Which AI Features in a Therapy EHR Are Worth Paying For

Every practice management vendor now has an AI section on the pricing page, and every demo of it goes beautifully: the note writes itself, the schedule fills itself, the claim scrubs itself. The question facing a practice owner is no longer whether AI features exist or whether they can look impressive for four minutes — it is which of them, at add-on prices, actually return more time than they consume in your practice, on your caseload, with your payers. That question has a usable answer, and it does not require guessing which vendor’s model is smarter. It requires one test applied consistently: for each feature, what step does it remove from your week, and what review step does it add? Features that remove a step you already perform are usually worth evaluating. Features that replace a step with a review step are worth it only when the loop is genuinely faster than the task. Features that add a step you never performed are usually a demo, not a tool. This article builds that test into a feature-by-feature map, a set of trial tests a vendor demo cannot fake, and the transparency and contract questions to settle before the line item lands on your invoice.

Strip the branding off any AI feature and it does one of three things to your workflow. It removes a step you currently perform by hand — scanning a waitlist for a fit, checking a claim against payer edits before submission, retyping a dictated sentence into a field. It replaces a step with a different step — you no longer draft the note, but now you review a draft you did not write. Or it adds a step that never existed — a dashboard of “insights” someone now has to read, interpret, and decide whether to trust. Vendors present all three as time savings. Only the first is unconditionally a time saving. The second is a trade whose value depends entirely on whether the new step is faster than the old one for your caseload. The third is a cost wearing a feature’s name tag.

The reason the second family needs honest math is that for anything generative — drafted notes, drafted messages, summarized evaluations — the review step is not optional and not vestigial. APTA’s practice advisory on AI-enabled ambient scribes states that AI-generated documentation may contain errors of omission or addition and must always be reviewed by the clinician. ASHA’s guidance on generative AI frames these tools as support for clinical work rather than a substitute for clinical reasoning, and points clinicians to their ethical obligation to evaluate any technology they use. AOTA’s journal has pressed the same point for occupational therapy. Whatever the tool drafts, the clinician who signs owns — so the honest unit of evaluation is never “minutes the draft saves.” It is net minutes: the time the feature removes, minus the review time it genuinely requires, measured across a real week. A drafted note that saves twelve minutes of typing but needs six minutes of careful correction is a six-minute feature, and should be priced as one.

The table below applies the test to the AI features that appear most often on therapy EHR and practice management pricing pages. It deliberately says nothing about any vendor’s implementation quality — that is what your trial measures. What it gives you is the right question to ask about each feature before the trial starts, so you evaluate the feature’s honest shape rather than its demo.

Two patterns in that table are worth making explicit. First, the features with the cleanest case — dictation cleanup, claim scrubbing, waitlist fill — are mostly not generative glamour. They remove mechanical steps and add thin review, which is why they tend to survive the net-minutes test even in skeptical practices. Second, the features with the biggest headline promise — ambient scribes, chart summarization — are exactly the ones whose value swings hardest on your own behavior, because the review step is clinical, unskippable, and easy to underprice at demo time. Neither pattern says yes or no to any feature. It says: know which kind you are buying, and test accordingly.

Treat “AI-powered” as an advertising claim, because that is what it is. The Federal Trade Commission has told businesses in plain terms that claims about what an AI product can do need substantiation like any other performance claim, and that exaggerating or overpromising what the technology delivers is the kind of deception it pursues. You do not need to litigate any of that from a buyer’s chair — you just need the posture it implies: the burden of proof sits with the vendor’s product in your hands, not with the vendor’s demo in theirs. The way to collect that proof is the same bottleneck-first trial method that works for practice software generally, narrowed to the AI line items.

1. 01

   Pick the three tasks the feature claims to attack, from your real week

   Not the vendor’s scenarios — yours. If you are evaluating an ambient scribe, that is your most common session type, your most complex one, and the one with the messiest room audio. If it is claim scrubbing, it is last month’s three most annoying denials. A feature evaluated on tasks you rarely do will look better than it is.

2. 02

   Recreate each task in the trial with fictional or de-identified cases

   No real patient information enters any system before the business associate agreement covering your account is signed — trials included. Role-play the session, reconstruct the claim, rebuild the message thread. The realism you need is workflow realism, not real PHI.

3. 03

   Time the full loop, including review and correction

   Start the clock when the task starts, and stop it when the output is genuinely done: note reviewed, corrected, and signed; claim verified against the payer rule; message read and sent. Then time the same task done your current way. The difference — net minutes — is the number the price has to justify.

4. 04

   Read the errors, not just the average

   One invented data point in a drafted note matters more than ten formatting quirks, because it tells you the review step can never be skimmed. Note what kind of errors appear, whether they cluster somewhere predictable, and whether the tool lets you fix the cause in a template or setting rather than re-fixing the symptom every day.

5. 05

   Price the winner at your volume, in writing

   Multiply net minutes by how often the task actually occurs, then set that against the add-on’s price at your clinician count — including what the price becomes at renewal and at twice your size. A feature that clears the bar for a six-clinician practice can fail it for a solo one, and the reverse.

You are allowed to ask how the AI works, and there is now a federal template for what a straight answer looks like. ONC’s HTI-1 final rule created the first algorithm-transparency requirements for health IT certified under the federal certification program: systems with predictive decision support must let users see “source attributes” describing the intervention — up to 31 of them, covering things like the data used to develop it, how it was validated, its known limitations, and how its performance is monitored. Many therapy-specific EHRs are not ONC-certified and are not bound by any of it. That does not make the rule irrelevant to you — it makes it a free, federally drafted question list. A vendor whose product is certified can show you the attributes; a vendor whose product is not certified can still answer the questions, and how they handle being asked is itself due-diligence data.

Nothing in the net-minutes math buys back a compliance failure, so run the gates before the scorecard. An AI vendor that receives protected health information on your behalf is a business associate under HIPAA, and HHS is unambiguous that a written business associate agreement must be in place before that information flows — there is no trial exception and no “we’re HIPAA compliant” exception. A feature that records session audio adds recording-consent questions that vary by state, covered in our consent walkthrough. And whatever any tool drafts, the clinician who signs is its author: the review obligations, and what payer reviewers actually punish in AI-shaped charts, are laid out in the sign-off article and the audit-risk article. If a vendor fails a gate, the feature’s brilliance is not a tiebreaker, because there is nothing to break: the answer is no.

Put the pieces in sequence and the decision stops being about AI at all. First, know where your hours go, because a feature can only pay for itself against a bottleneck you actually have. Second, run the gates — BAA, training-data terms, consent — and let a failed gate end the conversation. Third, trial the survivors against last week’s real tasks and keep the net-minutes numbers in writing. Fourth, buy the boring winners first: the step-removers with thin review usually fund themselves fastest, and they teach your team to trust or distrust the vendor’s AI before you bet your documentation workflow on the glamorous one. And revisit annually at renewal, because these features change faster than contracts do — this year’s demo-ware is sometimes next year’s genuine tool, and the reverse.

“

“An AI feature is not smart or dumb on your invoice. It is minutes in or minutes out, and only your week can tell you which.”