AI Agents for DentalPractice Management in 2026

If you have sat through more than two dental AI demos in 2026, you have seen the same four figures. A practice loses $47,000 a year to no-shows. Somewhere between 32 and 38 percent of inbound calls go unanswered. Each missed call is worth $250 to $350. And the whole thing is anchored by a citation reading something like "ADA Practice Transitions: 38% of new-patient calls unanswered."

We went looking for the sources. There are none. Not a study, not a dataset, not a methodology, not a year. The ADA attribution in particular could not be located in any ADA publication — which means the most authoritative-sounding number in the category is the one with the weakest provenance. These figures propagate because each vendor cites the previous vendor, and nobody walks the chain back to a primary document.

The dental and adjacent healthcare-AI figures that fail a citation check, and what a primary-source review actually returns.

There is one dental economics figure that survives review. The ADA Health Policy Institute reports general-practitioner net income of $215,320 for 2025. The widely quoted companion figure — "$965,660 in average gross billings" — is not ADA-verified, so it does not appear anywhere else in this article.

Real dental attendance data exists too. The peer-reviewed dental non-attendance rate is 14.3 percent overall, rising to 24 percent for adolescents (PMC12274092). Note how far that is from the numbers used to sell voice agents — and note that even the 14.3 percent is a literature average, not your practice.

The medical side of the industry is no cleaner. The "$200 per no-show" and "$150 billion a year" figures that anchor most healthcare scheduling pitches both trace to a single 2017 byline written by the chief medical officer of a scheduling vendor, with no study or methodology behind it. Every downstream citation — trade press included — leads back to that one article.

A vendor who invents your problem will invent your results. The practices that get real value out of dental AI are the ones that measured their own baseline before anyone showed them a slide.

— Frenchy Digital buyer's principle

Every number the category invents already exists inside your own systems. Your phone platform holds call-detail records. Your practice-management system holds appointment statuses, production per provider, recall aging and treatment-planned-not-scheduled dollars. None of it requires a vendor, a pilot or a purchase order to extract.

The eight metrics a dental practice can compute from its own data before speaking to a single AI vendor.

The sequencing matters more than the tooling. Measure before you change anything, or you will never be able to attribute a result to the agent rather than to the attention the project itself generated.

• Week 1 — Pull raw exports only: Call-detail records, appointment status report, production by provider, recall aging, treatment-planned-not-scheduled aging, claim adjustment reasons. No interpretation yet, and no process changes.
• Weeks 2–5 — Hold everything constant: Four clean weeks of baseline. Resist the urge to fix the obvious problems now; you need an untouched comparison period more than you need four weeks of improvement.
• Week 6 — Segment: Split every metric by hygiene versus restorative, by provider, by day of week, and by hour. Aggregate rates hide where the loss actually sits — a 14 percent broken rate concentrated in one hygiene column is a scheduling problem, not an AI problem.
• Week 7 — Price the gap in your own dollars: Multiply unfilled chair hours by your own production per scheduled hour, then discount by the share that historically gets backfilled. This produces a defensible number that is yours. It will not match any vendor's.
• Week 8 — Write the target and the guardrail before you shop: One target metric, one acceptable range, one guardrail metric that must not degrade (patient complaints, opt-out rate, staff escalations). Any vendor unwilling to be measured against a pre-agreed baseline is telling you something.

For context on what actually moves attendance, the evidence base is genuinely good — it just is not the evidence vendors quote. The strongest findings come from Cochrane and from a 2023 JAMIA systematic review, and the pattern is consistent: reminders work, targeting works better than volume, and financial penalties do nothing at all.

What the peer-reviewed literature says about reducing non-attendance. Note that texting performs as well as calling (Cochrane CD007458.pub3), and that no-show fees do not work.

This is where dentistry is genuinely ahead of most of medicine. Radiographic AI in dentistry is not a pilot category or a research curiosity — it is a regulated device category with a public clearance record you can read yourself. That record is also the most miscited thing in dental technology coverage, which makes getting it right a differentiator on its own.

Dental AI imaging 510(k) record as of August 2026, read from FDA's own database rather than from vendor material.

Pearl holds eight 510(k) clearances on record. The most recent is K250525, Second Opinion Panoramic, with a decision date of November 14, 2025, covering caries, periapical radiolucency and impacted third molars on panoramic radiographs for patients aged 16 and older.

Overjet holds eleven. The most recent is K253930 (Iris), decided April 10, 2026. Overjet's CBCT Assist is K251514, decided December 5, 2025.

You can verify all of this in under a minute. The 510(k) Premarket Notification database is searchable by applicant name and K-number, and the openFDA device API returns the same records programmatically. Ask every imaging vendor for their K-numbers in writing, then look them up before the second meeting.

A 510(k) is a premarket notification cleared on the basis of substantial equivalence to a legally marketed predicate device. That is a meaningful regulatory bar and it is worth insisting on. It is also routinely oversold, so here is the honest reading.

• It is a device authorization, not a performance promise: Clearance says the device is substantially equivalent to a predicate. It says nothing about how the software will perform on your sensors, your exposure settings, your image quality or your patient population.
• Cleared is not approved: "FDA-approved" describes a different pathway entirely. A vendor using that phrase for a 510(k) device is either careless or deliberately imprecise, and neither is what you want in a contract counterparty.
• It is not an endorsement: FDA does not rank, recommend or certify comparative quality between cleared devices. Two products with clearances for the same indication can perform very differently.
• The dentist remains the diagnostician: Cleared dental imaging AI is assistive. The AMA's framing for medicine — "augmented intelligence," assistive rather than autonomous — is the correct posture here too, and the AMA House of Delegates reaffirmed in June 2026 that AI is an assistive tool requiring clinician oversight, not an autonomous decision-maker.
• Ask about the PCCP: FDA finalized its Predetermined Change Control Plan guidance in December 2024. A PCCP is the mechanism by which a cleared device can ship model updates within an authorized envelope. Ask whether the product has one, and what falls outside it.

The related question — why is an imaging product a regulated device while a scheduling agent is not? — has a clean answer in the statute. FDA's Clinical Decision Support Software final guidance was issued January 6, 2026 and re-issued January 29, 2026 under docket FDA-2017-D-6569. The first of the four non-device criteria in section 520(o)(1)(E) excludes software intended to acquire, process or analyze a medical image or signal. Radiograph analysis fails that criterion by definition, which is precisely why Pearl and Overjet hold clearances.

Administrative software sits elsewhere. Software providing administrative support of a health care facility is excluded from the device definition under section 520(o)(1)(A). An agent that verifies benefits, works a recall list or answers the phone is generally outside device regulation — which is not the same as being unregulated, as the compliance section below makes clear.

One more note for buyers reading vendor claims about "FDA-aligned" AI governance: FDA's AI-Enabled Device Software Functions lifecycle guidance was issued in draft in January 2025 under docket FDA-2024-D-4488 and remains a draft as of August 2026. Draft guidance is not binding on anyone.

Here is where the picture inverts. Dentistry is ahead of medicine on imaging AI and meaningfully behind it on integration. In medicine, the federal certified-API criterion at § 170.315(g)(10)at least forces a standardized read API into every certified EHR — though note that the mandate "specifically exclude[s] 'write' capabilities," so even there every write is vendor-discretionary. Dental practice-management vendors sit under no equivalent federal API mandate at all.

The practical consequence is worth stating plainly: whether an agent can write an appointment into your schedule is a contract question, not a regulatory one. That makes the choice of practice-management system the single largest determinant of what an AI project will cost and how long it will take.

Dental practice-management integration paths as of August 2026. Open Dental is the only vendor in this table that publishes pricing.

Open Dental publishes API pricing per location: a free read-only tier limited to one request every five seconds, then paid tiers at $15, $30 and $35 per month, with developer approval stated at one to three business days. For a DSO operations lead, that is not a small detail — it means integration licensing can be modelled across fifty locations on a spreadsheet before a contract is signed.

Henry Schein One, which operates Dentrix, runs a gated API Exchange with more than 140 vendors, requiring SOC 2 Type II and OAuth2 and publishing no pricing. Integration is achievable, but the partner-program review is a schedule risk you must budget for separately from engineering time. Patterson's Eaglesoft is bridge-based rather than a modern REST API, so expect a different class of integration pattern. Curve Dental's API terms we could not verify at all — which is itself the finding, and the reason to ask for them in writing.

One further note on due diligence. In certified health IT, the HTI-1 rule obliges developers to publish transparency attributes for predictive decision-support interventions, which gives a medical practice a ready-made artifact to review. Where your dental software is not certified health IT, that artifact does not exist and you have to ask the vendor directly for model documentation, validation data and update policy.

Five workflows carry most of the realistic value in a dental practice. Each one is administrative, each one has a human checkpoint that is not optional, and each one has a specific failure mode when that checkpoint is removed.

The five dental workflows where administrative agents earn their cost — and the checkpoint each one requires.

Eligibility and benefit verification is usually the highest-value starting point, because it is high-volume, rules-based, and its errors are expensive and visible. It is also the workflow where the coverage landscape is least helpful: the CMS interoperability and prior authorization rule reaches Medicare Advantage, Medicaid and CHIP, and QHP issuers on the federally facilitated Exchanges — not commercial or ERISA plans. Most adult dental benefits sit in commercial plans, so the FHIR prior-authorization APIs arriving in 2027 will not cover them. Assume payer portals and clearinghouse traffic remain the substrate for the foreseeable future.

Human-in-the-loop is a compliance control, not a courtesy. If the reviewer cannot say how long they spent and what they rejected, there was no review — only a signature.

— Frenchy Digital delivery principle

Dentistry gets no HIPAA discount. Any vendor whose system creates, receives, maintains or transmits electronic protected health information on your behalf is a business associate and requires a signed BAA. The Security Rule as it stands today is what governs an AI agent touching ePHI — and there is no such thing as a HIPAA-certified vendor, so treat that phrase as a red flag rather than a credential.

The federal and state rules that actually bind an AI agent operating in a dental practice, and the boundaries of each.

The HIPAA Security Rule modernization proposalwas published January 6, 2025 at 90 FR 898 under RIN 0945-AA22, and comments closed March 7, 2025. It has since moved to Long-Term Actions on the Unified Agenda with a projected final action of July 2027 — the agency's own non-binding estimate. Build toward it if you like the controls; do not let a vendor sell it to you as a current obligation.

The TCPA healthcare exemption is where dental practices most often get into trouble, because recall and reactivation campaigns are exactly the kind of outreach that drifts across the line. The exemption at 47 CFR § 64.1200(a)(9)(iv) permits at most one message per day and three per week per patient, voice messages under one minute, texts of 160 characters or fewer, and an immediately honored opt-out — and it excludes telemarketing, billing and debt collection. A hygiene recall text can qualify. An outstanding-balance text does not.

AI voices carry their own overlay. Under FCC 24-17, adopted February 2, 2024, AI-generated voices are "artificial voices" for TCPA purposes, with damages of $500 per violation and $1,500 if willful. The FCC's separate AI-disclosure proposal, FCC 24-84, remains unfinalized. The one-to-one consent rule was vacated by the Eleventh Circuit in Insurance Marketing Coalition v. FCC(No. 24-10277, January 24, 2025), and the revocation "revoke-all" provision has been waived again through January 31, 2027 under DA 26-12.

On disclosure, three state rules matter most for a dental agent. Utah's AI Policy Act as amended by SB 226 is the most directly applicable rule in the country for a phone agent in a licensed practice: in a regulated occupation, disclose that the consumer is interacting with generative AI at the start of the interaction, orally when the interaction is verbal. Texas TRAIGA (HB 149), in force since January 1, 2026, requires clear and conspicuous disclosure of AI use in a patient's diagnosis or treatment before or at the time of the interaction; its companion SB 1188, effective September 1, 2025, requires practitioners to review AI-generated records and bars offshore storage of electronic medical records. California AB 489, effective January 1, 2026, bars AI from using terms, post-nominals or design elements implying licensure — each use a separate violation.

California AB 3030 is worth understanding for its boundary rather than its reach. It requires a disclaimer and human-contact instructions on GenAI communications containing patient clinical information — for audio, spoken at the start and at the end — but it expressly excludes appointment scheduling, billing and other clerical or business matters, and it is written for health facilities, clinics and physician offices. The useful lesson generalizes: the moment an agent stops booking and starts discussing findings, a different and stricter set of rules attaches.

That last item is not decoration. In its February 2026 compliance program guidance, HHS-OIGnamed querying clinicians through electronic record platforms — "including prompts generated by artificial intelligence algorithms" — as potentially abusive in the risk-adjustment context. The specific setting is Medicare Advantage, not dentistry, but the enforcement theory generalizes cleanly: a reviewer who accepts nearly every suggestion in half a second is not performing review, and automation bias is treated as a compliance failure rather than a human-factors curiosity.

If you only read one section before signing something, read this one. The evidence base for AI voice agents in healthcare is thin, and the risk evidence is currently stronger than the benefit evidence.

• There is no randomized trial: The peer-reviewed literature on healthcare voice agents consists of two 2026 single-site studies with no control arm: a Cureus report (2026;18(7):e112227) in neurology describing a backlog reduction over 98 percent, and Mount Sinai's "Sofiya" in npj Digital Medicine (July 4, 2026), covering 1,431 patients with completion moving from 86.4 to 87.9 percent. Nothing in JAMA, NEJM or JAMIA. Neither study is dental.
• Vendor performance numbers are self-reported and unaudited: Assort reports 97 percent resolution and 79 percent of referrals scheduled without staff; Notable reports 57 percent containment at Catholic Health; Hyro reports 85 percent resolution of routine interactions. These are vendor claims, not audited results, and none has been independently verified. Treat them as marketing until someone publishes a methodology.
• Transcription hallucination is a measured risk: The "Careless Whisper" study presented at ACM FAccT 2024 found roughly 1 percent of transcriptions were wholly hallucinated, and that 38 percent of those hallucinations contained explicit harms. In a practice, that is a note in a chart or a message in a patient record that nobody said.
• Latency claims are routinely misstated: Under 500 milliseconds voice-to-voice is the conversational target, and the bottleneck is model time-to-first-token at roughly 300 to 800 milliseconds. "Sub-75 millisecond" claims describe text-to-speech time-to-first-audio, not the round trip. No neutral healthcare voice-agent benchmark exists to arbitrate any of it.
• There is no reimbursement: In the CY2026 Medicare Physician Fee Schedule, CMS only solicited comment on paying for SaaS and software algorithms, stating they are not well accounted for, and finalized no general AI payment pathway. CPT Appendix S is a taxonomy of assistive, augmentative and autonomous AI services — not a payment schedule. Administrative AI is a cost-side play, full stop.
• Coding accuracy figures have no independent benchmark: Every published accuracy number for AI coding is vendor-produced. No neutral benchmark exists, so a claimed accuracy rate is an assertion about a dataset you cannot inspect.
• Small practices have small base rates: An agent that saves four hours a week in a two-chair practice is a real but modest result. Measured against a vendor's projection it will look like failure; measured against your own eight-week baseline it may be a clear win. Choose the comparison before you start.

These are our bands, stated plainly so you can compare them against anything else on your desk. They cover custom build work — the integration layer, the routing and escalation logic, the audit trail, the human-review workflow — not licence fees for third-party products.

Frenchy Digital 2026 cost bands for dental practice and DSO automation engagements.

• Senior-led rate: $150–$225 per hour. We do not staff junior engineers on regulated work.
• Ongoing retainer: $2,500–$9,500 per month, covering monitoring, model and dependency upgrades, incident response, and a quarterly review of the metrics you set in your baseline.
• 30-day post-launch warranty: On every engagement, without exception.
• Written scope in 5 business days: A fixed-price phased proposal within five business days of the discovery call. Phases are priced before they start, so asking questions does not cost you money.
• Full ownership transfers: Source code, prompts, integration configuration, audit-log schema and documentation transfer to your business at delivery. No vendor lock-in.

Budget separately for the things we do not control: practice-management API licensing (Open Dental's published per-location tiers at $15, $30 and $35 per month are the only transparent example in the category), telephony and messaging costs, model inference, and any FDA-cleared imaging product, which is a licence purchase from the device manufacturer and not something an agency builds.

These are the signals we tell every dental prospect to watch for, including the ones who go on to hire someone else. Most of them can be checked in the first meeting.

The Frenchy Digital red-flag checklist for dental AI buyers, 2026.

The vendor who cannot source their headline number, cannot name their integration model, and cannot show you where the message cap is enforced is asking you to take three separate risks on faith. Any one of them is survivable. All three together is a project that ends in a lawyer's office.

— Frenchy Digital buyer's principle

The build-versus-buy line in dentistry is unusually clean, because the regulatory boundary does most of the work of drawing it.

• Buy: FDA-cleared imaging AI: You cannot build this. Producing diagnostic radiographic software would require your own 510(k). Buy from a vendor whose clearance record you have verified in FDA's database, read the indications for use, and ask what their Predetermined Change Control Plan authorises.
• Buy: commodity plumbing: Telephony, speech-to-text and text-to-speech, SMS delivery and calendar primitives are commodities. There is no advantage in rebuilding them and no differentiation in owning them.
• Build: the integration layer: This encodes your practice-management system, your write permissions, your provider columns and your appointment types. It is also the piece that determines whether the whole project works, and it is where off-the-shelf products are most generic.
• Build: routing, escalation and message-cap enforcement: These encode your specific risk posture — which questions go to a human, which states require an oral disclosure, how many messages a patient may receive in a week. Policy documents do not enforce anything; code does.
• Build: the audit log and the QA sampling: Reviewer identity, reviewer time, accept and reject rates, and an append-only record of every agent action. This is what turns human-in-the-loop from a claim into evidence.
• Build: anything touching production and scheduling logic: This is where the return actually sits — which slot gets offered to whom, in what order, with what backfill priority. It is also the most practice-specific logic in the building.

Frenchy Digital is a senior-led, Black-owned agency in Los Angeles. We build administrative automation for regulated practices — dental practices and DSOs included — and we start from measurement rather than from a projection.

• Baseline first, always: We will not scope a workflow before we have seen eight weeks of your own call records, appointment statuses and production data. If the baseline does not justify the build, we say so and the engagement stops at discovery.
• Integration feasibility in writing: Before any build phase, a written finding on your specific practice-management system: which operations are available, which are read-only, what the licensing costs per location, and what the approval timeline looks like.
• Human checkpoints designed in: Every workflow ships with a named reviewer, a recorded review time, and tracked accept and reject rates. Human-in-the-loop is a control we build, not a sentence we add to a proposal.
• Compliance controls in code: Message caps, opt-out handling, oral AI disclosure at call open where a state requires it, clinical-question routing, and an append-only audit log. Enforced in the system, not in a training deck.
• Senior engineers only: $150–$225 per hour, fixed-price phases, and a written scope within 5 business days of the discovery call.
• You own everything: Source code, prompts, integration configuration, documentation and audit schema transfer to your business at delivery. Full IP transfer, no lock-in, 30-day post-launch warranty.

If you want a second opinion on a dental AI proposal already on your desk — the citations, the clearance numbers, the integration claims — book a call and bring the deck. Verifying those three things is an hour's work and it is worth doing before you sign. Book at calendly.com/frenchydigital/discovery-call or call +1 (424) 272-5601.