1. Five things "AI-powered" can mean

They differ enormously in how much work they remove, and the cheapest to build is the one most often put on the homepage.

Claim What it actually is Hours it gives back
A chat box A general assistant bolted on, usually without access to your data Close to none
Documentation A scribe drafting notes from the visit Real, and the easiest to measure
Data handling Reading an old system’s export, rebuilding forms, placing signature fields Large, but once, during onboarding
Reading results Structuring labs and flagging what is out of range Steady, and grows with lab volume
Suggestion Proposing a plan, a follow-up or an order for a human to approve Potentially the largest, and the one needing the most scrutiny

A product can honestly claim AI while doing only the first. That is the gap this post is about.

2. The evidence that settles each claim

Every one of these is answerable in a live session with your own data. None requires trusting a slide.

When a vendor says Ask them to show A weak answer sounds like
“Our AI knows your patients” A question about a real chart, answered with the source shown “It is trained on clinical data”
“AI drafts your notes” A draft from your own recording, numbers intact A pre-recorded sample visit
“AI handles your migration” Your actual export file read on the call “Our team handles it”
“AI reads your labs” A result arriving as discrete values that plot, not a PDF A PDF in an inbox with a summary
“AI suggests next steps” A suggestion with a citation to the chart line it came from A suggestion with no source

The last row is the one that matters most as these systems get more capable. A suggestion you cannot trace is one you have to re-derive yourself, which costs more time than it saves.

3. Why a chat box is not an AI EMR

The value is not the model, it is what the model is wired into. Two products can run the identical model and behave completely differently, because the question is what it can read and what it is allowed to do with the answer.

Consider one ordinary request: “book this patient’s follow-up”. A chat box can tell you that a follow-up is advisable. A system wired into scheduling can find a real slot with the right visit type and the right prescriber, present it, and book it when a human taps once. The sentence a vendor shows you is the same in both demos. The work removed is not remotely the same.

So when you are shown an AI feature, ask what happens after the answer appears. If the next step is a human retyping it into another screen, the AI is decorative and the seams are still where the time goes.

4. Four refusals to get in writing

What a clinical AI is forbidden from doing matters more than its feature list, and it belongs in the contract rather than the brochure.

The refusal Why it has to be explicit
It never files anything a clinician has not approved A note, order or message that posts itself is an unsigned clinical act
It never changes a dose or a diagnosis Suggestion and decision are different jobs, and only one is yours to delegate
It never states a clinical fact without a source An unciteable claim has to be re-derived, which is slower than not having it
It never sees another clinic’s data, and yours never trains a shared model Your BAA covers your patients; it does not make their records training material

A vendor who agrees to all four in writing has told you more about its engineering than any demo will. One who treats the fourth as unusual has told you something too.

5. The data questions, in order

Ask these before the feature questions, because a no here ends the conversation regardless of how good the product is.

Question What a good answer looks like
Will you sign a BAA? Yes, as standard, offered without being chased
Does our data train your models? A written no, or a precise description of what you have agreed to
How long is our data retained by the AI layer? A number, and ideally one you can shorten
Where is it processed? A straight answer, and in writing if residency matters to you
What happens on the day we leave? The same export you would get without AI, in a format you can read

There is no HIPAA certification for software and nobody holds one. What a vendor can do is sign a business associate agreement and build to the requirements — which is what we do, and what you should expect as a floor rather than a feature.

6. How AI features get priced

AI has a per-use cost to the vendor, so it is rarely bundled forever, and the structure tells you what to expect.

Structure What it means for you
Included in the plan Simplest; check whether an allowance sits underneath it
A flat add-on Predictable, and usually the easiest to budget
Per provider Scales with hiring, which is the plan
Per use or per hour Scales with volume, which is also the plan

Ours is published rather than quoted: one AI scribe seat is included with every Prescribing plan and extra seats are $49 a month, with an allowance of 100 scribed hours per seat per month and $0.50 an hour beyond it. Whatever a vendor’s shape is, the three numbers to extract are the allowance, the overage rate and the unit it is counted in. A headline price without those three is not a price.

For how the underlying platform gets priced, independent of AI, see clinic software pricing models.

7. Frequently asked questions

What makes an EMR actually AI-powered?
That the AI is wired into the rest of the system rather than sitting beside it. A chat box that cannot read your charts or act on your schedule is a different product from one that drafts a note, reads a lab into discrete values, or proposes a real appointment slot for a human to approve.

Is AI in an EMR HIPAA compliant?
Compliance is a property of how the vendor handles your data, not of the model. Ask whether they sign a BAA, whether your data trains their models, how long the AI layer retains it, and where it is processed. No software holds a HIPAA certification, because none exists.

Should AI be allowed to write in a patient chart?
It can draft. It should not file. Any note, order or message that reaches the record without a clinician approving it is an unsigned clinical act, and that boundary belongs in your contract.

How much does an AI EMR cost?
The platform and the AI are usually priced separately. Aminova publishes $199 a month including one prescriber and $99 for each additional one, with one AI scribe seat included and extra seats at $49 a month.

How do I test an EMR vendor’s AI claims?
Make them use your data on the call: your export file, your recording, your lab result. Every honest AI claim survives that; the decorative ones do not.