Lesson objective
Identify the four capabilities that sit behind most dental practice AI products — image analysis, speech to text, retrieval and tool use — and state the control each one needs.
Plain-language explanation
Why this matters now
A dental practice does not mainly buy chat windows. It buys software with AI capabilities wired into it: something that looks at a radiograph, something that turns speech into a note, something that answers from a document set, and something that can act — send, book, post, submit. Each of those is a different capability with a different failure and a different control. If a team only knows the chat window, it will apply a chat window's habits to a system that can move money or send a message to a patient.
The dental picture is not speculative. The ADA's SCDI White Paper No. 1106, Overview of Artificial and Augmented Intelligence Uses in Dentistry (approved by the ADA SCDI on 30 December 2022, prepared by SCDI Working Group 13.8), describes AI across clinical disciplines and imaging, and also in non-clinical areas "with a focus on payor topics, such as claims processing, payment integrity and quality assurance, and on dental practice administrative issues involving claims preparation, including attachments". It states that the primary use of AI systems is currently in the payer claim review process: validating the type and content of attachments and screening radiographic images, with claims that meet payer guidelines completing processing and others routed to dental consultants. Two things follow. First, the capability is already pointed at your submissions from the other side of the transaction. Second, the white paper's account describes what payers do; it is not permission for a practice to use AI to shape a claim. Note also its status and its age: a white paper binds nobody, its regulatory and market description dates from 2022, and its Working Group 13.8 voting members included representatives of two makers of FDA-cleared dental radiograph AI, so it is not independent evaluation.
A question to sit with
Of the AI features in your practice, which ones can only tell someone something, and which ones can do something — send, book, post or submit — without a person in between?
What goes wrong
Assuming every capability is the chat window. A team trained only on "check the draft" has no habit for a system that books an appointment on its own.
Treating a box on an image as a finding. Image analysis produces a marked region and a score. That is an input to a licensed dentist's reading, never a finding, never a sentence to a patient, and never a line in a treatment presentation before the dentist has read the images.
Believing a transcript is a record. Speech to text produces text that resembles what was said. It mishears names, numbers, tooth references and negations, and it has no idea which errors matter.
Assuming retrieval means correct. A system that answers from a document set can still pull the wrong document, an outdated version, or a fragment that reads differently in context, and then write fluently around it.
Forgetting that a document the system reads is untrusted input. Content the system reads can carry wording aimed at the system rather than at a person.
Giving an agent more reach than the failure it will have. Reach, not intelligence, decides how bad a wrong output is.
The practical answer
Learn the four capabilities as four different questions.
| Capability | What it does | Typical practice use | How it fails | The control |
|---|---|---|---|---|
| Image analysis | Marks regions in an image and scores them | Software that draws boxes or overlays on radiographs | Marks something that is not what it is scored as; performance varies with the equipment and population it was built on | The licensed dentist reads the images and owns every finding; nothing marked goes to a patient or into a plan before that |
| Speech to text | Converts audio into text, sometimes into a structured draft | Dictation; chairside audio turned into a draft note | Mishears names, numbers, tooth references and negations; drops what it could not hear | The dentist reviews, corrects and adopts the draft; the recording is patient information and is treated as such |
| Retrieval | Finds passages in a document set and answers from them | An assistant answering from your policies, plan booklets or manuals | Retrieves the wrong or outdated passage, then writes confidently around it | Require the quoted passage and its source and date; open the source and read it |
| Tool use and agents | Calls other systems: sends, books, posts, writes, submits | Messaging automations; scheduling assistants; workflows that write into other systems | Executes a wrong output instead of merely displaying it | Limit what it may change; keep a named person's approval in front of anything that reaches a patient, the chart or a payer |
Image analysis is the one to be most careful about, because it is the one that is regulated. FDA publishes a resource list, Artificial Intelligence-Enabled Medical Devices, which it describes as "a resource intended to identify AI-enabled medical devices that are authorized for marketing in the United States". FDA states the list "is not a comprehensive resource" and that devices were identified "primarily based on the use of AI-related terms", and that it will continue to be updated periodically. Read on 15 September 2026, with the page stating "Content current as of: 09/04/2026", the list includes dental entries under product code MYN (medical image analyzer): among them Second Opinion (Pearl Inc., K210365, decided 03/04/2022), Overjet Caries Assist (K212519, 05/10/2022), Videa Caries Assist (VideaHealth, K213795, 04/21/2022), Overjet Periapical Radiolucency Assist (K231678, 09/21/2023), Overjet Caries Assist-Pediatric (K233738, 03/04/2024), Velmeni for Dentists (K240003, 08/30/2024), Better Diagnostics Caries Assist (K241725, 03/11/2025), Videa Dental AI (K251002, 09/19/2025) and Diagnocat (DGNCT LLC, K252934, 01/15/2026). The oldest dental radiograph entry is LOGICON CARIES DETECTOR (Carestream Dental LLC), P980025, decided 09/02/1998 — a premarket approval, not a 510(k) clearance. Most dental radiograph analysis entries list the Radiology panel rather than the Dental panel.
Those names appear here for one reason only: to show that dental radiograph AI sits inside FDA's device framework. They are not a recommendation, a ranking or a comparison, and no product is endorsed by this course. Equally, absence from the list does not mean a product is not a device or is not authorised. Module 8 covers what a cleared indication actually says and what the trials found; this lesson stops at the capability and the control.
Retrieval changes what you should ask for. When a product answers from your documents, the useful request is not "is this right?" but "which passage says so, from which document, of what date?" If the system cannot show you the passage, treat the answer as generated text and check it as such.
Tool use is where reach and risk meet. Sort every feature that can act into three bands, and let the band set the control.
| Band | What it can do | Control |
|---|---|---|
| Read only | Summarises, extracts, suggests; changes nothing | Verification by the person who uses it |
| Drafts for a person | Prepares a message, note, claim or letter that a person must send, adopt or submit | A named approver, every time; nothing auto-sends |
| Acts | Sends, books, posts, writes into the chart, submits | Restrict it; require approval in front of anything reaching a patient, the chart or a payer, and keep a stop switch and a log |
Remember this
- Four capabilities sit behind most dental AI products: image analysis, speech to text, retrieval and tool use.
- A marked region on a radiograph is an input to the licensed dentist's reading, never a finding and never a sentence to a patient.
- A transcript is not a record; the draft becomes a record only when the dentist reviews, corrects and adopts it.
- With retrieval, ask for the passage, the document and the date, then open the document.
- Reach sets risk: read-only, drafts-for-a-person and acts need three different controls, and documents an agent reads are untrusted input.
Dental practice example
Brightfield Pediatric Dentistry is a fictional two-dentist practice that turns on four AI features in one quarter, and practice administrator Colm Deverell writes a single line for each before anything goes live. The first is an imaging assist that overlays marked regions on bitewings. Colm's line reads: reads images, changes nothing, output visible to Dr Amara Solano only. Two weeks in, an assistant asks whether she may point at an overlay while seating a patient. The answer is no: the overlay goes to Dr Solano, who reads the images and owns whatever is said to the family. The second is a dictation tool that turns Dr Solano's chairside speech into a draft note. On its third day it renders "no caries noted on 3" as "caries noted on 3", and a tooth reference as a room number. Dr Solano catches both because she reads and corrects every draft before adopting it; Colm writes into the workflow that the draft is not the record until she does, and that the audio file is patient information stored the way the practice stores patient information. The third is a retrieval assistant loaded with the practice's own policies and three plan booklets. Asked about a fictional Northgate Family plan's missed-appointment policy, it answers confidently and cites "Northgate Member Handbook, page 14". Colm opens the handbook: page 14 is about coordination of benefits, and the missed-appointment wording is in a 2024 version that the plan replaced. He changes the standing instruction so the assistant must quote the passage with the document title and the document date, and he removes the superseded booklet from the set. The fourth is a scheduling automation that can text short-notice openings. Colm puts it in the "acts" band: it may text only from three approved templates, only to patients flagged for the fill list, never to a patient with an unresolved balance dispute, and it stops on a switch the front desk controls. In its first week it proposes a message to a family whose child was seen that morning; the front desk declines it, and the decline is logged. Nothing in Colm's four lines concerns how clever the tools are. Each line says what the feature may read, what it may change, and who approves.
Responsible-use warning
Do not let a marked region on an image, a transcript, or a confident retrieval answer travel any further than the person who owns it. The dentist reads the images and owns the finding; the dentist reviews, corrects and adopts the note; the person who uses a retrieval answer opens the cited document and reads the passage. And never let a feature that can act reach a patient, the chart or a payer without a named approver in front of it. A system that can only display a wrong answer creates a correction. A system that can send one creates a patient message you cannot take back.
Human-review requirement
The licensed dentist reads the radiographs, owns every finding, and owns any patient communication that carries clinical information; ADA policy states an AI system in clinical dental practice should be supervised by a dentist and its limitations acknowledged. The dentist also reviews, corrects and adopts any note produced from dictation before it becomes part of the record. The practice manager or named administrative owner owns retrieval answers used in administrative work, and confirms the cited passage was opened and read. For any feature that can act, the reviewer confirms a named approver stands in front of anything reaching a patient, the chart or a payer, and that the approval is recorded.
Practical exercise
Sort the AI features in your practice by capability and by reach. Use your own software and fictional patients only; do not put patient information into any tool for this exercise.
- List every AI feature your practice uses or is considering, including features inside software you already pay for.
- For each one, name the capability it mainly relies on: image analysis, speech to text, retrieval, tool use, or plain text generation.
- For each one, write the failure that capability has, in one sentence, in your own words.
- Put each feature in a reach band: read only, drafts for a person, or acts.
- For every feature in the "acts" band, write down exactly what it can change, who approves before it acts, and how it is stopped.
- For every retrieval feature, write the standing instruction you will require: quote the passage, name the document, give the document date.
- For every image analysis feature, write one sentence naming the licensed dentist who reads the images and owns the finding, and one sentence forbidding anyone else from relaying a marked region to a patient.
- Identify one feature whose reach exceeds its control today, and draft the change you would propose to the practice owner.
- Keep the sorted list; Module 11 turns it into a workflow and Module 12 revisits the "acts" band.
Knowledge check
-
A radiograph program marks a surface and gives it a score. Who may act on that, and how?
Answer guidance: Only the licensed dentist, who reads the images and owns the finding. The marked region is an input to that reading. No hygienist, assistant, coordinator or front-desk member relays it to a patient as a finding, and it does not appear in a treatment presentation before the dentist has read the images.
-
What does the ADA white paper record about where AI is mainly used today in dentistry, and what limits that statement?
Answer guidance: It states the primary use is in the payer claim review process: validating attachments and screening radiographic images, with conforming claims completing processing and others routed to dental consultants. It is a voluntary white paper approved on 30 December 2022, its market description is from 2022, and vendor representatives sat on the working group, so it is not independent evaluation and is not permission for a practice to use AI to shape a claim.
-
Why is an inbound document that an AI system reads treated as untrusted input?
Answer guidance: NIST AI 100-2 E2025 explains that generative models combine the data and instruction channels, so a third party can manipulate resources the system reads, and the primary user is often the one harmed. Because mitigations are largely empirical without provable guarantees, the practical control is to limit what the system may do rather than to rely on filtering.
-
What are the three reach bands, and what control does each need?
Answer guidance: Read only, which needs verification by the user; drafts for a person, which needs a named approver every time and no auto-send; and acts, which needs restricted permissions, approval in front of anything reaching a patient, the chart or a payer, a stop switch and a log.
Sources and further reading
- American Dental Association, Standards Committee on Dental Informatics — ADA SCDI White Paper No. 1106: Dentistry — Overview of Artificial and Augmented Intelligence Uses in Dentistry (2022-12-30). Accessed 2026-09-15.
- U.S. Food and Drug Administration — Artificial Intelligence-Enabled Medical Devices. Accessed 2026-09-15.
- National Institute of Standards and Technology — Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations (NIST AI 100-2 E2025). Accessed 2026-09-15.
- American Dental Association, House of Delegates — Policy Statement on the Use of Augmented Intelligence in Dentistry (Trans.2021:282), as printed in ADA Current Policies, Adopted 1954-2025. Accessed 2026-09-15.