AI Agents for Patient Schedulingand No-Show Reduction

Search for the cost of patient no-shows and you will find the same two figures everywhere: a no-show costs a practice $200, and no-shows cost the US health system $150 billion a year. They appear in vendor decks, in analyst summaries, in business-press articles, and in the return-on-investment models that scheduling software is sold on. They are not research findings.

Both figures trace to a single 2017 opinion byline written by the chief medical officer of a patient-scheduling software vendor. There is no study behind them. No dataset, no sample, no stated methodology, no peer review. When you follow the citation chain on any article that repeats them — including the business-press pieces that lend them credibility — the chain terminates at that one byline. A number repeated ten thousand times is still one number, and one number with no method behind it is a marketing claim.

The same is true of the specialty table that circulates alongside them, the one that puts dermatology at 30 percent and sleep medicine at 39 percent. We could not trace it to any primary source. It appears on vendor blogs with no attribution and gets copied forward. So do the phone-operations benchmarks that show up in voice-AI marketing: 7 percent call abandonment, a 4.4-minute average hold, 85 percent of callers who never call back. None of them resolve to a measurement study.

Audit of the four figures most commonly used to justify scheduling automation spend, 2026.

We will not build a business case on a number we cannot trace to a method. Neither should you — and the first thing to ask any scheduling vendor is where their headline figure came from.

— Frenchy Digital principle

The cost of a no-show is not a property of healthcare. It is a property of your schedule. In a practice running at capacity with a functioning waitlist, the marginal cost of a missed appointment approaches zero, because the slot backfills. In an under-booked practice with no waitlist, it approaches the full contribution margin of the visit. Those two practices can sit on the same street, in the same specialty, and their real numbers differ by an order of magnitude.

Here is the calculation, using inputs you already have. Do it per visit type — a new-patient consult and an established follow-up are not the same asset.

Seven-step method for computing a practice's own no-show exposure from contribution margin.

Segment the result. In most practices we audit, a small number of visit types — high-margin procedures, long new-patient blocks, imaging-adjacent slots — carry the overwhelming majority of the exposure, while the bulk of missed appointments are short follow-ups that cost very little to lose. Automating reminders uniformly across all of them spends the same effort on both. Targeting does not.

There are real benchmarks. The problem is that they get quoted without their scope, which is how a safety-net figure ends up in a suburban orthopedic practice's board deck.

No-show benchmarks with their scope stated — the scope is the point.

The most-cited figure in the field is the 23 percent average from Dantas and colleagues (2018), a review of 105 studies. It is a legitimate piece of work and it is routinely misused. The review is global, it spans public and safety-net systems, and it includes settings whose access conditions have nothing in common with a US private practice. Presenting 23 percent as "the industry average" to a US group inflates the projected benefit of any intervention by roughly a factor of three before the pilot even starts.

Meanwhile MGMA's single-specialty aggregate sits at 6.81 percent for 2023, against roughly 7 percent in 2019. Flat. Whatever the last five years of scheduling technology accomplished, it did not move that aggregate.

The instructive comparison is inside a single system. A VA cluster-randomized trial reported 11.9 percent in primary care and 18.0 percent in mental health — same organization, same technology, same reminder infrastructure, six percentage points apart. An urban FQHC network in New York reported 41.6 percent. General dental practice sits near 14.3 percent, with adolescents at 24 percent.

No-show rate is a function of population, not of specialty glamour — which means the honest benchmark for your practice is your own last twelve months, segmented.

— The lesson in one sentence

This is the table the rest of the article exists to support. Read the "outcome measured" column carefully — some of these risk ratios are on attendance, where higher is better, and some are on missed appointments, where lower is better. Vendors mix the two, which makes weak results look strong.

Randomized and systematic-review evidence on interventions to reduce missed appointments, with outcome direction stated explicitly.

The Cochrane review of mobile phone messaging reminders (CD007458.pub3) is the anchor. SMS reminders beat no reminder at a risk ratio of 1.14 (95% CI 1.03–1.26) on attendance, with pooled attendance moving from 67.8 percent to 78.6 percent. That is a real, replicated effect — and it is also the effect nearly every US practice has already captured, because nearly every US practice already sends reminders. The marginal value of the eleventh reminder channel is not the same as the value of the first.

The same review puts SMS against telephone reminders at RR 0.99. There is no meaningful difference. A text costs a fraction of a cent; a staffed reminder call costs a minute or two of labor plus the callback tail. If you are still calling your whole panel to remind them, you are paying for a result you can buy at a thousandth of the price.

Where the picture changes is targeting. The 2023 JAMIA rapid systematic review of predictive-model-based interventions reports RR ~0.61 for ML-targeted phone calls and RR 0.55 for patient navigators on missed appointments, versus RR ~0.91 for ML-targeted texts. The same review rates predictive overbooking as very low certainty — the reviewers were not confident it did anything.

Put the numbers next to each other and the conclusion writes itself. Untargeted SMS to everyone: RR 1.14 on attendance, essentially free, and already deployed almost everywhere. Targeted phone calls: RR ~0.61 on missed appointments, expensive per unit. Navigators: RR 0.55, more expensive still. The gap between the cheap universal intervention and the expensive selective one is not explained by the channel. It is explained by who receives it.

So the design of a scheduling agent that actually earns its cost is not "replace texts with AI calls." It is a two-tier allocation:

• Tier one — everyone gets the automated text: Within the TCPA ceiling, on a cadence you can defend. Near-zero marginal cost, well-evidenced effect, no staff time. This is the floor, not the strategy.
• Tier two — scarce human contact goes where the model points: Staffed calls and navigator time allocated to the appointments with the highest predicted risk and the highest contribution margin. This is where RR 0.61 and RR 0.55 live.
• Tier three — the freed capacity gets recycled: Cancellations captured in-channel and offered automatically to a waitlist. Roughly one in nine offers converts, and those patients get seen materially sooner.
• What does not get a tier: Overbooking (very low certainty) and no-show fees (null result). Neither belongs in a payback model.

There is a second-order benefit worth naming. Cancellations are more valuable than prevented no-shows, because a cancellation arrives with lead time and a no-show does not. An agent that makes cancelling trivially easy — same channel, no hold music, no guilt — converts silent no-shows into recoverable slots. Practices often resist this because it feels like inviting cancellations. The schedule disagrees.

No-shows are a symptom. The condition underneath them is access, and the access data is worse than the no-show data. AMN Healthcare's 2025 survey of 1,391 physician offices across 15 metropolitan markets found an average new-patient appointment wait of 31 days — up 19 percent since 2022 and 48 percent since 2004.

AMN Healthcare 2025 physician appointment wait times, 1,391 offices across 15 metros.

Third-next-available — the operational measure that actually reflects capacity, rather than the first cancellation hole in the calendar — runs a mean of 11.0 days (SD 6.0). And waiting is not neutral: research published in JAMA Network Open (2022) found that each additional week of wait reduced good-or-excellent patient ratings of wait time by 7.35 percent. Access is a satisfaction variable and a retention variable, not only a revenue one.

The demand-versus-supply picture for self-scheduling is where the opportunity is most explicit.

Patient demand for digital scheduling against provider supply, 2025 surveys.

A scheduling agent is administrative automation. It books, confirms, reminds, cancels, rebooks, and backfills. It does not triage, it does not advise, and it does not make clinical decisions — which is the framing the AMA captures with the term "augmented intelligence": assistive by design, not autonomous. Every workflow below runs under human escalation, and every clinical question routes to a person.

The nine stages of an administrative scheduling agent and the control on each.

One compliance point is worth stating precisely because it is so often gotten backwards. California's AB 3030, which governs generative AI in patient communications, applies to clinical information and expressly excludes "administrative matters, including appointment scheduling, billing, or other clerical or business matters." A scheduling agent that books visits sits outside AB 3030. The same agent, the moment it discusses symptoms or results, sits inside it — with disclosure obligations that for audio must be spoken at both the start and the end of the interaction. The boundary is the content, not the technology.

Separately, Utah's AI Policy Act as amended by SB 226 requires a person operating in a regulated occupation — health care included — to disclose generative AI at the start of an interaction, orally if the interaction is verbal. If you deploy a voice agent in Utah, the disclosure is not optional and it is not a footer.

The healthcare exemption in the Telephone Consumer Protection Act is what makes automated appointment reminders workable — and it is narrower than most reminder configurations assume.

TCPA healthcare exemption constraints applicable to automated appointment reminders, 47 CFR 64.1200.

The frequency and length limits are the easy part; a well-built agent enforces them in code and logs every send. The trap is content. The exemption does not cover telemarketing, billing, or debt collection. A reminder that helpfully adds "you have a balance of $42 due at check-in" has just left the exemption, and TCPA damages are $500 per violation and $1,500 for willful violations — per message, across a panel, which is how a well-intentioned template becomes a class action.

For voice, the FCC's Declaratory Ruling 24-17 (CG Docket 23-362) established that AI-generated voices are "artificial voices" for TCPA purposes. That ruling remains in force. The separate AI-disclosure rulemaking proposed in August 2024 has not been finalized, so the disclosure obligations that bind a voice agent today come from state law rather than the FCC.

• Build the cadence limit into the agent, not the playbook: One message per day and three per week per patient, enforced at the send layer across every channel and every campaign, including recall and marketing systems that share the same phone number.
• Keep billing content in a separate, consented channel: Reminders and balances must not share a template. The exemption is content-scoped.
• Honor opt-out immediately and prove it: Timestamped suppression, effective across all downstream systems, with an audit trail you can produce.
• Log every send: Recipient, timestamp, channel, template version, character count, and consent state. TCPA defense is a records problem.

Here is the load-bearing architectural fact, and it surprises almost every buyer: the federal certified-API mandate is read-only. ASTP/ONC's §170.315(g)(10) test method requires health IT modules to support API-enabled read services and states that those services "specifically exclude 'write' capabilities."

The consequence is direct. Every write your scheduling agent performs — booking, cancelling, rescheduling — is vendor-discretionary. It is a commercial and technical negotiation with each EHR vendor and each customer organization, not a regulated guarantee you can assume across a multi-site rollout.

Epic is the case worth knowing in detail, because it does expose real scheduling writes. Epic on FHIR documents Appointment.$find and Appointment.$book (STU3), alongside Schedule.Read and Slot.Read, with R4 variants. Booking is slot-constrained: the agent searches real availability and books into an existing slot rather than writing an arbitrary calendar entry. That constraint is a feature — it makes the EHR the source of truth and prevents the agent from inventing capacity.

The gate is per customer, not per app. Each organization must sign the open.epic API Subscription Agreement, have staff holding the "Able to Purchase Apps" security point, download by Client ID, and set a client secret. Apps are immutable once marked production-ready — changes require a new app record, so your release process has to plan for it. Epic's sandbox, client-ID registration, and SMART and CDS Hooks documentation are free; Epic states that direct access to its software is not needed to develop, test, deploy, or support a product. The API launch guide is the authoritative reference.

A predictive no-show model is trained on historical attendance. Think about what that history contains. It contains who has a car and who depends on a bus that runs every 40 minutes. It contains who can take a Tuesday morning off without losing pay and who cannot. It contains who has childcare and who brings a toddler to a specialist appointment. The model does not learn "unreliability." It learns disadvantage, and it outputs it as a risk score.

That is not a reason to avoid the model. Targeting is, on the evidence, the highest-value thing you can do with a scheduling system. It is a reason to write down — before deployment, in policy, not in a hallway conversation — what the score is allowed to change.

Permitted and prohibited uses of a no-show risk score, stated as policy.

Pointed at extra help, a no-show model is an equity intervention. Pointed at access — denied bookings, required deposits, systematic double-booking — the same model becomes a machine for concentrating harm on the patients with the least slack in their lives.

— The distinction that matters

And this is not confined to predictive models. The same 2024 JMIR study of automated waitlist backfill — a feature that looks entirely benign, because it only ever offers an earlier appointment — found that patients aged 65 and older, patients of other ethnicity relative to White patients, and primarily Chinese-speaking and other non-English-speaking patients were less likely to accept an offer.

Think about why. A short-notice offer rewards whoever can respond fastest, and the ability to respond fastest is itself a form of privilege: schedule flexibility, a job that tolerates interruption, childcare that can move, a phone you read in your first language, no need to arrange a ride or an interpreter. So a backfill feature that reports a clean aggregate lift is simultaneously redistributing access toward the patients who already had the easiest time getting seen. Nothing in the feature is discriminatory by design. The redistribution happens anyway, because the mechanism is speed of response.

There is a useful due-diligence artifact here. Under ASTP/ONC's HTI-1 rule, certified health IT developers must publish 31 source attributes for Predictive Decision Support Interventions across nine categories — including fairness in development, external validation, quantitative performance, and ongoing maintenance — and must run Intervention Risk Management for each one, with a summary posted publicly. The rule binds developers, not practices. But if the no-show model you are being sold is embedded in certified health IT, that published transparency data is your ready-made vendor questionnaire, already answered, already public. Ask for it by name.

An honest scoping conversation includes the parts of this that do not work, or are not known to work. Here they are.

• There is no randomized evidence for AI voice scheduling agents: As of August 2026 there is no RCT, and nothing in JAMA, NEJM, or JAMIA. The peer-reviewed literature is two 2026 single-site studies with no control arm: a Cureus neurology report describing a backlog reduction, and a Mount Sinai study of 1,431 patients in npj Digital Medicine where appointment completion moved from 86.4% to 87.9%.
• Vendor performance claims are self-reported and unaudited: The resolution, containment, and booking percentages circulating in this market come from vendors. None have been independently audited. Ask for the denominator, the definition of 'resolved,' and the exclusion criteria before treating any of them as a benchmark.
• The reminder evidence is strong but old and moderate at best: The Cochrane finding is robust and replicated, but its certainty is low to moderate and the trials predate modern channel mixes. It tells you reminders work. It does not tell you your eleventh channel will.
• Overbooking has very low certainty and real externalities: The JAMIA reviewers were uncertain it affected attendance at all, and the cost of being wrong lands on the patients who did show up and on the clinicians running behind.
• No-show fees produced a null result: 5% versus 5% in a Danish randomized trial. If a payback model includes fee revenue or fee-driven attendance gains, it is modeling something that was tested and did not happen.
• Waitlist backfill is miscounted constantly: 11% offer acceptance and a median 14 days earlier is real value. It is not no-show reduction, and counting it as such double-counts the benefit in almost every ROI model we review.
• Speech transcription is a genuine risk surface: The 'Careless Whisper' study at ACM FAccT 2024 found roughly 1% of transcriptions wholly hallucinated, with 38% of hallucinations containing explicit harms. A voice agent's transcript must never write an unreviewed value into a clinical field.
• Ceiling effects are the most common cause of a disappointing pilot: A practice already at 6–7% no-shows with universal reminders has very little headroom. The honest projection there is capacity recovery and staff time, not attendance.
• Write access is not guaranteed anywhere: The certified-API mandate is read-only. Any multi-site plan that assumed uniform booking writes should be re-baselined before contracts are signed.

These are Frenchy Digital's own scoping bands for scheduling and patient-access automation in 2026. They are consistent across every engagement we quote, and every one of them is a fixed-price phased proposal rather than an open meter.

Frenchy Digital cost bands for scheduling and patient-access automation, 2026.

Senior-led delivery is priced at $150–$225 per hour. Ongoing operations run $2,500–$9,500 per month depending on channel volume, number of sites, and whether a risk model is in production and therefore requires monitoring. Every engagement includes a 30-day post-launch warranty, and you receive a written scope and a fixed-price phased proposal within 5 business days of the discovery call.

Most of these are visible in the first sales deck, before anyone touches your data.

Frenchy Digital buyer's checklist for scheduling and patient-access automation, 2026.

Ask one question early: "Which randomized study supports the largest number in this proposal?" The answer, or the absence of one, tells you more about the vendor than a six-week pilot will.

— Frenchy Digital buyer's principle

Frenchy Digital is a senior-led, Black-owned Los Angeles agency. On patient-access work we hold to six commitments:

How Frenchy Digital scopes and delivers scheduling automation, 2026.

• Week 1–2 — measurement: Your no-show data segmented by visit type, lead time, provider, channel and payer; backfill rate; third-next-available; contribution margin per visit type. This produces your number, not an industry number.
• Week 2–4 — evidence-weighted scope: We map each proposed intervention to its evidence base and its cost, and we tell you which parts of the vendor pitch you should decline.
• Week 4–9 — single workflow, shipped: One workflow end to end with TCPA cadence enforced in code, consent and send logging, human escalation, and a concurrent control group so the result is attributable.
• Week 9–16 — integration and targeting: EHR write path behind a swappable adapter, waitlist backfill, and targeted outreach governed by a written model-use policy.
• Ongoing — monitoring: Drift and calibration monitoring on any production risk model, cadence and consent audits, quarterly review against the pre-registered outcome.