Top 10 AI Agents forDental Practices in 2026

The “$200 per no-show” figure that anchors almost every dental AI sales conversation was never a dental number. It is a physician per-hour slot figure, invented without methodology by a hospital scheduling-software vendor in 2016, and silently relabelled as dental production loss years later. Everything a practice owner is told about the economics of missed appointments rests on it, and it does not hold weight. There is no methodology behind it, and we do not print it.

The earliest traceable appearance is a Healthcare Finance News report of 17 November 2016, describing a vendor report. Its wording is precise, and worth reading closely, because the precision is exactly what later disappeared.

Missed and open appointments cost the healthcare industry $150 billion annually, according to a new vendor report that recommends automating the process. Providers have no-show rates between five and 30 percent nationwide. Each 60-minute open or no-show slot typically costs physicians $200, according to the SCI Solutions report.

— Healthcare Finance News, 17 November 2016, reporting the SCI Solutions report

Both figures are vendor-published, and we refuse them. Read the qualifier: costs physicians, per 60-minute slot. It is not a dental figure, it is not a per-appointment figure, and SCI Solutions was a hospital ordering, referral-management and network-scheduling platform — a company selling software to reduce missed appointments. Dental hygiene and restorative production per hour bear no defined relationship to a physician's hour.

The bylined origin: a vendor's CMO, on the vendor's own blog

The figure has an author. Jamie Gier, Chief Marketing Officer of SCI Solutions, published “No Joke. Missed Appointments Cost the U.S. Healthcare System $150B” on SCI's own blog on 3 January 2017. The first Internet Archive capture is dated 4 January 2017 — one day later — so the date is genuine rather than backdated. The domain is now dead. The same author restated the claim in Health Management Technology on 26 April 2017, and SCI Solutions' own Director of Marketing and Communications promoted it again on 25 May 2017.

So the sequence is: a number authored by the Chief Marketing Officer of a company that sells software to reduce missed appointments, published on that company's blog, placed in a trade magazine, then quoted back by that company's marketing director. No sample. No N. No time period. No geography. No definition of “no-show.” No denominator. No data source.

The attribution mutates into a different industry entirely

What happened next is a textbook case of citogenesis, and it is documented. By 25 March 2019, Becker's ASC Review was crediting the $150 billion figure not to SCI Solutions at all, but to Medbridge Transport, a company that provides transportation services to ASCs in Houston , citing a Medbridge whitepaper. We could not locate that whitepaper, which leaves the figure untraceable at its root. A second vendor, in a different industry, had become the apparent source. The same Becker's piece carried a companion figure: on average, no-shows cost a single-physician medical practice $150,000 annually.

The crossing point into dentistry is identifiable too. On 25 October 2022, Solutionreach content published on Patterson Dental's “Off the Cusp” blog asserted that one study foundno-shows and late cancellations can cost a practice $150,000 to $300,000 annually per provider — with no study named and no link. Solutionreach sells appointment-reminder software. Patterson Dental, which hosted the post, sells dental practice-management software and supplies. The physician figure had become a dental figure, and the doubling to $300,000 arrived with no new evidence at all.

And the dental-specific variant has no ancestry whatsoever

“$47,000 a year lost to no-shows” is the figure most often quoted at dental practice owners today, and we could not locate a single source behind it. The earliest instance found anywhere is a blog post dated 23 March 2026on the site of a dental practice coaching and seminars business, marketing those services. No source, no study, no dataset, no sample, no period, no definition. Three targeted searches surfaced only that one page — which is itself unusual, because most bad statistics leave a propagation trail. This one reads as a one-off invention rather than a propagated statistic.

We are careful about how we state that. We do not claim to have traced it. The provenance trace for the dental no-show cost family did not complete, and the absence of a located origin is not proof of anything beyond “no source is locatable.” That distinction is the difference between an evidence-based refusal and a rhetorical one. The same coaching page attributes a claim about no-show rates and profitability to an “ADA 2024 Health Policy Institute report”; ADA HPI's own published research index lists State of the Dental Economy, Dental Care Market, Dental Practice Research, Dentist Workforce, Dental Education and Coverage/Access/Outcomes — and no no-show or missed-appointment cost research. That is not an exhaustive crawl, so the accurate statement is that it was not found on ADA's own index, not that it does not exist.

There is one finding here that is safe to print flatly, because it was tested directly: a full-text search of the 2026 ADA Survey of Dental Practice workbook for “no-show,” “broken appointment,” “missed appointment,” “case acceptance,” “denial,” “denied,” “claim” and “treatment plan” returns zero hits on every term. The profession's flagship economic survey does not measure any of this. Any according-to-the-ADA citation for a no-show cost, a case-acceptance benchmark or a denial rate is citing a document that does not exist.

Claims identified and refused during this review, checked 23 August 2026. Each is named rather than silently omitted, so a reader can recognise it in a sales deck.

What honest missed-appointment evidence actually looks like

Real, defined, published figures do exist — they are just smaller, narrower and less useful for selling software. Discepolo and colleagues published a retrospective two-year cohort in JDR Clinical & Translational Research in 2023 covering 30,095 visits at an urban paediatric hospital dental clinic, finding 30.9% missed care opportunities, explicitly defined as appointments not attended or cancelled and not rescheduled prior to the initial scheduled visit. A 2025 study of 7,379 visits in an academic setting found 14.3% overall. Both are safety-net or academic settings. Neither is a private-practice benchmark, and no such published benchmark appears to exist.

The most operationally useful finding in that literature is not a cost at all. Discepolo found that wait times over 21 days carried an odds ratio of 4.07for a missed visit. The lever the evidence supports is shortening time-to-appointment, not sending more reminders — and that is a materially different product decision. We build the worked example later in this article on that mechanism, because it is the one with published support behind it.

One more boundary is worth stating now. ADA reports median gross billings per visit for solo general practitioners at $480.75 — and “gross billings” means fees charged, not collected. That is an upper bound on the value of one missed visit, before insurance write-offs. It is precisely why a $200-per-no-show or $47,000-per-year figure needs a denominator nobody publishes.

If you want the wider framing before the vendor detail, our cross-industry review of AI agents in 2026 applies the same refusal-to-score-marketing method across ten verticals, and the same pattern holds everywhere: the read surface is wide, the write surface is narrow, and the numbers that sell the software are the ones nobody has audited. The equivalent ranking for medical practices is worth reading alongside this one, because medicine has a federal certified-API floor and dentistry has none — which turns out to change almost everything.

We scored six attributes, every one of which you can re-check at a public URL, and we refused to score anything a vendor publishes about its own performance. That refusal is not modesty. In this market, every accuracy, case-acceptance, deflection and ROI figure in circulation was published by the company selling the software, usually without a methodology, a denominator or an independent reviewer.

The independent benchmark that does exist — and what it found

No independent benchmark of the commercial products in this ranking exists — nobody has tested Pearl against Overjet against Videa against Denti.AI. That is a precise claim, and it is not the same as saying no independent evidence exists. Dentistry is unusual among the verticals we have reviewed, because here the independent literature evaluates platforms, not only underlying models. It is worse than the marketing and better than nothing, and every practice owner considering imaging AI should know the number.

Abbott, Saikia and Anthonappa of the UWA Dental School published a systematic review and meta-analysis in the Journal of Evidence-Based Dental Practice in 2025 (25(1):102077) , screening 2,538 studies across eight databases from January 2000 to March 2024 and including 45, with QUADAS-2 and CLAIM quality assessment. Their finding is the single most valuable number available to a dental AI buyer.

A total of 21 different AI platforms were reported. The accuracy ranged from 41.5% to 98.6% across reported AI platforms.

— Abbott LP, Saikia A, Anthonappa RP, J Evid Based Dent Pract 2025;25(1):102077

Two further independent findings belong beside it, and omitting either would be dishonest. A 2026 meta-analysis in the same journal pooled sensitivity at 0.84 and specificity at 0.93 for proximal caries detection — and stated verbatim that most studies (22 of 28) showed high risk of bias and certainty of evidence was low, graded with GRADE. Quoting 0.84 and 0.93 without that caveat is the same sin as quoting a vendor. Across three 2025–26 meta-analyses the pooled range is sensitivity 76–84% and specificity 91–94%; quote the range across reviews rather than a single number.

And a 2026 systematic review of bias, fairness and equity in dental imaging AI in the International Journal of Medical Informatics reported that 0 of 10 studies conducted subgroup analysis, 2 of 10 provided partial demographic reporting, and 8 of 10 provided no meaningful demographic reporting , concluding that AI systems used in dental imaging show strong technical capability but lack adequate evaluation of bias, fairness and equity. That is the honest answer to “will it work on my patient population?” — the published literature does not say, and no vendor's marketing addresses it.

The ADA's own standards body says the comparison basis does not exist

This is not our editorial position. It is the profession's. ADA Technical Report No. 1109:2025, Dentistry — Evaluation of Dental Image Analysis Systems Using Augmented/Artificial Intelligence, calls for an independent validation dataset based on known diagnoses from validated sources and kept by a third party that is not an AI manufacturer, so that all users, developers and approval agencies can compare each proposed AI algorithm for accuracy and specificity. The ADA is stating, in a 2025 technical report, that no manufacturer-neutral basis for comparing dental AI accuracy currently exists.

Two precision notes, because this document is routinely misidentified. The commonly cited “ADA Technical Report No. 1106 on AI in dentistry” is a 2022 White Paper, not a technical report. And 1109 is informative, not normative — the normative document is ANSI/ADA Standard No. 1110-1:2025, Validation Dataset Guidance for Image Analysis Systems Using Artificial Intelligence, Part 1, which ADA describes as the first US standard on AI in dentistry approved by ANSI.

The independent economics point the same direction as the profession's scepticism. A 2026 paper in Caries Research found that AI can promote unnecessary interventions for early-stage caries lesions, and that greater sensitivity did not lead to better economic outcomes. And Moezzi and Hofmann in Dentistry Journal (2026;14(6):341) published an independently funded peer-reviewed analysis of exactly these vendors' marketing claims. That aligns with what dentists themselves report: an ADA Health Policy Institute panel poll in Q2 2026 found 43.3% currently use AI for at least one task but 82.6% do not use and do not plan to use it for treatment recommendations, with sceptics citing concerns about misdiagnoses and overtreatment. Peer review and practitioner sentiment agree, and both disagree with the marketing.

What happened the one time a regulator audited a vendor's AI metrics

There is a stronger reason than professional caution to hold vendor performance numbers at arm's length. In In the Matter of Presto Automation Inc., Securities Act Release No. 11352, Exchange Act Release No. 102177, Admin. Proc. File No. 3-22413, entered 14 January 2025, the SEC addressed a drive-thru voice AI product the company had told investors delivered over 94% accuracy even in noisy environments and 95% to 99% automated order completion.

The product “lacked the capability to take orders on their own and required substantial human involvement,” with “human order takers located abroad (primarily in the Philippines and India), who processed the vast majority of drive-thru orders.”

— SEC, In the Matter of Presto Automation Inc., Release No. 33-11352 (14 January 2025), ¶3

Four things must be said precisely, because this order is frequently mis-summarised. Presto consented without admitting or denying the findings. The remedy was a cease-and-desist order with no civil penalty — it was not a fine. The findings are against Presto only; the order's “Supplier A” is Hi Auto, against whom the SEC made no findings whatsoever. And Presto Automation Inc. is a different entity from Presto Phoenix Inc., which acquired assets in December 2024.

The two findings that did the work map directly onto dentistry: claiming a third party's model as proprietary, and hiding humans behind an “automated” product. A dental AI vendor reselling a foundation model as its own, or a “fully automated” charting or phone product with offshore reviewers behind it, sits in exactly that shape. The order also confirms that negligence sufficesunder Securities Act section 17(a)(2) — a vendor need not have intended to mislead.

Here is the entire ranking in one table, with a citation or the literal words “not publicly disclosed” in every cell. Nothing in it is our opinion of clinical or operational output quality, because output quality is exactly what nobody has measured head-to-head. The ordering reflects how much due diligence each vendor makes possible before a sales call, which is the only honest proxy available.

All cells read off the vendor's own pages, the FDA 510(k) database and public developer portals on 23 August 2026. Blank is never assumed — where a vendor does not publish something, the cell says so. Independent evaluation is omitted as a column because the answer is the same for all ten: none located.

One disclosure belongs with the table rather than buried. Adit, an all-in-one dental platform that launched an AI Front Desk Agent in February 2026, appears on Henry Schein One's published “Unauthorized Vendors” list. That is a commercial party's allegation about a competitor — Henry Schein One's own wording is that listed vendors may be utilizing an unauthorized connection— and it is not a finding by any court or regulator. We address that list in full in the binding-constraint section, because it is one of the most useful public documents in dentistry and almost no practice owner has read it.

Four of the ten are radiograph analysis products, and all four are copilots rather than agents — a distinction their marketing tends to blur. None of them is authorised to diagnose, all of them are cleared as an aid to a licensed clinician, and the gap between the marketed product and the cleared indication is the single most useful thing to check before a demo.

One more clearance is worth naming because it shows the gap at its widest. Diagnocat's K252934, cleared 15 January 2026, is indicated to aid in the detection of periapical radiolucency on permanent teeth captured on maxillofacial CBCT images, using scans that were previously acquired for clinically justified purposes independent of Diagnocat, read by professionals with at least two years of clinical experience reading CBCT scans, as a second read, in patients 22 years of age or older. One finding, one modality, only on scans already taken for other reasons. Note too that the applicant is DGNCT, LLCof Miami — there is no 510(k) filed by any entity called Diagnocat Ltd or Diagnocat Inc, and both applicant searches return zero.

The six front-office products are where the word “agent” is doing the most work, and where the compliance gaps are widest. Several of these dial or text patients, which puts them squarely inside the TCPA; several claim to write into the practice management system, which is the contested capability; and most publish nothing at all about how they handle protected health information.

A ranking article is a vendor roster, and a stale roster is the fastest way to tell that nobody re-checked. Fourteen candidates were checked for 2026 corporate status. Here is what we removed, and the reason for each — because naming an exclusion is more useful to a practice owner than three more table rows.

• Toothpik — does not exist: No US dental-practice AI agent vendor of this name could be confirmed. The only near match is Toothpic, a UK-based teledentistry and photo-triage app: a different product, a different market, and not verified as a 2026 practice-facing agent. It appears in other rankings. It does not go in this table.
• DentalBee — acquired, and the name is gone: Overjet announced the acquisition on 2 December 2025, terms undisclosed. The voice-enabled clinical documentation and hands-free perio charting now ship as Overjet Voice. Ranking DentalBee as an independent vendor in 2026 would be listing a product that no longer exists under that name.
• TrueLark — acquired, completed May 2025: Weave announced the acquisition on 5 May 2025 and completed it on 19 May 2025. TrueLark is the AI front-desk layer inside Weave today. It is covered in Weave's entry, not as a separate row.
• VideaHealth the brand — retired; the entity is not: The rename to Videa was announced 21 April 2026. Only the brand changed: the legal entity remains VideaHealth, Inc., confirmed in the company's own privacy policy. The product is ranked; the dead brand name is not used; and 'Videa, Inc.' is a company that does not exist.
• Vyne Dental — present, but a litigant in the case that is this article's spine: National Electronic Attachment, Inc. d/b/a Vyne Dental is active and owned by The Jordan Company. It is a party to the federal litigation over Dentrix write access, so ranking it alongside the other side would be indefensible. One disambiguation: there is an unrelated London fintech also called Vyne, acquired by Tarabut in 2024. Several aggregators conflate them. That is a different company.
• Curve Dental — no published developer route at all: CD Newco, LLC is active and has launched AI products through 2025-26 including Ask CurveAI and a voice-driven Curve Care+. But /api, /developers, /integrations and /partners all 404; developer, api and docs subdomains fail DNS; and a sweep of 776 sitemap URLs returns zero developer paths. Its only public statement about an API refers to rigorous standards for quality, security, reliability, interoperability and customer experience — standards that are nowhere published, with no application route.
• Adit — ranked with disclosure rather than silently dropped: Adit is active in Houston and launched an AI Front Desk Agent in February 2026. It is mostly classic automation and communications software with an agent module added, and its own pricing page shows no plan prices. It also appears on Henry Schein One's published Unauthorized Vendors list — an allegation by a commercial party, not a finding by any court or regulator, and it is in company there with NexHealth, Sikka and Kolla. A third-party directory lists a monthly figure for Adit; that is not the vendor's published price and we do not print it.

There is also a set of vendors that most published roundups miss entirely, and that a buyer should know exist, because they are confirmed by the strongest possible evidence: a current FDA clearance filed under a named legal entity. Velmeni, Inc. (K252953, cleared 11 May 2026, with an authorised predetermined change control plan), Better Diagnostics AI Corp. (K260851, 15 July 2026), Dentsply Sirona for DS Core (K260785, 23 April 2026), Nobel Biocare for DTX Studio (K252086, 17 November 2025), and DGNCT, LLCfor Diagnocat (K252934, 15 January 2026). A clearance proves the device exists and the entity filed. It does not prove the company is trading, staffed or selling — confirm commercial status before you shortlist any of them.

Three names that appear in dental AI roundups have no 510(k) clearances at all: Retrace, Adent.io and Dentem, each returning zero applicant matches. And Orca Dental AI holds the CephX clearances, neither of which covers caries, periapical radiolucency or bone level — they are orthodontic and cephalometric. Counting a denoiser, an image-quality checker or a cephalometric tool toward a “caries AI clearance” total inflates the number.

Dentistry has no federal certified-API mandate at all, so every write an agent performs into a dental practice management system is vendor-discretionary, commercially priced, and — as of six weeks before this article was published — actively litigated. This is the spine of the market, and almost nobody explains it to a practice owner at the point of sale.

Start with what medicine has and dentistry does not. Under the ONC certification criterion 45 CFR 170.315(g)(10), a certified health IT module must respond to requests and search requests for patient data. Its entire data surface is retrieval. There is no create, update or write requirement anywhere in it. But it carries a documentation obligation that is devastating by contrast: the documentation must be available via a publicly accessible hyperlink without any preconditions or additional steps. Certified health IT must publish its API docs for free, with no gate. Dentrix charges $5,000 to see the write list and refuses to publish its schema at all.

And nothing is moving toward dentistry. A keyword sweep of ASTP/ONC's December 2025 deregulatory proposed rule — 485,058 characters — found the word “dental” exactly once, in a historical recital about a 2013–14 data-set revision, with zero occurrences of “dentist” or “oral health.” A January 2026 imaging-interoperability RFI contains zero occurrences of all three. The direction of travel is the opposite of extension: the proposed rule would remove criteria and narrow the programme onto APIs.

The clearest proof that federal certification is not what produces openness here: Open Dental deliberately exited the ONC certification programme. Its own page states that it was EHR certified from August 2011 through December 2018 and has decided not to develop certified 2015-edition software. The most open API in dentistry belongs to a vendor that affirmatively left the federal scheme. Its openness is entirely voluntary.

What each system actually charges, and what it actually permits

The correction that matters most here: dental PMS vendors do not “publish nothing.” Henry Schein One, Planet DDS, Open Dental and CareStack all publish extensive documentation including write operations, and four publish exact fees. The real finding is sharper. Write access is priced, contractually gated, categorically denied to competitors, and there is no regulator to appeal to.

Every cell verified against the vendor's own public pages on 23 August 2026. The spread across this table — free-and-documented in one system, $10,000 to enter plus an unpublished royalty in another, and no published route at all in two more — is the binding constraint made concrete.

Look at the Open Dental row against the Dentrix rows. Open Dental publishes full CRUD across roughly 130 resources, and its pricing is quotable verbatim: Free — Read All, then $15, $30 or $35 per location per month for progressively wider write access, with the note that pricing may change over time but that the company tries to keep increases below inflation. Reading is free. Writing costs $15 to $35 per location per month. Open Dental has priced exactly the capability that Henry Schein One charges an entry fee plus an undisclosed royalty for, and has gone to federal court over. Same capability, three regimes. An operator choosing a practice management system is choosing an agent-integration regime, and almost nobody tells them that.

Three published gates on Dentrix Ascend deserve individual attention, because the third is the one operators never anticipate. A signed agreement after application review. Production access requiring a minimum of 1,000 API calls across intended production endpoints plus verification by the API team. And per-practice vendor enablement, where vendors appear in the list by default as off until someone enables them. Even an approved vendor is switched off at your practice until you turn it on.

The gatekeeper's own published terms

Henry Schein One's public developer FAQ is one of the most useful documents in dentistry, and it is not behind a login. It publishes the fee structure — a one-time registration and set-up fee of $5,000 for read and $5,000 for write, plus a monthly royalty based on API categories selected, with no start-up pricing and no access prior to joining, and a singular pricing model that applies to all. It states that there is no certification process available to API members, which contradicts an archived 2024 program page advertising a certification tier. It states that the licence agreement prohibits transferring, selling, distributing, disclosing, sublicensing or lending the API key — which forecloses the aggregator and middleware business model by contract. And it states, plainly:

What are “protected” categories? Certain software types (e.g., patient financing, credit card processing, insurance claim processing) or non-commercial API accounts aren't allowed in the program or marketplace.

— Henry Schein One, API Exchange customer FAQs, read 2026-08-23

Henry Schein One sells products in all three of those categories. We state the two facts side by side and draw no legal conclusion, because no regulator or court has found one and none has jurisdiction to. But it is the whole argument in a single page: an “integration” in dental is a commercial permission that the PMS vendor grants, prices, withholds from competitors and can withdraw — not a regulated interface. Its own page tells customers what happens if a vendor is not approved: the processes will no longer be automated and may need to be done manually.

The list every practice owner should read and almost none has

Henry Schein One publishes a page headed “Unauthorized Vendors” naming 57 companies that, in its own words, may be utilizing an unauthorized connection to Henry Schein One, adding that this is not a complete list. The names include NexHealth, Sikka, Kolla, Adit, DentalHQ, Swell, Kleer, Rhinogram, Simplifeye, Solventum — and Vyne.

Be precise and fair about what that is. It is an allegation by a commercial party about its own competitorsand about companies integrating with its product. It is not a finding by any court or regulator, and nobody on that list has been found to have done anything unlawful. Exactly one of the 57 has been adjudicated at all — Vyne, at the preliminary-injunction stage, where the court granted both sides relief in part and then suspended the operative paragraph.

The litigation, with the status exactly right

National Electronic Attachment, Inc. d/b/a Vyne Dental v. Henry Schein One, LLC, No. 1:25-cv-03246-MJM (D. Md.), before Judge Matthew J. Maddox. On 31 July 2026 the court issued a 94-page Memorandum Opinion and an Order granting both sides' cross-motions for a preliminary injunction in part and denying them in part. Paragraph 8(a) barred Vyne from marketing, selling, distributing or supporting software capable of write access to Dentrix databases. Paragraph 8(b) barred Henry Schein One from disabling Vyne's printer driver or denying it access to Dentrix databases on practice computers.

ORDERED that paragraph 8(a) of the Court's Order of July 31, 2026, Dkt. 190, is SUSPENDED pending this Court's consideration of Plaintiff's Motion for a Stay Pending Appeal and for Clarification.

— Order, ECF 197, National Electronic Attachment, Inc. v. Henry Schein One, LLC, No. 1:25-cv-03246-MJM (D. Md.), signed 11 August 2026

Get every one of these right, because content farms are getting them wrong. Only paragraph 8(a) — the restraint on Vyne — is suspended. Paragraph 8(b), which restrains Henry Schein One, was not suspended. As of 23 August 2026, the PMS vendor is restrained and the third party is not. The stay motion was fully briefed and undecided; an appeal is docketed at the Fourth Circuit as No. 26-2085, opened 11 August 2026; a Rule 65(c) bond hearing was set for 16 September 2026. Suspended pending consideration of a stay motion is not stayed, not vacated and not dissolved.This is a preliminary injunction, not a final judgment. Nobody has won. Henry Schein One's own customer-facing page still describes the injunction as operative, so do not use that page for the court's status — use it only for the company's published programme terms.

Two further precision points. The court made no finding that Henry Schein One committed information blocking— it cited the 21st Century Cures Act as a public-interest consideration and nothing more. And there is no standalone information-blocking cause of actionin the case, because the Cures Act creates no private right of action; Vyne pleads it as context. Vyne's counts are the Computer Fraud and Abuse Act, Lanham Act false advertising, tortious interference, unfair competition and defamation. Henry Schein One's motion to dismiss Vyne's amended complaint was denied, and four of its own counterclaim causes were dismissed without prejudice, meaning they can be repleaded.

The only forum where dental PMS write access has actually been adjudicated is private commercial litigation under the CFAA and the Lanham Act — not antitrust, and not information blocking. The court did lean on Real Time Medical Systems v. PointClickCare Technologies, affirmed at 131 F.4th 205 (4th Cir. 2025), for the proposition that the public benefits from fostering access to medical records consistent with the Cures Act. That is the strongest available citation that courts will protect third-party data access in health IT — and it is a skilled-nursing case, not a dental one.

One date trap worth naming: several sites date this injunction 12 May 2026. That is wrong. The Memorandum Opinion and Order are both signed 31 July 2026. The May date belongs to an industry analyst's blog post about an earlier letter order signalling the injunction's shape. Anything you read that cites the May date was not checked.

For the architectural consequences of all this — where PHI actually sits, what a business associate agreement has to cover, and how to design an agent so that a revoked integration is an inconvenience rather than an outage — our guide to HIPAA-compliant AI agent architecture works through the design patterns in detail.

Filed, proposed, entered, in force, stayed, enjoined, vacated and on appeal are different states, and a practice owner acts differently depending on which one applies. Here is what actually binds a dental practice deploying an AI agent in August 2026.

• HIPAA — the 2003 Security Rule governs, and the NPRM is still only proposed: The HIPAA Security Rule NPRM published 6 January 2025 at 90 FR 898 under RIN 0945-AA22 is a proposed rule. Querying the Federal Register by RIN returns exactly one document and it is that proposal. There is no final rule as of 23 August 2026. Do not write about the NPRM's requirements as if they are in force. What is in force today is what has always been in force: a dental practice that bills electronically is a covered entity, an AI vendor that creates, receives, maintains or transmits PHI on its behalf is a business associate, and a BAA is required under 45 CFR 164.502(e), 164.308(b) and 164.314(a). That is why the BAA column in the comparison table is a real scoring criterion and not a nicety.
• OCR has already settled with a dental software vendor — March 2026: HHS Office for Civil Rights announced a Resolution Agreement with MMG Fusion, LLC on 5 March 2026. OCR's own recital describes MMG as a software company that focuses on helping oral healthcare professionals to market, manage and grow their practices, and states that MMG is a business associate as defined at 45 CFR 160.103. The covered conduct was impermissible use and disclosure, failure to conduct an accurate and thorough risk analysis, and failure to notify affected covered entities of a breach; PHI had been accessed in December 2020 and published on the dark web. The resolution amount was $10,000 plus a corrective action plan. Frame it precisely: the agreement is expressly not an admission of liability. It is a settlement, not a finding of violation, and a resolution amount under a voluntary settlement is not a fine.
• California AB 489 — in force since 1 January 2026, and it reaches dentistry squarely: Business and Professions Code sections 4999.8 to 4999.9 prohibit the use of any term, letter or phrase in the advertising or functionality of an AI system that indicates or implies that care, advice, reports or assessments are being provided by a natural person holding the appropriate licence. Health care profession is defined as any profession licensed or regulated under BPC Division 2, which contains the Dental Practice Act. Violations fall to the appropriate licensing board, meaning the Dental Board of California can enforce it, and each use of a prohibited term is a separate violation. Several agents in this market carry human first names and coworker framing; how yours introduces itself on a live call is a configuration choice you own.
• California AB 3030 — probably does NOT reach an ordinary private dental office: A great deal of published content asserts otherwise, so this correction is worth having. Health and Safety Code section 1339.75, effective 1 January 2025, requires a GenAI disclaimer for communications pertaining to patient clinical information — but its covered entities are a health facility, clinic, physician's office or office of a group practice, and the statute defines physician's office as an office of a physician in solo practice and office of a group practice as two or more physicians. HSC section 1206(a) exempts from clinic licensure any place operated as a clinic or office by licensed health care practitioners for the practice of their profession, which is where an ordinary private dental office sits. The enforcement subdivision routes to health facilities, licensed clinics and physicians via the Medical and Osteopathic Boards — there is no dental board pathway at all. A dental practice operating as a licensed clinic is a different matter.
• Texas HB 149 (TRAIGA) — in force 1 January 2026, and it creates a real duty: Read the structure carefully, because section 552.051(b) binds only governmental agencies. The provision that reaches a private dentist is section 552.051(f): if an AI system is used in relation to health care service or treatment, the provider shall provide the disclosure to the recipient not later than the date the service or treatment is first provided, except in an emergency. Health care services covers diagnosis, prevention or treatment of a human disease or impairment provided by a licensed individual. The disclosure must be clear and conspicuous, in plain language, with no dark patterns, and may be delivered via a hyperlink to a separate page. The Attorney General has exclusive enforcement authority, there is no private right of action, and there is a 60-day cure period.
• TCPA — the artificial-voice classification and written consent both stand: FCC Declaratory Ruling 24-17, adopted unanimously and released 8 February 2024, holds that AI-generated voices are artificial voices under the TCPA, and expressly refuses any carve out of technologies that purport to provide the equivalent of a live agent. Callers must obtain prior express consent, prior express written consent for marketing, identify themselves, disclose the responsible party and offer opt-out. Insurance Marketing Coalition v. FCC (11th Cir., 24 January 2025) vacated only Part III.D of the 2023 Order — the one-to-one consent provision — and the court's own footnote states that the 2012 Order is not at issue. Anyone telling you the TCPA got easier for AI callers has inverted the law. One dental-specific nuance: appointment reminders and recall are generally treated differently from marketing, but a reactivation campaign to lapsed patients — which is exactly what several of these agents sell — is much closer to marketing than to a reminder. That is a question for your counsel, not settled ground.
• State practice acts — diagnosis is reserved, and that is the real answer: The cleanest structural answer to 'can an AI diagnose?' is not AI law at all. California Business and Professions Code section 1625 defines dentistry as the diagnosis or treatment of diseases and lesions of the human teeth and associated structures, and reaches anyone who performs or offers to perform a diagnosis of any kind. Florida Statutes section 466.003(3) covers examination, diagnosis, treatment planning and care, and section 466.003(9) requires that a dentist diagnoses the condition to be treated. New York Education Law section 6601 names diagnosing. We verified those three. We did not verify Texas, Ohio, Illinois or Georgia practice-act language, and we do not assert it.
• Teledentistry — one verified federal constraint on state boards: In the Matter of Board of Dental Examiners of Alabama, FTC Docket No. C-4757, Decision and Order issued 20 December 2021 and terminating 20 December 2031. The FTC challenged the Board's interpretation of its rule to prohibit non-dentists from performing digital scans without on-site dentist supervision, alleging it unreasonably excluded emerging competition from teledentistry platforms under FTC Act section 5. Frame it correctly: this is a consent order in which the Board admitted jurisdictional facts only, with no admission that the law was violated. The governing precedent is North Carolina State Board of Dental Examiners v. FTC, 574 U.S. 494 (2015). Why it matters for AI: it is the clearest available statement that a state dental board may not use a supervision rule to block remote scan-based workflows.

This is an illustrative scenario, not a client result, and every input is a figure named and dated earlier in this article. We build it that way deliberately: an article that refuses an unsourceable industry statistic and then asserts an unverifiable client outcome has destroyed its own standing.

What this looks like in a real engagement

We can describe scope rather than outcomes here, and we will keep to that. For BH Dental Corp, a Beverly Hills practice , Frenchy Digital delivered a practice website with NexHealth scheduling integration, HIPAA-compliant patient intake forms, insurance verification information and patient portal integration, alongside service pages and local search work. We publish no performance figures for that engagement in this article, because this is an article about refusing unverifiable numbers and it would be absurd to make an exception for our own.

But the scope itself illustrates the point precisely. The scheduling integration runs through NexHealth— the middleware layer that agent vendors across this market depend on, that documents its write endpoints publicly, that offers free self-serve developer signup, and that appears on Henry Schein One's published unauthorized-vendor list. That is the actual architecture of dental scheduling in 2026: a practice, a middleware layer, and a practice management vendor whose position on that layer is published and worth reading before you build on it.

In dental deployments, the failure is almost never the model. It is the write path, the permission, or a patient population the cleared indication never covered. Here is what actually goes wrong, how you detect it, and what your rollback looks like.

• The vendor is switched off at your practice and nobody told you: On Dentrix Ascend, vendors appear in the enablement list as off by default until someone at the practice turns them on. An approved vendor with a signed agreement and a paid royalty still does nothing until that toggle flips. Detection: a booking or write count that reads zero on day one. Rollback: it is a toggle, but find out who at your practice owns it before go-live, not after.
• A destructive write does exactly what the documentation says it does: Kolla's public documentation is admirably candid: replacing insurance benefits under a coverage replaces all of them, benefits not included in the request are deleted, and this is a destructive operation — and for some systems, Eaglesoft among them, benefits are shared at the plan level and the operation may affect all patients on the same plan. Detection: a benefits diff before and after any bulk operation. Rollback: a database backup taken before the run, because the API will not give you one.
• Payments post FIFO against the wrong balance: Most practice management systems apply a payment first-in-first-out against outstanding balances. An agent posting payments at volume will do exactly that, correctly and unhelpfully, against ledger lines a human would have allocated differently. Detection: an aging report that improves in total while individual accounts drift. Rollback: manual reallocation, which is slow.
• The patient is outside the cleared age range and nobody checked: This one is easy to miss and hard to defend. Videa Caries Assist is cleared for bitewings only, patients 22 and older. Both Denti.AI devices are adults 22 and older without remaining primary dentition. Overjet Caries Assist-Pediatric is cleared for 4 to 11 only. Videa Dental Assist covers 3 and older. Overjet Caries Assist moved from 18+ to 12+ between versions. 'These tools work on kids' is false for roughly half the roster. Detection: cross-reference your patient age distribution against the specific K-number your vendor is operating under — and ask which version.
• The reading protocol is wrong, so the workflow is wrong: Second reader and concurrent read are not interchangeable, and FDA requires the protocol to be labelled. Pearl's original Second Opinion is a second reader: the dentist reads unaided first, then the AI output is revealed. All Overjet CADe devices, Videa Caries Assist, Velmeni, Better Diagnostics, Nobel Biocare and DS Core Detect are concurrent read. Building a chairside workflow around the wrong protocol undermines the clearance you paid for. Detection: read the indications for use, not the sales deck.
• Your agent vendor's name appears on your PMS vendor's list: Lists change, and neither party will email you when they do. Detection: a calendar reminder to re-read both the authorized-partner page and the unauthorized-vendor page quarterly. Rollback: Henry Schein One's own customer page states the consequence plainly — the processes will no longer be automated and may need to be done manually. Keep the manual path warm and staffed enough to absorb that for a week.
• Prompt injection — unsolved, and your phone and chat agents are exposed: Any agent that reads untrusted input — a patient's free-text message, a web-chat entry, a scanned document, an inbound email — can be steered by that input. There is no reliable defence today. The correct posture is blast-radius reduction, never 'solved': scope the agent's credentials to the minimum resources it needs, keep every write behind a human commit step where the write is consequential, log every action with the input that triggered it, and rate-limit anything that touches money or the clinical record. Detection: an action log you actually read, and alerting on any write outside the expected distribution.

The pattern across all seven is the same. These are not model failures, and no amount of accuracy would prevent any of them. They are permission failures, protocol failures and population failures — which is precisely why this ranking scores permissions, protocols and populations rather than accuracy.

The boundary in dentistry is not a matter of taste. It is set by three verified constraints that all point the same way. First, every cleared dental imaging AI is an aid, and 21 CFR 892.2070 states that the device is not intended to replace the review by a qualified radiologist, and is not intended to be used for triage, or to recommend diagnosis. Second, diagnosis is the practice of dentistry and is reserved to licensees under state practice acts — an AI cannot lawfully be the diagnostician. Third, California AB 489 makes it a separate violation each time an AI system implies it is a licensed professional.

Two marketing patterns in this roster sit close to that last line: human first names — Annie, Aimee — and “digital coworker” framing. We state the rule and accuse no vendor of violating it, because the compliance question is how the agent introduces itself on a live call — and that is a configuration choice the practice makes and owns, not a property of the product. Write the introduction script yourself, put it in the implementation checklist, and listen to ten recorded calls in the first week.

The shape that survives scrutiny is simple to state. An agent may prepare, gather, draft, surface and propose. A licensed person commits anything that enters the clinical record, changes a clinical decision, or reaches a patient under the practice's name. That is not caution for its own sake. It is the only architecture that survives a prompt injection, a board complaint and a bad week — and it happens to be the one that actually ships.

For most single-location practices, the right first move is not a custom build at all. It is to find out what your practice management vendor actually permits, price that access, buy a product that publishes a BAA, and measure it against your own baseline for a quarter. If after that a workflow still needs something no product covers, these are our bands.

Senior-led delivery is $150–$225 per hour and retainers run $2,500–$9,500 per month. Every engagement carries a 30-day post-launch warranty, full source-code and IP ownership transfers to you, and you receive a fixed-price phased proposal within 5 business days of the discovery call. Frenchy Digital is a senior-led Black-owned Los Angeles agency; you can reach us at +1 (424) 272-5601 or book directly at calendly.com/frenchydigital/discovery-call .

These bands are ours and are not an industry benchmark. We publish them because a practice owner comparing a $28,000 single-workflow build against a $199-a-month subscription and a $420-a-year API fee deserves all three numbers on the same page. Most of the time the subscription wins. The build wins when the workflow is genuinely specific to your practice, when the integration your product needs does not exist, or when you have decided that owning the code matters more than owning the fastest path.

Every one of these is checkable in under five minutes, and every one is drawn from something we actually found during this review.

1. 1.Infrastructure certifications recited as the vendor's own: If a security page says ISO 27001 and SOC 2 Type 2 in a sentence whose subject is the cloud provider, those are the cloud provider's certifications, not the vendor's. Ask what the vendor's own legal entity holds, and ask for the report. There is a live example in this roster and it takes thirty seconds to verify.
2. 2."HIPAA-ready" and "designed to support HIPAA compliance": Those are support claims, not compliance claims, and the distinction is deliberate. Ask the one question that resolves it: will you sign a BAA, and may I read it before we contract? Only two of eleven vendors here publish the document at all.
3. 3.A BAA you are not allowed to read until after you sign: Read the AI clause specifically. The one public BAA in this roster permits PHI use within internal and external Artificial Intelligence systems and models. Another is silent on AI and requires return-or-destruction with no copies retained. Ask every vendor for that exact sentence in their own agreement. If they will not show you the document before contracting, that is the answer.
4. 4.An integration list with no API documentation behind it: A page of logos is not documentation. Ask which named programme the integration runs through — Dentrix API Exchange, an Ascend tier, Open Dental's API, a middleware layer — and get it in writing. Then ask whether it reads or writes. Read-only is a materially different product, and most of the labour saving in scheduling and revenue cycle is in the write.
5. 5.A headline accuracy or production number with no methodology: Six of eleven vendors reviewed publish one. Ask for the denominator, the date, and whether the study's subjects were customers. A joint study with a customer is not an independent benchmark, no matter what the press release calls it. And hold any figure against the independent floor: 41.5%, across 21 platforms, peer-reviewed.
6. 6.A marketing brand presented as FDA-cleared when the K-number covers something else: Ask for the K-number, then read the indications for use yourself in the FDA database. Two heavily marketed products in this category have no clearance under their marketed names at all. Several cleared devices are far narrower than the product sold around them — one covers a single finding, on one modality, only on scans taken for other reasons.
7. 7.An "autonomous" agent with no human-in-the-loop description anywhere on the site: The SEC's Presto order is what happens when that gap turns out to be load-bearing. Ask where the human is, what they see, and what the agent does when it is not confident. A vendor that cannot answer that has not thought about it, or does not want to.
8. 8.Pricing that cannot be obtained without a demo, on a product sold per location: Nine of eleven vendors reviewed publish no price. That is normal in this market and it is still a cost to you: it makes budgeting a negotiation and makes comparison impossible. Ask for per-location pricing in writing, ask what happens at renewal, and ask what happens to your rate if the company is acquired — a question with unusual bite right now.

A ranking that refuses vendor claims and then hides its own gaps has learned nothing. Here is what this article does not know.

• No independent benchmark of any commercial dental AI product exists: Nobody has tested these products head to head. What exists is peer-reviewed meta-analytic evaluation of dental radiograph AI as a modality — Abbott 2025 across 21 platforms, Guo 2026, Wang 2026 — and the ADA's own Technical Report 1109:2025 explicitly calling for a validation dataset kept by a third party that is not an AI manufacturer. Hold that distinction exactly: independent evidence exists; a product benchmark does not.
• The evidence base that does exist is graded low-certainty: Twenty-two of 28 studies at high risk of bias with low certainty overall (Guo 2026), and 0 of 10 dental imaging AI studies performing any demographic subgroup analysis (Suganya 2026). Nobody has shown these tools perform equally across patient populations, and no vendor's marketing addresses it.
• The $47,000 no-show origin was not traced to completion: We refuse the figure regardless — it has no locatable source and it is on our permanent refusal list. But we do not claim to have traced it, and the absence of a located origin is not proof of anything beyond 'no source is locatable.' Search access during this research was heavily degraded, with several engines serving bot challenges, so the working path was one search provider plus archive snapshots.
• Compliance postures were read off vendor pages on a single date: Absence of a SOC 2, HITRUST or BAA statement means it was not published on 23 August 2026 — not that the vendor lacks one. Several may well hold certificates behind a sales NDA. What we can report is what is published, and we report exactly that.
• Funding figures for private companies come from aggregators, not filings: Only Weave has SEC-filed financials. Round sizes and valuations elsewhere come from company announcements and data aggregators. Where a figure conflicted with a company's own scale claims, we recorded funding as not confirmed from primary record rather than printing a number.
• We could not verify which dental systems hold ONC certification: The Certified Health IT Product List REST API failed or timed out on every attempt across this research. We therefore publish no count of certified dental practice management systems. Dentrix Enterprise, aimed at community health centres, markets HL7 and FHIR services and is the plausible exception — check it directly rather than relying on us.
• Section 1557's litigation posture and state dental board AI rulemaking are both unchecked: The court orders touching 45 CFR 92.210 were never pulled, so we say nothing about the injunctions. And no state dental board's rulemaking was checked in any state, so we do not claim none has acted on AI. Both are open items, and naming them is more useful than a confident sentence we cannot support.
• Advertising and regulatory enforcement against dental AI vendors was not searched: We do not write that no scrutiny has occurred, because we did not look. Likewise, no DOJ or FTC antitrust matter on dental PMS interoperability was located, but that search was not run to exhaustion — treat it as not located, never as does not exist.
• The litigation is unresolved and moving: A preliminary injunction was entered 31 July 2026 and its central paragraph suspended 11 August 2026, with a stay motion undecided and an appeal docketed at the Fourth Circuit. Anything written about it must carry the check date, and ours is 23 August 2026. Re-check before you rely on it.

None of this argues against buying. It argues for buying the way you would buy any other clinical or operational system: find out what you are permitted to do before you find out what the model can do, measure your own baseline first, run one workflow, keep the human commit step, put the compliance answers in the contract rather than the email, and re-check the vendor's published facts at renewal. The practices that get value out of this technology are the ones that read the developer terms before they read the sales deck.