The Benchmarks Everyone Quotes and Nobody Can Source
Start here, because it changes how you read everything else. The statistics that anchor almost every AI voice-agent pitch to a medical practice cannot be sourced. Not "sourced to an old study" — sourced at all. There is no survey instrument, no sample, no methodology, no publication behind them.
We went looking for the origin of each one before writing this article. Every trail ends at a voice-AI vendor blog citing another voice-AI vendor blog.
| Circulating claim | Where the trail actually ends | Verdict |
|---|---|---|
| “7% of patient calls are abandoned” | Voice-AI vendor blogs only. No survey instrument, no sample, no methodology | Do not use |
| “4.4-minute average hold time” | Repeated across vendor sites; never traced to a study or dataset | Do not use |
| “60% of callers hang up after one minute” | Vendor marketing. No published measurement | Do not use |
| “85% of patients never call back after one bad experience” | Vendor marketing. No published measurement | Do not use |
| “54% of patients book by phone, 46% online” | Vendor marketing. Conflicts with the access surveys that do publish methodology | Do not use |
| “$200 per no-show” / “$150 billion a year” | A single 2017 byline by the chief medical officer of a scheduling vendor. No study, no methodology behind it | Do not use |
The unsourceable phone benchmarks in the AI voice-agent category, traced as of August 2026.
There are real access statistics — they are just less dramatic. The AMN Healthcare 2025 physician appointment wait-time survey, which surveyed 1,391 offices across 15 metropolitan markets, found an average new-patient wait of 31 days, up 19% versus 2022 and 48% versus 2004. That is a documented access problem with a documented methodology. Build the business case on that, not on a number someone made up about hold times.
What the Peer-Reviewed Evidence Actually Shows
As of August 2026, the peer-reviewed literature on AI voice agents answering the phone at a medical practice consists of two single-site studies, neither with a control arm. That is the whole evidence base. There is no randomized controlled trial, and nothing has appeared in JAMA, the New England Journal of Medicine, or JAMIA.
| Study | Design | Scale | Reported result | Limitation |
|---|---|---|---|---|
| Cureus 2026;18(7):e112227 — high-volume outpatient neurology practice | Single-site, retrospective, no control arm | One practice | Scheduling backlog reduced by more than 98% | No comparison group; operational metrics collected by the site itself |
| Mount Sinai “Sofiya,” npj Digital Medicine, July 4, 2026 | Two-phase single-site pilot, no control arm | 1,431 patients | Pre-procedure call completion moved from 86.4% to 87.9% | A 1.5-point change, no randomization, one department, one health system |
| Randomized controlled trials of AI front-desk voice agents | — | None published | — | As of August 2026 there is no RCT, and nothing in JAMA, NEJM or JAMIA |
| “Careless Whisper,” ACM FAccT 2024 — speech-to-text hallucination audit | Systematic transcription audit | Large transcript sample | ~1% of transcriptions wholly hallucinated; 38% of hallucinations contained explicit harms | Measures the transcription layer, not a whole agent — but it is the layer every agent sits on |
The complete published evidence base for AI front-desk voice agents, August 2026.
The Cureus neurology study reports a scheduling backlog cut of more than 98% at a single high-volume outpatient practice. That is a large operational number, and it is also exactly the kind of number a single site with no comparison group produces when it changes several things at once.
Mount Sinai's "Sofiya," published in npj Digital Medicine on July 4, 2026, is the more careful of the two. Across 1,431 patients receiving pre-procedure calls, completion moved from 86.4% to 87.9%. That is a real, measured, modest improvement — and it is worth noticing how far it sits from the numbers on a vendor's homepage.
A 1.5-point improvement in call completion at one health system is a legitimate finding. A 97% resolution rate with no denominator is a marketing asset. They are not the same category of claim and should never appear on the same slide.
— Frenchy Digital evaluation principle
The one piece of rigorous evidence in this space is about risk, not benefit. The "Careless Whisper" audit presented at ACM FAccT 2024 found that roughly 1% of speech-to-text transcriptions were wholly hallucinated — invented content in an otherwise plausible transcript — and that 38% of those hallucinations contained explicit harms. Speech recognition is the foundation every voice agent stands on. Plan around that finding rather than hoping your vendor's model is the exception.
Vendor Performance Claims, Labeled Honestly
Vendors in this category publish impressive numbers. None of them are audited, and none publish a methodology you can replicate. That does not make them false. It makes them unverifiable, which is a different problem and one you should price into the contract rather than into the forecast.
| Source | Claim | Evidence status |
|---|---|---|
| Assort Health | 97% of calls resolved; 79% of referrals scheduled without staff involvement | Self-reported. Not audited. No published denominator or adjudication method |
| Notable | 57% containment at Catholic Health | Self-reported customer figure. Not audited |
| Hyro | 85% of routine interactions resolved | Self-reported. Not audited |
| Any “sub-75ms latency” claim | Marketed as response speed | Measures text-to-speech time-to-first-audio, not voice-to-voice round trip |
| Any “benchmark-leading” claim | Implies a shared, comparable test | No neutral healthcare voice-agent benchmark exists to lead |
Publicly circulated vendor performance claims for healthcare voice agents, with their evidentiary status, August 2026.
The three questions that separate a real metric from a marketing metric
What is the denominator? A 97% resolution rate is meaningless until you know which calls were counted. Were after-hours calls included? Wrong numbers? Calls that transferred within five seconds? Vendors frequently exclude the hard population and report on the easy one.
Who decided it was resolved? If the agent classifies its own outcome, the metric measures the classifier, not the outcome. Ask whether a human adjudicated a sample, how large that sample was, and what the disagreement rate was.
Over what period, at how many sites? A containment rate from one enthusiastic launch site in month one is not a forecast for your six-location practice in month nine.
The 2026 Market and the Pricing Problem
Capital has arrived in this category well ahead of the evidence. Assort Health raised a $120M Series C on June 24, 2026, led by Menlo Ventures at a $1.2 billion valuation, bringing it to $222 million raised. Hyro raised $45 million in October 2025. Hello Patient raised a $22.5 million Series A in September 2025, and Prosper AI a $30 million Series A led by a16z in June 2026. Parakeet raised a $3 million seed.
| Company | 2026 funding signal | Published pricing |
|---|---|---|
| Assort Health | $120M Series C on June 24, 2026 at a $1.2B valuation (Menlo Ventures); $222M raised to date | None published |
| Hyro | $45M, October 2025 | None published |
| Hello Patient | $22.5M Series A, September 2025 | None published |
| Prosper AI | $30M Series A, June 2026 (a16z) | None published |
| Parakeet | $3M seed | None published |
| Zocdoc “Zo” | — | $2 per autonomously booked appointment, no upfront fee |
Healthcare voice-agent funding and pricing transparency, 2025–2026.
The practical defense is to force the conversation into per-unit terms before you sign. Ask every vendor for cost per handled call and cost per completed booking at your actual call volume, with the containment assumption stated explicitly. Then model the same volume at half the claimed containment rate and see whether the deal still works. If it only works at the vendor's number, you are buying their forecast, not a system.
What to Automate First — and What Never
The single best predictor of whether a front-desk voice agent succeeds is scope discipline. Practices that automate three bounded, verifiable workflows tend to keep the system. Practices that let the agent "handle patient calls" tend to remove it within two quarters.
Here is how we sequence it. The test for the first column is simple: is the task bounded, is the outcome verifiable in a system of record, and is the harm ceiling low if the agent gets it wrong?
| Front-desk workflow | Automate now? | Human required | Why |
|---|---|---|---|
| Booking, rescheduling, cancellation | Yes — start here | Escalate on any clinical question | Bounded task, verifiable outcome in the scheduling system, low harm ceiling |
| Hours, directions, parking, accepted insurance, forms | Yes — start here | No | Static knowledge. A wrong answer is correctable, not dangerous |
| Appointment reminders and confirmations | Yes, with TCPA caps enforced in code | No | Covered by the healthcare exemption only inside strict frequency and length limits |
| Waitlist backfill after a cancellation | Yes | No | Outbound to patients already on the list, with a measurable acceptance rate |
| New-patient demographic and insurance intake | Yes, as a draft | Staff verifies before it lands in the chart | Speech recognition errors on names, member IDs and dates are common |
| Prescription refill requests | Intake only | Clinician approves every refill | Taking the request is administrative. Granting it is not |
| Eligibility and benefits questions | Partial | Staff confirms before any financial commitment | Payer data goes stale. A confident wrong answer becomes a billing dispute |
| Symptom questions, triage, “should I come in?” | No | Licensed clinician | Clinical. It also pulls the agent inside AB 3030 and Texas TRAIGA |
| Test-result delivery | No | Clinician | Clinical information. Disclosure duties and harm ceiling both jump |
| Past-due balance and collections calls | No | Staff | Expressly outside the TCPA healthcare exemption |
| Recall campaigns, service promotion, marketing | No | Staff, with prior express written consent | Telemarketing. The healthcare exemption does not reach it |
| After-hours emergencies | No | Answering service or on-call clinician | The agent must recognize and hand off, never assess |
Frenchy Digital sequencing guide for AI front-desk voice agents in outpatient practices, 2026.
Note the pattern in the bottom half of that table. Every workflow we hold back is held back for one of two reasons: it is clinical, or it falls outside the TCPA healthcare exemption. Those are the two boundaries that actually matter, and both are covered below.
An AI front-desk agent handles administrative work under human review. It does not assess symptoms, it does not decide urgency, and it does not practice medicine. If a workflow requires clinical judgment, the agent's only correct behavior is to route it to a person.
— Frenchy Digital scoping principle
TCPA: The Outbound Rules That Sink Practices
Inbound calls are the easy half. The legal exposure lives in outbound — reminders, confirmations, waitlist offers, recalls, balance notices — and it is the part vendors demo least.
The threshold fact: on February 2, 2024 the FCC adopted Declaratory Ruling FCC 24-17 in CG Docket 23-362, confirming that AI-generated voices are "artificial voices" under the Telephone Consumer Protection Act. It remains in force. Your agent is inside the TCPA regardless of how natural it sounds.
| Rule | What it says | Where it bites |
|---|---|---|
| Healthcare exemption — 47 CFR §64.1200(a)(9)(iv) | Prerecorded and artificial-voice healthcare calls to a wireless number are exempt from prior express consent, within limits | The limits are the entire rule |
| Frequency cap | One message per day, maximum three per week, per patient | Reminder plus confirmation plus survey plus refill nudge in one week is four |
| Length cap | Voice calls one minute or less; texts 160 characters or fewer | A conversational AI greeting eats the budget before the message starts |
| Opt-out | Immediate opt-out mechanism required in every message | Must be honored across every channel, not only the one it arrived on |
| Excluded content | Telemarketing, advertising, billing, accounting and debt collection are NOT exempt | This is where practices actually get sued |
| AI voices | Declaratory Ruling FCC 24-17, adopted February 2, 2024 — AI-generated voices are “artificial voices” under the TCPA | Your agent is regulated whether or not it sounds synthetic |
| Damages | $500 per violation; $1,500 per willful or knowing violation | Per call and per text, and commonly pleaded as a class action |
| AI-disclosure NPRM — FCC 24-84 (August 2024) | Proposed AI-call disclosure requirements. Still unfinalized as of August 2026 | Design for it. Do not describe it as law |
| One-to-one consent rule | Vacated in Insurance Marketing Coalition v. FCC, 11th Cir. No. 24-10277, January 24, 2025 | The consent regime vendors sold against in 2024 is not in force |
TCPA constraints on AI voice agents calling patients, as of August 2026. See 47 CFR §64.1200.
Two more things worth knowing because vendors get them wrong in both directions. First, the one-to-one consent rule that generated so much compliance advice in 2024 was vacated by the Eleventh Circuit in Insurance Marketing Coalition v. FCC, No. 24-10277, decided January 24, 2025. It is not in force. Second, the FCC's AI-disclosure proposal, NPRM FCC 24-84 from August 2024, is still unfinalized as of August 2026. Build so you can turn a disclosure on, but do not let anyone sell you compliance with a rule that does not exist yet.
What TCPA compliance looks like in code, not in a policy document
Frequency caps enforced at the send layer, not in a runbook: a per-patient counter that hard-blocks a fourth contact in a rolling week and a second in a day, across voice and SMS together. Length caps enforced as a validation step that refuses to dispatch a script over one minute of synthesized audio or a text over 160 characters.
Opt-out honored globally within seconds and propagated to every channel and every campaign, with the timestamp retained. A content classifier or a hard campaign-type flag that prevents billing, collections and marketing content from ever entering an exemption-based send path.
And a retained, exportable log of every outbound attempt with its consent basis, campaign type, duration and disposition. If a TCPA claim ever arrives, that log is the defense. Practices that cannot produce it settle.
Disclosure Law: Utah, California, Texas
"Am I talking to a robot?" is now a legal question in several states, and the answers differ enough that a single national script will be wrong somewhere.
| Jurisdiction | What triggers it | What you must do | Exposure |
|---|---|---|---|
| Utah — AI Policy Act as amended by SB 226 (effective May 7, 2025) | A regulated occupation (health care is one) plus a high-risk generative-AI interaction | Disclose at the START of the interaction — spoken aloud if the interaction is verbal | Up to $2,500 per violation (Division of Consumer Protection); up to $5,000 per violation (AG or courts) |
| California — AB 3030 (effective January 1, 2025) | GenAI communication containing patient CLINICAL information. Expressly excludes scheduling, billing and clerical matters | Disclaimer that the communication is AI-generated, plus clear instructions for reaching a human provider. In audio, spoken at the start AND at the end | No standalone fine. Enforced through facility and clinic licensure and the Medical Board of California |
| California — AB 489 (effective January 1, 2026) | AI using terms, post-nominals or design elements implying licensure | The agent may not be named or described as a doctor, nurse or clinician | Each use is a separate violation; licensing-board enforcement |
| California — SB 243 (operative January 1, 2026) | A reasonable person could be misled into believing they are speaking with a human | Clear and conspicuous AI notification; crisis protocol for self-harm content | Private right of action: actual damages or $1,000 per violation, plus fees |
| Texas — HB 149 (TRAIGA, in force January 1, 2026) | Use of an AI system in a patient's diagnosis or treatment | Clear and conspicuous disclosure to the patient or their representative before or at the time of the interaction | Attorney General-exclusive enforcement |
State AI-disclosure obligations most relevant to a medical front-desk voice agent, August 2026.
Utah is the sharpest rule and the one to design against. Under the Utah AI Policy Act as amended by SB 226, effective May 7, 2025, a person providing services in a regulated occupation — health care qualifies — must disclose that the consumer is interacting with generative AI at the start of the interaction, and orally if the interaction is verbal. Not in a portal. Not in a privacy notice. Spoken, at the top of the call. The Division of Consumer Protection can assess up to $2,500 per violation and the Attorney General or a court up to $5,000 per violation.
California AB 3030 is the rule most often misapplied. It does not reach a scheduling agent: the statute expressly excludes appointment scheduling, billing and other clerical or business matters. Vendors sometimes cite it as a reason to buy their compliance module for a booking bot. That is the wrong reading. But the moment the same agent discusses clinical information — symptoms, results, medications — AB 3030 applies, and in an audio interaction the AI disclaimer must be spoken at the start and at the end, along with clear instructions for reaching a human provider. There is an exemption where a licensed human provider reads and reviews the communication.
California AB 489 constrains naming. Effective January 1, 2026, it bars AI from using terms, post-nominals or design elements implying licensure, with each use a separate violation. "Nurse Ava" is not a branding decision in California; it is a compliance decision. Texas TRAIGA (HB 149, in force since January 1, 2026) attaches when AI is used in a patient's diagnosis or treatment, requiring clear and conspicuous disclosure before or at the time of the interaction — another reason to keep the agent structurally out of clinical territory.
Engineering Reality: Latency, Escalation, Records
A phone call is an unforgiving interface. There is no spinner, no "typing" indicator, and no way to re-read the last message. Silence is the only feedback channel, and patients interpret it as a dropped call.
The conversational target is under 500 milliseconds voice-to-voice — from the moment the patient stops speaking to the first audio of the reply. The dominant cost inside that budget is the language model's time-to-first-token, typically 300 to 800 milliseconds. Everything else is comparatively cheap.
| Stage | Typical budget | Notes |
|---|---|---|
| Endpointing / turn detection | Tens of milliseconds, plus the silence you deliberately wait out | Aggressive endpointing talks over patients. Slow endpointing feels dead |
| Streaming speech-to-text | Overlaps the utterance; the final transcript lands shortly after speech ends | Streaming recognition hides most of this cost |
| LLM time-to-first-token | 300–800 ms | The bottleneck. Everything else is rounding error beside it |
| Text-to-speech time-to-first-audio | Tens of milliseconds on current engines | This is the number marketed as “sub-75ms.” It is not the round trip |
| Telephony and network transport | Varies by carrier path and codec | Frequently the difference between a good demo and a bad production call |
| Target: total voice-to-voice | Under 500 ms | Past roughly a second, callers start talking over the agent |
Where the voice-to-voice latency budget actually goes in a production healthcare voice agent.
Escalation to a human is a safety control, not a fallback
Hard triggers. Any clinical question. Any mention of chest pain, difficulty breathing, severe bleeding, self-harm or another red-flag symptom. Any request the agent has failed twice. These fire on vocabulary matching, not on model judgment — you do not want the transfer decision to depend on the same reasoning that just failed.
Caller-initiated. A spoken phrase and a keypress that always work, announced in the first ten seconds. Honored on the first request, not the third. Nothing damages trust faster than an agent that argues with a patient asking for a person.
Time-boxed. If a call exceeds a set duration without reaching a completed task, transfer it. Long calls are failing calls.
Warm, with context. Transfer the transcript, the caller's identity and the attempted task to the staff member. A cold hand-off that makes the patient start over is worse than never having answered. And publish the escalation rate on a weekly dashboard — an escalation rate that is suspiciously low usually means the agent is failing silently rather than succeeding.
After-hours deserves its own boundary. An agent can take a message, book into tomorrow's grid, answer hours and directions, and route an urgent caller to the on-call line. It should not decide whether something is urgent. Encode the escalation path so that any ambiguity resolves toward a human, and make sure the after-hours behavior is tested as deliberately as the daytime behavior — that is when the calls that matter most arrive.
Limitations and Failure Modes
These are the failure modes we plan for on every build, because they show up on every build. None of them are reasons not to deploy a voice agent. All of them are reasons to deploy one with instrumentation.
| Failure mode | What it looks like in production | Control |
|---|---|---|
| Accents, dysarthria, aphasia, stutter | Recognition accuracy degrades, escalation rate rises, callers repeat themselves | Measure recognition and escalation by caller cohort. Make the human path instant and obvious |
| Speakerphone, car noise, background speech | Endpointing errors — the agent interrupts or waits forever | Tune endpointing per deployment. Add a barge-in-friendly script |
| Transcription hallucination | Fabricated content in a record that looks authoritative | The agent's transcript is a draft. A human confirms anything consequential |
| Stale payer or schedule data | Confident wrong answers about coverage or availability | Read live where possible. Never let the agent commit the practice financially |
| Ambiguous clinical framing | A scheduling question that is really a symptom question | Hard escalation triggers on symptom vocabulary, not on model judgment |
| Silent degradation after a model or prompt update | Yesterday's working flows quietly regress | Golden call sets replayed on every change, with pass thresholds in CI |
| Callers who simply want a person | Frustration, complaints, lost appointments | Announce the human path in the first ten seconds and honor it on the first request |
Known failure modes for healthcare voice agents and the controls that contain them.
The accent and speech-difference problem deserves particular attention, because it has an equity dimension that is easy to miss in an aggregate containment metric. Recognition accuracy degrades on accented speech, on dysarthria and aphasia after stroke, on speech affected by Parkinson's disease, and on stuttering. Those are disproportionately the patients for whom a phone call is the primary access route. An agent that quietly performs worse for them while the overall containment rate looks healthy is a system that has shifted burden onto the people least able to absorb it. Measure by cohort or you will not see it.
One more limitation that is organizational rather than technical: a voice agent shifts work rather than eliminating it. Calls the agent handles disappear from the queue; calls it escalates arrive at staff pre-loaded with context but also pre-loaded with a frustrated caller. Budget for the retraining, the new dashboards, and the person who owns the agent's behavior. Systems without an owner drift.
The EHR Write Ceiling
The difference between an agent that books an appointment and an agent that emails a task to your front desk is most of the value. Whether you get the first one is largely determined by your EHR vendor, not by your voice-agent vendor — and not by federal rules.
The federal certified-API requirement at 45 CFR §170.315(g)(10) requires health IT modules to support API-enabled read services, and those services expressly exclude write capabilities. The interoperability mandate you have heard about does not give your agent the right to write anything.
| Constraint | What it actually is | What it means for your build |
|---|---|---|
| Federal certified-API requirement, 45 CFR §170.315(g)(10) | Read services only — write capabilities are expressly excluded | Every write your agent performs is vendor-discretionary, not something you can demand |
| Epic scheduling operations | Appointment.$find and Appointment.$book, plus Schedule and Slot reads (STU3, with R4 variants) | Real booking is possible and slot-constrained |
| Epic app lifecycle | Apps are immutable once marked production-ready; changes require a new app record | Version your agent's integration deliberately — you cannot hot-patch it |
| Per-customer access | Each organization signs its own API subscription agreement and issues its own client ID and secret | The gate is per-customer, not per-app. Budget onboarding time per site |
| Anything else | Write paths outside documented operations | Assume a task queue for staff rather than a direct write, until proven otherwise |
The write ceiling for AI agents integrating with certified EHRs, 2026.
The good news is that scheduling is one of the places where real write paths do exist. Epic on FHIR documents Appointment.$find and Appointment.$book along with Schedule and Slot reads, so slot-constrained booking is genuinely achievable. Epic's own documentation and sandbox are free; the gate is per-customer rather than per-app, since each organization signs its own API subscription agreement and issues its own client ID and secret. Budget onboarding time per location accordingly, and note that Epic apps are immutable once marked production-ready — changes require a new app record, so plan your versioning before you ship.
Cost Bands and Timelines
Frenchy Digital publishes its bands because the rest of this category does not. These are the ranges we scope AI front-desk work into, with the assumptions behind each tier stated plainly.
| Engagement | Range | Timeline | Typical scope |
|---|---|---|---|
| Discovery + workflow audit | $9k–$22k | 2–4 weeks | Call-flow mapping, volume and containment baseline, state disclosure posture, TCPA review, EHR write feasibility, written scope |
| Single-workflow agent (intake, reminders, eligibility) | $25k–$65k | 4–9 weeks | One bounded workflow end to end, escalation design, disclosure scripting, logging, golden call set, staff training |
| Multi-workflow practice automation with EHR integration | $65k–$160k | 9–16 weeks | Booking, reschedule, reminders and intake with live scheduling reads and writes, human-in-the-loop queues, dashboards |
| Multi-site / regulated build, HIPAA posture + HITL + audit logging | $160k–$400k+ | 14–24 weeks | Multi-location routing, BAA-covered infrastructure, encryption and access controls, full audit trail, per-site onboarding |
Frenchy Digital cost bands for AI voice-agent and practice-automation work, 2026.
Senior-led delivery runs $150 to $225 per hour. Ongoing retainers, which cover model and prompt updates, golden-call-set expansion, regulatory monitoring, incident response and a quarterly review, run $2,500 to $9,500 per month depending on scope and site count. Every engagement carries a 30-day post-launch warranty, and you receive a written scope and fixed-price phased proposal within 5 business days of the discovery call.
Red Flags When Buying a Voice Agent
We tell every practice this list even when they end up buying from someone else. If a vendor trips three of these, walk.
| Red flag | Why it matters |
|---|---|
| Quotes 7% abandonment, 4.4-minute holds or 85%-never-call-back | None of those can be sourced. If slide one is unsourced, assume the performance slide is too |
| Reports containment or resolution without a denominator | “97% resolved” of what? Ask for the call population, the exclusions, and who adjudicated “resolved” |
| Advertises “sub-75ms” response time | That is text-to-speech time-to-first-audio. Ask for measured voice-to-voice p50 and p95 on your carrier path |
| Claims to beat an industry benchmark | There is no neutral healthcare voice-agent benchmark. Ask them to name it |
| Will not sign a BAA, or calls itself “HIPAA-certified” | No such certification exists. A vendor touching PHI is a business associate and the BAA is not optional |
| No published pricing and no per-unit metric | Insist on cost per handled call and cost per completed booking, then model it against your real volume |
| Outbound reminders without frequency and length caps in code | The caps are one per day, three per week, one minute, 160 characters. A policy document does not stop a scheduler |
| Bundles billing or recall marketing into the “healthcare” exemption | Both are expressly excluded. This is the most common way a practice buys a TCPA lawsuit |
| Cannot show the disclosure script and exactly where it fires | In Utah it must be spoken at the start of the call. “It is in the terms of service” is not compliance |
| Names the agent something clinical | California AB 489 treats each licensure-implying use as a separate violation |
| No escalation path except “leave a message” | Escalation is a safety control. It needs warm transfer, business-hours rules and a measured rate |
| Treats the agent transcript as the record | Transcription hallucination is documented. The transcript is a draft a human confirms |
| Refuses to hand over prompts, recordings, transcripts and logs | You need them to audit, to defend a TCPA claim, and to leave |
The Frenchy Digital red-flag checklist for healthcare voice-agent buyers, 2026.
A vendor who cannot source their market statistics, cannot define their containment denominator, and cannot show you where the disclosure fires is asking you to accept their word on three separate things you could have verified. That is the pattern, not a coincidence.
— Frenchy Digital buyer's principle
How Frenchy Digital Builds One
Frenchy Digital is a senior-led, Black-owned Los Angeles agency. We build administrative automation for medical practices — booking, reminders, intake, eligibility — under human review. We do not build clinical decision tools, and we are explicit with every client that an AI agent does not practice medicine. The American Medical Association's framing is the right one: augmented intelligence, assistive by design, enhancing human judgment rather than replacing it.
- Week 1–2: measure before you automate: We instrument your current call flows first — volume by hour, abandonment, transfer paths, task mix, after-hours pattern. You cannot claim a containment improvement without a baseline, and almost nobody has one.
- Week 1–2: legal posture, in writing: State disclosure duties for every state you call into, TCPA campaign-type mapping for every outbound flow, BAA scope, and a confirmed answer on EHR write access. This is a deliverable, not a conversation.
- Week 2–4: one workflow, end to end: A single bounded workflow shipped to production with escalation, disclosure, logging and a golden call set. One workflow working beats five workflows demoing.
- Ongoing: golden call sets in CI: Thirty to two hundred real, de-identified call scenarios with expected outcomes, replayed on every prompt, model or integration change. This is what stops a model upgrade from silently breaking your booking flow.
- Ongoing: cohort-level measurement: Recognition accuracy and escalation rate broken out by caller cohort, not just in aggregate, so degradation for accented and non-standard speech is visible rather than averaged away.
- Human-in-the-loop as a control, not a courtesy: Anything consequential is confirmed by a person before it lands in a system of record. Review that is a rubber stamp is not review — design the queue so the reviewer has enough context and enough time to actually disagree.
- Handover: you own everything: Source code, prompts, call recordings, transcripts, evaluation sets, dashboards and cloud accounts transfer to your practice. Full IP ownership, 30-day post-launch warranty, no lock-in.
If you want a grounded read on whether a voice agent fits your practice — including an honest answer when it does not — book a free 60-minute discovery call at calendly.com/frenchydigital/discovery-call or call +1 (424) 272-5601. You leave with a written scope and a fixed-price phased proposal within 5 business days.
Get a Straight Answer on AI Voice Agents for Your Practice
Book a free 60-minute discovery call with Frenchy Digital, a senior-led Black-owned LA agency. We map your call flows, EHR write path, disclosure duties and TCPA exposure — then send a written scope and fixed-price phased proposal within 5 business days.
Thinking About an AI Phone Agent for Your Practice?
Book a free 60-minute discovery call with Frenchy Digital. We map your call flows, your EHR write path, your state disclosure duties and your TCPA exposure — then send a written scope and fixed-price phased proposal within 5 business days.
1517 S Bentley Ave Unit 204, Los Angeles CA 90025
Frequently Asked Questions
Sources & References
- 1FCC Declaratory Ruling 24-17 — AI-generated voices are “artificial voices” under the TCPA (CG Docket 23-362)↗
- 247 CFR §64.1200 — Delivery restrictions, including the healthcare exemption at (a)(9)(iv)↗
- 3Utah SB 226 (2025) — Artificial Intelligence Consumer Protection Amendments↗
- 4California AB 3030 — GenAI in patient communications (Health & Safety Code §1339.75)↗
- 5California AB 489 — Board advisory on AI implying licensure↗
- 6Texas HB 149 (TRAIGA) — health care provider AI disclosure duty↗
- 7Cureus 2026;18(7):e112227 — AI-driven call center operations in a high-volume neurology practice↗
- 8npj Digital Medicine — Mount Sinai “Sofiya” pre-procedure voice agent, July 4, 2026↗
- 9“Careless Whisper: Speech-to-Text Hallucination Harms” — ACM FAccT 2024↗
- 10ASTP/ONC — Standardized API for Patient and Population Services, §170.315(g)(10) (read-only)↗
- 11Epic on FHIR — scheduling operations and app launch documentation↗
- 12AMA — Augmented Intelligence in Medicine↗
- 13AMN Healthcare 2025 Survey of Physician Appointment Wait Times↗
- 14HHS — HIPAA Privacy and Security Rules↗
