The Dental AI Paradox: Cheaper Expertise, More Powerful Platforms

In April, Henry Schein One opened the Model Context Protocol (MCP) layer of Dentrix Ascend, allowing AI agents to query practice data and perform operational work inside the practice-management system. The company described a future in which AI does not merely suggest an action—it verifies insurance, reconciles claims, manages recalls, and follows up with patients from inside the system of record.[1]

Read one way, that is the opposite of enclosure. MCP is an open protocol. Opening it invites outside developers to build against practice data rather than waiting for one vendor to ship every feature. Extensibility of that kind is a genuine argument against the fear that dental software is closing in on itself.

But extensibility is not portability. An open interface to data that still lives in one vendor's system makes the system more useful without making it easier to leave. The more third-party agents a practice builds against that interface, the more of its operating knowledge comes to depend on it. Opening a protocol and deepening a dependency are not opposites. They can be the same move.

Two months later, the European Commission opened an antitrust investigation into whether Align Technology had created a closed ecosystem between its iTero scanners and Invisalign. Align disputes the allegations, and no wrongdoing has been established.[2][3]

These developments concern different companies and different issues. Together, they reveal the central tension shaping dental AI.

AI is no longer arriving as an isolated software feature. It is becoming infrastructure.

Once a technology becomes infrastructure, the critical question changes from "How intelligent is it?" to "Who controls it?"

A small practice or laboratory can now reach capabilities that once required specialists, larger teams, or expensive infrastructure. But as AI becomes embedded in practice management, imaging, CAD/CAM, claims, manufacturing, and patient acquisition, the cost of reaching that capability may fall while the cost of switching systems rises.

That is the dental AI paradox:

The tools are becoming more democratic while the platforms are becoming more powerful.

What AI is making cheaper to reach

The first wave of dental AI is fundamentally a labor and capacity story.

The pressure is measurable. In the 2025 Dental Economics–Levin Group Annual Practice Survey, 58% of practices reported higher overhead than the prior year, with an average increase of 4.5%, and 55% named rising overhead as their single biggest challenge. Average production per doctor was statistically flat—$1,001,807 against $1,004,178 the year before. Seventy-six percent had at least one unfilled position, 49% were raising base compensation to compete for staff, and 50.5% reported that at least one payer lowered reimbursement during the year.[4]

That is the arithmetic driving interest in AI: costs rising, labor scarce and getting more expensive, production flat, reimbursement flat or falling. Practices and laboratories are under pressure to serve more patients or process more cases without adding proportional labor.

The same survey reported that 43% of responding practices were using AI clinically or administratively, and that more than 89% believed there was a shortage of dental staff available for hire. The sample leaned heavily toward independent practices—87% of respondents were owners or partners in independent practices, roughly 10% worked for DSOs or large groups—and the published findings do not disclose the total number of respondents. It should not be interpreted as a national adoption rate. But it illustrates what is driving interest: efficiency rather than novelty.[4]

One caution belongs here rather than in a footnote. The pressure on practice economics is documented. What AI does to that pressure is not. There is no published pricing, wage, or cost-per-case evidence showing that dental AI lowers a practice's or laboratory's total cost of delivering care. What the clinical literature supports is narrower and more specific: AI can make certain bounded capabilities easier to reach without hiring for them. Whether easier access translates into lower cost is an open question, and this article treats it as one.

The strongest evidence supports AI for bounded tasks—problems with relatively clear inputs, outputs, and opportunities for human review.

In a randomized trial involving 30 dentists and 50 panoramic radiographs, AI assistance improved diagnostic accuracy for periapical radiolucencies from 91.6% to 93.3%. The improvement was modest, but false-positive diagnoses fell from 4.3% to 2.0%, and junior dentists benefited most.[5]

Sensitivity did not improve—and the level is the finding. Sensitivity was 46.0% unaided and 45.8% with AI assistance. Readers missed more than half of the lesions either way. AI assistance made them more precise about what they did see. It did not make them see more.[5]

That is a narrow, real, and instructive result. The AI did not become the dentist. It helped less-experienced dentists perform a specific task more consistently, while leaving the harder problem—detection—essentially untouched. Any claim that AI is substituting for clinical expertise has to explain that sensitivity figure.

Similar work is underway in:

• Radiographic second review
• Image-quality assessment
• Aligner monitoring
• Clinical documentation
• Prescription interpretation
• Initial crown design
• Case classification
• Production scheduling
• Insurance verification
• Patient communication

A 2026 retrospective multicenter study of DentalMonitoring analyzed 3,323 assessments from 623 patients and found strong performance in identifying aligner tracking problems. The system was particularly effective at ruling out meaningful misfits. However, the study used data from the company's own data pool, and its authors called for broader independent validation.[6]

Reviews of AI-assisted crown design likewise report shorter design times and generally acceptable fit and morphology. But much of the evidence remains in vitro, and long-term clinical performance is still poorly documented.[7]

The pattern is becoming clear:

AI lowers the cost of routine execution. It does not eliminate the need for professional judgment.

Where the evidence stops

Dental AI studies frequently report impressive sensitivity, specificity, or accuracy. Those numbers matter, but only if the system performs similarly outside the environment in which it was developed.

A 2026 meta-analysis covered 27 studies and 60,857 radiographic images. The studies varied substantially in populations, imaging systems, clinical tasks, and validation methods, and many were retrospective and lacked strong external testing. Against that heterogeneity, the pooled figures—sensitivity of 0.85 and specificity of 0.94—describe an average across incommensurable settings more than a performance expectation for any particular product.[8]

One cross-national caries-risk study demonstrates why this matters. Its model performed moderately well on its internal data but fell to little better than chance when tested on a different national population. Sensitivity dropped dramatically. This was a questionnaire-based research model—not a commercial radiographic product—but it provides a useful warning about assuming that performance transfers across populations and settings.[9]

There is a second gap, less often noted. The dental AI literature measures diagnostic and design performance. It does not measure economics. Studies report accuracy, sensitivity, design time, and agreement rates; they do not report cost per case, staffing effects, subscription costs against labor saved, or return on investment. Anyone claiming that AI reduces the cost of dental care—including anyone making the accessibility argument in this article—is extrapolating beyond the published evidence.

FDA authorization is another meaningful but frequently misunderstood signal. It indicates that a product met applicable premarket requirements for its intended use. It does not prove superior patient outcomes, widespread adoption, or financial return. The FDA also notes that its public list of AI-enabled devices is not comprehensive.[10]

The appropriate posture is neither resistance nor blind enthusiasm.

It is selective adoption: use AI where the task is defined, the evidence is relevant, the errors can be detected, and a qualified professional remains accountable.

What platforms are making harder to leave

The individual AI model may eventually become a commodity. The workflow around it may be harder to replace.

Henry Schein One says its technology supports more than 100,000 global locations, including approximately 90% of the largest DSOs. These are company-reported figures, but they illustrate the advantage of placing AI inside a system that already contains scheduling, treatment, insurance, claims, communications, and financial data.[1]

Align describes an integrated digital platform connecting Invisalign, iTero scanners, and exocad CAD/CAM software.[11] Reporting fourth-quarter and full-year 2025 results, the company cited more than 121,000 active iTero scanner units and more than 70,000 exocad CAD/CAM licenses.[12] Align disputes the European Commission's allegations, and no wrongdoing has been established; the scale figures describe integration, not misconduct.

Dentsply Sirona is pursuing a similar platform logic through DS Core, which connects scanning, design, equipment, and treatment workflows across its portfolio.[13]

None of this proves that dental AI has already consolidated into a durable oligopoly. But the conditions for concentration are forming:

• Data accumulates inside core platforms.
• AI becomes more useful when connected to that data.
• More workflows become dependent on the platform.
• Integrations increase the cost and disruption of switching.
• The platform gains leverage over pricing, access, and product selection.

The model may become a commodity. The workflow may become the moat.

The cost of leaving

Interoperability can sound like a technical concern. For a practice or laboratory owner, it is an economic concern.

Patient records, images, design histories, production presets, communications, and analytics are the accumulated operating knowledge of a business. If they cannot be exported in usable formats, changing vendors becomes expensive—even when the nominal subscription can be canceled.

The American Dental Association has warned that dental imaging remains fragmented by proprietary formats, inconsistent implementations of DICOM, incomplete metadata, and disconnected exchange pathways. It has called for open export specifications and interoperable application programming interfaces.[14]

ISO 18618:2025, the third edition of the standard and published in August 2025, specifies an XML format for transferring dental case data and CAD/CAM data between software systems.[15] It is a real remedy for a specific problem, and it is worth noting how specific that problem is: the standard governs case and design data exchange between CAD/CAM systems. It does not cover patient records, imaging metadata, practice-management data, or the communications and analytics history that make a platform hard to leave. And a published standard creates portability only when vendors implement it consistently and customers insist on it.

The ADA has also urged, in comments on a federal interoperability and prior-authorization proposal, that many dental practice-management systems lack modern interfaces for structured health-data exchange. Tooth numbering, odontograms, surface-level restorations, periodontal measurements, and imaging metadata remain inconsistently supported. Small practices may bear a disproportionate burden when integrations require middleware, manual workarounds, or outside technical support.[16]

This is the practical meaning of platform power:

Whether AI lowers the cost of expertise is still unproven. That platforms can raise the cost of exit is already documented.

Dental laboratories show both possible futures

Consider a five-person dental laboratory.

AI could plausibly help that laboratory review incoming scans, interpret prescriptions, generate routine crown designs, identify incomplete cases, route work, forecast capacity, and respond to customers. In principle, the laboratory could increase output without immediately hiring another CAD technician. The published evidence supports the design-time component of that list; the rest is vendor claim and reasonable extrapolation, not demonstrated result.

That is the accessibility case.

Now imagine that its case portal, scanner connections, design history, manufacturing presets, customer communication, and pricing all reside inside one vendor's cloud. If the vendor changes its fees, restricts an integration, or discontinues a feature, moving to another system may mean rebuilding years of workflow knowledge.

That is dependency.

One small process-mining study points to a laboratory opportunity that may not be design speed alone. It examined ten custom-abutment orders and found that nine involved at least one iteration. Repeated design reviews, rescans, incomplete information, and differences in interpretation created delays and additional work. Ten orders cannot establish a general pattern, but the mechanism it describes is familiar to anyone who has run a laboratory.[17]

AI can potentially detect those problems before they move downstream.

The laboratory of the future may not merely generate designs faster. It may identify which cases are incomplete, which scans are questionable, which cases require expert review, and where production is likely to stall.

Routine morphology, intake, classification, and scheduling are the most plausible candidates for automation. Complex occlusion, full-arch rehabilitation, implant complications, dentures, shade interpretation, esthetics, and final quality control are likely to remain more dependent on skilled technicians.

The technician's value will migrate from producing every step manually toward supervising systems, resolving exceptions, managing quality, and applying judgment where standardized models are weakest.

The same AI can therefore strengthen an independent laboratory—or turn it into a replaceable production node inside someone else's ecosystem. Who captures the value will depend on ownership, differentiation, and bargaining power.

The new divide in dentistry

The future divide will not simply be between organizations that use AI and those that do not.

It will be between organizations that control their AI-enabled workflows and organizations whose workflows are controlled by someone else.

Implementation capacity will matter as much as access to the tool.

Large organizations can standardize data, negotiate integrations, employ compliance teams, and test systems across thousands of cases. A solo practice or small laboratory may have access to the same software but lack the time, expertise, and negotiating power required to use it safely.

Education is already lagging. A 2026 survey of dental students found that 61% had received no formal AI training, while 87.8% identified insufficient training as a barrier. The study has drawn published methodological criticism, and its precise figures should be read with that exchange in mind, but the direction it reports is consistent with the broader picture.[18]

AI literacy is becoming part of professional competence. Dentists and technicians will need to understand not only how to operate a system, but how to question its evidence, recognize its failure modes, and preserve their ability to disagree with it.

Democratizing care is harder than democratizing interpretation

AI may help extend screening, monitoring, education, and specialist support into underserved communities.

That matters. As of December 31, 2025, approximately 63.7 million people in the United States lived in designated dental Health Professional Shortage Areas, and roughly 33% of identified need was being met.[19]

But AI cannot perform a restoration, extract a tooth, treat an emergency, or create a local dental workforce.

It can make interpretation more portable. It cannot make physical care frictionless.

The same distinction applies to patients. A multicentre survey of patients reported broad willingness to accept AI as a complementary diagnostic tool, alongside continued preference for human oversight, privacy protection, and transparent communication.[20]

Patients appear willing to accept AI behind the dentist. They are less willing to accept AI instead of the dentist.