Buying an AI Scribe With Fewer Than Ten Providers

Because you are shopping in a different half of the market than the case studies describe, and most published advice was written for the other half.

Ambient documentation vendors have split into two shapes. Enterprise products sell to health systems through a security review, a legal negotiation and an implementation team, with pricing on request and contracts measured in years. Self-serve products sell to individual clinicians and small groups with a card, a published price and a monthly term. The technology in the two is often comparable. The buying process is not, and neither is the leverage.

With fewer than ten providers you have almost no negotiating leverage on an enterprise contract and you will pay for a procurement process you do not need. What you do have is the ability to decide quickly, switch cheaply and pilot without a committee. Buy in a way that uses those advantages instead of imitating a health system. The general market view is on the AI medical scribes comparison, and the workflow itself is described on the ambient documentation page.

Self-serve, in almost every case under ten providers. The exception is a practice that needs deep write-back into a specific EHR and cannot get it any other way.

The row that decides it for most small practices is time to first note. A self-serve product tells you within a week whether ambient documentation works for your specialty, your accents and your room acoustics. An enterprise product tells you the same thing three months and one legal review later, and by then you have spent more on the evaluation than a year of the alternative.

The row that occasionally overrides that is EHR integration. If copy and paste into your system is genuinely unworkable, integration depth becomes the deciding factor, and it is worth reading the EHR integration due diligence questions before you shortlist rather than after.

At the self-serve end, published per-clinician monthly pricing is common and generally sits in the low hundreds of dollars. Several vendors publish a list price on their website, some publish a lower rate for annual commitment, and some publish nothing at all.

At the enterprise end, published pricing is not available. Vendors selling to health systems price per clinician with volume tiers, and frequently add an implementation or integration fee. Any figure you read in a blog post for enterprise ambient documentation is either a leak, an estimate, or a number from a different contract shape, and it should not go into your model.

What we can say usefully is what the shape of the cost is rather than its level. Per clinician per month, billed on named seats rather than usage, is the dominant model. Watch for three things that change the effective price: minimum seat counts that force you to buy for clinicians who will not use it, per-encounter overage above an included volume, and integration fees quoted separately from the subscription.

Whether it pays for itself is a different calculation and one you can run before you talk to anyone. The AI scribe ROI calculator takes your visit volume, your documentation time and your loaded cost per clinician hour and gives you the break-even. For most practices the answer turns on whether reclaimed time gets converted into visits or into going home earlier, and both are legitimate answers as long as you pick one before you buy. A like-for-like view of published rates sits on the scribe pricing comparison.

Three. Not one, because you will not know what normal looks like. Not seven, because a small practice does not have the hours and the marginal vendor teaches you nothing.

Build the three from different shapes rather than from a ranking. One vendor with a native or app-store integration into your specific EHR. One general-purpose ambient product with published pricing and a free or cheap trial. One specialty-oriented product if your specialty has one, because behavioural health, dermatology and orthopaedics have documentation patterns that generic models handle unevenly.

Screen them on paper before any demo, using four questions you can answer from their public material: is per-clinician pricing published, do they state that they sign a business associate agreement, do they state a position on training on customer data, and do they name an integration with your EHR. A vendor that fails two of those four does not need an hour of your week. This is the same screen the pre-demo HIPAA checklist applies in more detail.

Then demo two and pilot one. Piloting two products simultaneously in a practice with six clinicians produces confusion rather than a comparison, because you cannot hold the clinician constant.

Four weeks, three clinicians, one number. Small practices tend to run pilots that are too short and too enthusiastic, and then generalise from the doctor who was always going to like it.

Pick three clinicians: two who volunteered and one who did not. The sceptic is the most valuable participant in the pilot, because rollout is decided by people like them and not by the early adopter. If your sceptic cannot be persuaded to try it for four weeks, that is worth knowing before you commit the practice.

Measure one thing consistently: minutes spent on documentation per clinic session, self-reported at the end of each session, recorded for two weeks before the pilot starts and for the four weeks of the pilot itself. Self-reported is fine at this scale. What is not fine is having no before figure, which is the single most common pilot failure and the reason so many practices cannot say whether the tool worked.

Then watch three qualitative signals: how often the clinician edits more than a couple of lines, whether note quality holds up in your specialty's harder encounters, and whether anyone stops using it in week three. Silent abandonment in week three is the classic pattern, and it usually means the editing burden exceeded the time saved. The metric design that scales beyond a scribe pilot is covered in the pilot metrics post.

A signed business associate agreement, and a clear written answer on whether your data trains shared models.

Encounter audio and the resulting note are protected health information, so any vendor processing them is a business associate under HIPAA and must sign a business associate agreement. A vendor that will not, at any price, is not a candidate. This is not negotiable and it does not scale down with practice size: the obligation is identical for a solo practice and a health system, which is the point small practices most often get wrong.

The second question is training rights. Some vendors use customer data to improve shared models by default, some do not, and some allow you to opt out. All three are legitimate commercial positions. What matters is that you know which one you signed, in writing, before the first encounter is recorded. The full question set is in the fifteen questions for AI vendor RFPs, and the shortlist view of vendor posture is on the HIPAA-compliant AI tools page.

One thing that is genuinely different at small scale: you will not amend the standard terms. Self-serve vendors do not negotiate, and pretending otherwise wastes weeks. Your decision is to read the terms and accept or decline them, which makes reading them the entire exercise.

Less than vendors imply and more than practices plan for. There is no implementation project at this size, but there are three things that decide whether adoption sticks.

The first is the note template. Out of the box the draft will not match how your clinicians write, and the fix is usually a template or a few instructions rather than a support ticket. Budget an hour per clinician in week one for this and it stops being the reason someone quits in week three.

The second is patient consent language. You need a sentence at the start of the encounter, the same sentence every time, and it should be written down rather than improvised. Where you practise, that sentence may also be a legal requirement rather than a courtesy: Texas requires clear and conspicuous plain language disclosure that a patient is interacting with an AI system, and other states have their own duties. The current map is in the state AI law post.

The third is the review discipline. Somebody has to be able to say, if asked, that no note reaches the chart unread. That is a clinical safety position first and a regulatory one second, and it is also what keeps you outside the harder end of the patient-communication rules. Write it down in one paragraph and put a date on it.

Five, and four of them are about getting out rather than getting in.

1. Silent training rights. Terms that permit use of your data to improve the service, with no carve-out for shared models and no opt-out. The phrase to look for is broad and permissive, and the absence of a data section is worse than an unfavourable one.
2. Auto-renewal with a long notice window. An annual term that renews unless cancelled 90 days in advance is a two year commitment wearing a one year label.
3. Exit without export. If the contract does not say you get your notes and transcripts back in a usable format, and how quickly, assume you do not. Notes in the EHR are yours. Anything held only in the vendor's system may not be.
4. Seat minimums above your headcount. A ten seat minimum sold to a six provider practice is a 40 percent price increase described as a discount tier.
5. Price protection that runs one way. Terms allowing the vendor to change pricing on notice, in a contract you cannot exit on the same notice, is the combination to refuse.

The pattern behind all five is that small practices sign for the product and get hurt by the term. Read the termination clause first and the feature list second. If you want a second opinion before signing, our vendor selection engagement is built around exactly this decision, and it is deliberately capable of concluding that the incumbent workflow is fine and you should not buy anything yet. Practice-specific context sits on the private practice page.