AI Agents for General Dentists: Treatment Plans and Recall Follow-Up
An AI agent can track unaccepted treatment plans and overdue recalls for a general dental office, draft follow-ups and route exceptions to staff. Dentists keep every clinical decision.

A general dental office can use an AI agent to track treatment plans the patient has not accepted and recall visits that are overdue, then prompt the front desk or the patient at the right time with the right message. The agent works from records the office already controls and stops at the edge of clinical judgment: it never changes a diagnosis, a treatment sequence or a recommendation. The value comes from consistent follow-up with clear permissions, not from software that decides what a patient needs.
Key takeaways
- Start with two queues only: unaccepted treatment plans and overdue recalls, each with a named human owner.
- Give the agent read access to status and dates, and let it write only drafts and tasks until the office approves more.
- Keep every diagnosis, sequencing decision and clinical message with the dentist of record.
- Judge the work with a short weekly review of open items, contacts made and outcomes, not with a dashboard that nobody acts on.
- Expand scope in small steps once the first queue runs cleanly for a sustained period.
Why unaccepted plans and overdue recalls slip
Most general practices do the clinical work carefully and then lose track of what happens afterward. A treatment plan is presented at the chair, the patient says they will think about it, and the status stays open in the practice management system while the schedule fills with new patients. Recall lists are often generated in batches, so a patient who was due in the spring may not appear until someone runs the report again. Nobody has decided to neglect these patients. The follow-up simply has no owner, no rhythm and no record of what was already tried.
That is the gap an agent can fill. It has no hands in the mouth and no authority over the diagnosis. Its job is narrow: read status and dates, notice what has gone quiet, and put the next action in front of a person. Offices that already use reminder software will recognize part of this, but a reminder sends a fixed message on a fixed schedule. An agent can check context first, such as whether the plan changed, whether the recall was already booked, or whether the patient replied with a question that needs a clinician. For the insurance side of the same workflow, see Superintelligent Agents for Dental Offices: Recalls and Insurance Admin. A parallel case in another recall-driven practice is covered in AI Agent for a Chiropractic Practice: Intake and Visit Recalls.
Where the agent's authority should stop
Write the boundary before the workflow. The useful question is not what the agent could do with patient data, but which actions the dentist would be comfortable seeing logged under the practice's name.
The agent can do the following:
- List open treatment plans by presentation date, status and last recorded contact.
- Flag plans with no contact inside the follow-up window the dentist has approved.
- Flag recalls past their due date, grouped by appointment type.
- Draft a reminder from approved wording that names the presented treatment and offers a booking route.
- Create a task for the coordinator with the relevant record attached.
- Log every draft, send, reply and outcome with a timestamp.
The agent must not do the following:
- Change a diagnosis, a treatment sequence or the priority of clinical work.
- Suggest a clinical alternative or interpret radiographs, charting or photographs.
- Tell a patient whether treatment is still needed, urgent or optional.
- Offer discounts, payment exceptions or financing terms the office has not approved.
- Answer a message that describes pain, swelling, bleeding or any new symptom. It escalates that message at once to a clinical person.
The last item deserves the hardest testing, because a patient reply is where administrative and clinical questions mix. A message that begins as a question about scheduling can end as a symptom report.
A step-by-step plan for the first queue
Start with one queue. Unaccepted treatment plans are often the better first choice because the cost of a wrong message is higher and the benefit of consistent follow-up is easy to see. Use this sequence:
- Step 1: Name one owner. Choose the coordinator or front desk lead who will review the queue each day and approve drafts.
- Step 2: Write the rules with the dentist. Define what counts as a stale plan, which contact channels are allowed, how many attempts are reasonable, and when a conversation goes to the dentist.
- Step 3: Grant scoped access. Connect the agent to treatment and scheduling records through a credential limited to reading those fields and writing tasks and drafts.
- Step 4: Run in draft mode. The agent prepares every message, and a person approves each one before it leaves the office.
- Step 5: Review a sample each week. Compare drafts against the rules, log every mismatch, and adjust the wording or the flags.
- Step 6: Expand only after a clean run. Allow routine, template-based reminders to send without approval only for the cases the review has shown to be reliable.
Recalls follow the same sequence with a separate rule set. The dentist sets the recall intervals, and the agent applies them as written rather than deciding that a patient should come sooner.
An illustrative scenario for an overdue recall
Illustrative scenario: a general practice has a patient whose routine recall passed without a booking, and a treatment plan for a filling presented at the previous visit is still unaccepted. The agent places both items in the coordinator's queue. It notices a recent message from the patient asking about parking hours, which tells it the patient is reachable. It drafts one message that mentions the recall window and the open treatment, using approved wording, and does not describe the filling as necessary or urgent. The coordinator reviews the draft, edits one sentence and sends it.
Two days later the patient replies that one side has been sensitive to cold. The agent does not answer that message. It routes the reply to the dentist of record with the patient's notes attached, marks the thread as clinical, and tells the coordinator it needs a call. The coordinator phones the patient the same day.
The agent improved timing and context in this example. Every clinical judgment stayed with the dentist, and the record shows who saw what and when.
Permissions, records and audit trails
A dental office handles sensitive health information, so access design matters as much as message design. The OWASP authorization guidance describes a principle worth copying: deny by default and grant only what a role needs. For this workflow that means read access to specific status and date fields, write access to drafts and tasks, and no ability to edit the clinical chart. The NIST AI Risk Management Framework offers a useful structure for the review cycle: map the risks of the use, measure how the agent behaves, manage the controls and assign clear accountability.
Treat every inbound patient message as untrusted input. A reply can contain text that tries to override the agent's instructions. The OWASP guidance on prompt injection recommends keeping instructions separate from data, limiting what the agent can do with untrusted content, and requiring a person to approve consequential actions. The same design principle appears in the broader agent literature, where Anthropic describes simple, well-bounded patterns as the stronger starting point before adding autonomy.
Keep an audit log for each item with the source record, the rule that triggered the flag, the draft, the approver and the final outcome. When a question comes up months later, the log should answer it without anyone reconstructing events from memory. Privacy obligations depend on where the practice operates and on its arrangements with software vendors, so confirm them with the practice's compliance advisor before connecting any system.
How Be Human runs this in managed operations
Be Human designs this kind of agent as a managed role rather than a one-time setup. A first engagement defines the single role, the memory and tools it needs, the approval boundaries and an observable success criterion, such as the share of flagged plans that receive a reviewed contact within the approved window. Managed operations can then cover monitoring of the queue, recurring evaluations of drafts against the dentist's rules, permission reviews when the practice management system or messaging tool changes, incident response when something goes wrong, and controlled improvements to templates. The service is remote and available in English and Portuguese.
If you want to find the first task in your practice that is worth handing to an agent, take the free Be Human AI diagnostic.
Frequently asked questions
Can an AI agent recommend treatment to a patient? No, and it should not be configured to try. Diagnosis, sequencing and clinical recommendations belong to the dentist of record. The agent can carry approved plan wording and presentation dates into a reminder, and it routes any question about whether treatment is still needed to a clinical person.
Does this replace the front desk or the treatment coordinator? No. The agent prepares the queue, drafts the message and logs the outcome, while a person approves outbound contact in the early stages and handles bookings, insurance questions and anything emotionally sensitive. Its purpose is to make the coordinator's day more predictable and less dependent on memory.
How does an office keep patient data safe with an agent? Start with scoped access: the agent should see only the fields it needs, use credentials that can be revoked, and write to drafts and tasks rather than to clinical records. Treat patient messages as untrusted input, since a reply can contain instructions. Review privacy obligations for the jurisdiction where the practice operates with its own compliance advisor.
What should we measure in the first months? Measure the queue itself: how many plans and recalls are open, how many received a reviewed contact, how many were answered or booked, and how many drafts staff corrected. Corrections are the most useful signal because they show where the rules are unclear or wrong. Avoid vanity metrics that do not lead to a decision about the rules, the templates or the scope.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


