Ambient AI Scribe Contract and a Portal Backlog

Cost: $299 × 8 clinicians × 12 months = $28,704 per year. Break-even: at $95 net collections per visit, that is 28,704 ÷ 95 ≈ 302 additional visits per year, or about 0.15 visits per clinician per day - roughly seven minutes of recovered time if the saved hour is actually converted into clinical capacity.

That break-even is trivially low, which means the ROI is not the real question. The real questions are:

1. Does saved time convert? If clinicians use the recovered 90 minutes to leave earlier, the practice captures retention and burnout value but no revenue. That can still justify $28,704 - replacing one physician costs far more - but it should be stated as the rationale rather than implied.
2. Does the practice have the demand to fill added capacity? If third-next-available is already long, yes. If not, the ROI is zero regardless of hours saved.
3. What does the manual paste cost? With no EHR integration, each note requires review, correction, and copy-paste. If that consumes 15 minutes, you have bought a 75-minute saving, not 90, and you have introduced a new documentation error pathway.

Recommendation: do a single-clinician, 60-day pilot with baseline and post time measurement, a short termination-for-convenience right, and no auto-renewal. Sign nothing 12 months long on a vendor's time-savings claim.

Privacy and data rights - the core problem.

• 'De-identified' is a defined standard: either the HIPAA safe harbor (strip the 18 identifiers, including dates and any unique codes) or expert determination with documented methodology. The contract says 'de-identified' without specifying which, and the practice cannot verify compliance after the fact. Require the methodology in writing and a right to audit.
• 'Aggregated data for any lawful purpose' is effectively an unlimited license. Aggregated data can be re-identified when the cell sizes are small - a solo specialty practice's records are not statistically anonymous. Strike it or narrow it to de-identified, aggregate-only, no re-identification, no use to train competing products sold to this practice's competitors.
• The BAA must be a real BAA: permitted uses, minimum necessary, subcontractor flow-down, breach notification timelines shorter than the 60-day regulatory outer limit, return or destruction on termination, and no sale of PHI.

Contract. Auto-renewal on a 12-month term for an unproven product is unacceptable; demand month-to-month after a 60-day pilot, a termination-for-convenience clause, a service-level commitment, and clarity on who owns the generated note (the practice and patient, not the vendor).

Documentation. The clinician must review and edit every note and attest to it before signing. Billing levels must reflect the clinician's own medical decision-making - a fluent AI narrative that describes history or complexity the clinician did not actually obtain or consider is upcoding, whatever the software produced. Configure the EHR audit trail to preserve who drafted and who edited.

It is a workflow and staffing problem wearing a technology costume, and buying a scribe will not touch it. But it does carry compliance weight: HIPAA gives patients a right of access to their records within 30 days, and while routine clinical questions in a portal are not themselves a records request, a triage system that lets 22% of messages age past three business days will eventually swallow a genuine access request - and it is already producing front-desk complaints, which means patients are routing around the channel the practice spent money to build.

Diagnose before prescribing:

• Volume per clinician per day, and what fraction are clinical questions versus scheduling, refills, and billing.
• Which messages are going unanswered - is it one clinician at 40% and seven at 10%? A single outlier needs a conversation, not software.
• Time-of-day and day-of-week patterns, to see whether the backlog is a coverage hole (rounds, OR days, vacations).

Interventions in order of cost and speed: a written message policy with a published response standard and auto-acknowledgment; routing rules that send refills and scheduling to staff rather than the clinician; a shared triage pool staffed by an RN or MA with escalation criteria; and non-clinician coverage during known absences. Charge for portal messages only after checking payer rules and the patient-relations cost - many practices regret it.

Set a target (e.g., 95% of clinical messages answered within one business day) and report it monthly. If the volume is genuinely beyond capacity, the honest answer is another clinical FTE, and the AI scribe money is better spent there.