decision made before recording is enabled.

A practice can clear one and not the other, and clearing the vendor agreement is the one that creates false confidence.

Score the transcript, not the audio

This is the design decision that makes the whole thing tractable, and I arrived at it the hard way.

Vendors implementing a covered configuration commonly restrict access to call audio while leaving the machine transcript available as ordinary text on the lead record. When I confirmed that in writing with one vendor’s support team, it changed the architecture: the audio path came out entirely and the pipeline was rebuilt to score text.

That turned out to be better on every axis. Text is cheaper to process, it can be scored by a model running locally, and the residual data at rest is smaller. Verify it with your own vendor in writing rather than assuming, because the exact split between what is withheld and what remains reachable differs between products and between plan tiers.

Where the model runs matters more than which model

If scoring goes to a hosted AI provider, that provider is receiving patient call content and becomes another vendor needing an agreement, on a plan that offers one.

Running the model locally removes the question rather than answering it. A small open-weights model on a machine the practice or its operator controls, with constrained decoding so the output is always a valid record, handles a rubric like this without difficulty. No vendor, no API key, no per-call cost, and one fewer agreement to negotiate.

Two things I wouldn’t skip. First, keep a human reading the output for a long time before anything automated acts on it. A small model’s stated confidence isn’t reliable; I have measured one scoring correctly on nine of ten rubric cases while reporting maximum confidence on the one it got wrong. You can’t build a threshold on a number like that, so the control is human review, not a cutoff.

Second, run in observe mode first, and make sure observe mode genuinely writes nothing. An observation run that marks records as processed will cause the live run to skip them forever, which is a silent, permanent undercount that looks like the scoring simply found nothing.

Store less than you think you need

The instinct is to keep the reasoning so a human can audit a score. Then you look at what the log contains and it’s the caller’s number beside a description of their complaint, sitting in a file on a machine, which is a small clinical record you didn’t mean to create.

Keep the decision. Keep a keyed fingerprint instead of the raw identifier, so duplicates can still be detected without the number being readable. Drop the free-text reasoning once a human has reviewed it, or do not write it at all. Set a retention period and schedule the deletion job at the same time you schedule the scoring job, because a purge script that exists but is never scheduled is how a system quietly accumulates months of transcripts.

That last one isn’t hypothetical. An unscheduled purge is the most common reason a pipeline that was designed to hold days of data is holding a quarter of it.

Expect it to be quiet

The volume surprises people. A single-specialty practice can run a meaningful ad budget and still produce a handful of scoreable paid calls a month, because most calls aren’t from ads and most ad clicks do not call.

Two implications. Value-based bidding needs far more volume than a practice this size generates, so the honest answer for most is that scoring informs human decisions rather than feeding an algorithm, at least for a long while.

And a quiet night has to be distinguishable from a broken one. If the pipeline is silent when it works and silent when it fails, nobody finds out it stopped. Make it report that it ran and found nothing, rather than saying nothing at all.

The order I would build it in

Get the vendor agreement and the covered configuration working first, and find out what that configuration breaks. Settle the recording consent question with counsel. Then score in observe mode with a human reading every output against the rubric, and only after the rubric is signed off by whoever owns intake does anything get written back to the ad account.

Most of the value arrives at the observation stage, when a practice finds out what its paid calls actually consist of. That alone usually changes the keyword list.