Build vs buy an AI medical scribe: an honest comparison

Short answer: buy a product if your burden is a clinical note. Build if your burden is a pile of structured forms specific to your setting, if you need EMR integration no vendor supports, if per-encounter pricing has become a dominant line item, or if data residency rules out sending audio to a third party.

We build custom scribes, and we still tell most people to buy one. Here is the honest comparison.

Three options, not two: product, API, or built

The framing is usually “buy a finished product or build from nothing”, and it skips the middle option most teams actually end up in. There are three:

Buy a product. Abridge, Nuance DAX, Suki, Freed. A clinician-facing application, priced per seat, with the EMR integrations the vendor chose to build.

Build on an ambient scribe API. Some vendors expose the transcription and note-generation as an API you call from your own product. You own the interface, the workflow and the EMR integration; they own the model. It is faster than building in-house and leaves you with a per-call cost that scales with usage and a dependency you cannot tune.

Build in-house. Speech-to-text from a provider like Deepgram, your own diarisation, extraction and form mapping, your own confidence scoring and review interface. Highest effort, no per-seat ceiling, and the only option that lets you fill forms specific to your organisation.

Most of what follows compares the first and third, because they bracket the decision. If you are weighing an ambient scribe API against building in-house, read the build column: the API removes the speech and note-generation work, roughly the front third, and leaves you owning the extraction, form mapping, review workflow and integration, which is where the schedule actually goes.

Product versus custom, side by side

	Buy a product	Build custom
Examples	Abridge, Nuance DAX, Suki, Freed	A system built around your workflow
Time to first use	Days	Weeks
Cost model	Per clinician or per encounter, ongoing	One-time build, then hosting
Output	Clinical note, usually SOAP	Whatever your chart requires
Custom forms	Rarely, and rarely all of them	The specific forms you must file
EMR integration	Whatever the vendor supports	Whatever your EMR exposes
Data residency	Vendor’s infrastructure	Yours, if required
Specialty vocabulary	Broad coverage, limited tuning	Tuned to your specialty
Who fixes a bad note	Support ticket	You, same day
Ongoing cost at scale	Grows with headcount	Roughly flat

Where products genuinely win

Be clear about this before spending money on engineering.

Speed. A product works this week. A build works next month. If the documentation burden is actively driving clinician attrition, that difference matters more than feature fit.

Breadth of speech coverage. Vendors have trained on volumes of clinical audio no single practice can assemble. On general consultation transcription they are very good.

Someone else’s problem. Model updates, uptime, drift and compliance paperwork all sit with the vendor. That has real value.

Price at small scale. For a handful of clinicians, per-seat pricing is cheaper than any build. The crossover only arrives with volume.

If your requirement is “turn this consultation into a good note,” buy the product. You will not beat them on price and you probably will not beat them on quality.

Products win more often than a firm offering custom scribe development is supposed to say, and the cases below are where they stop winning.

Where building wins

Your forms are not a SOAP note. This is the most common reason our clients build. Products produce a note; many settings must file ten or more mandatory structured forms. A note does not populate those, so the clinician re-enters everything and the promised saving evaporates. On a production build for a healthcare client, one consultation populates 10+ mandatory forms with the clinician entering data zero times.

Per-encounter pricing has become the line item. Per-seat or per-encounter costs scale linearly forever. A build is largely fixed. There is a volume at which the arithmetic reverses — work out where yours sits before assuming either.

Your EMR is not on the supported list. Legacy or in-house systems are common in healthcare and rarely a vendor priority. A custom build integrates with whatever your EMR actually exposes.

Audio cannot leave your infrastructure. Some jurisdictions and institutional policies forbid sending patient audio to a third party. That single constraint eliminates most products.

You need control over the uncertainty threshold. Products decide what to flag. If you need a specific field always reviewed, that is a build decision.

The cost question

Nobody can quote a scribe build without seeing the forms — and anyone who does is guessing.

What drives cost is how many forms need populating and how unusual they are. Transcription is close to a commodity; the expensive engineering is fact-to-field mapping and the review interface. A scoped build against a defined set of forms is a matter of weeks, not months.

What drives product cost is clinician count and encounter volume. Model both over three years. The crossover is usually further out than vendors imply and closer than engineers assume.

What both options must have

Whichever way you go, these are non-negotiable:

Per-field confidence scoring. A system at 98% accuracy that knows which 2% it is unsure about is safe. A system at 99% with no confidence signal is not, because nothing separates the wrong answers from the right ones.
Nothing written unreviewed. The scribe drafts; the clinician signs.
Targeted review. If clinicians must re-read every generated note in full, the saving never materialises.
Diarisation that holds up on overlapping speech. “I have been taking ibuprofen” means something completely different depending on who said it.
HIPAA discipline across every hop — BAAs including the model provider, encryption in transit and at rest, role-based access, consent capture, audit logs.

When evaluating a vendor, the question is not “how accurate is it” but “what happens on the cases it gets wrong”.

A decision shortcut

Build if two or more of these are true:

You must file structured forms, not just a note
Per-encounter cost is already material, or will be within a year
Your EMR is unsupported by the vendors you have evaluated
Audio cannot leave your infrastructure
You need specialty-specific extraction a general product does not prioritise

Fewer than two? Buy the product. That is genuinely the advice we give, and we give it often.

Tools referenced

EpochC does AI medical scribe development. See the scribe case study — 10+ forms auto-filled per consultation — the clinical documentation guide, or book a call and we will tell you honestly which side of this you fall on.