Why Practices Need Human Oversight for AI-Drafted Patient Communications

AI can speed up consultation summaries, insurance claims, and follow-up messages—but only careful human review keeps a fast draft from becoming a costly mistake. Five tips to use AI safely in your practice.

By Gleb Tsipursky, PhD

Orthodontists can save time on administrative writing with AI, but an efficient draft can still create an expensive mistake. A polished AI-generated message may omit an important condition, invent a treatment detail, or promise something the orthodontist never approved. Practices need a simple way to use AI for repetitive communication while keeping patient care, billing accuracy, and professional judgment under human control.

Here are five tips to help your practice get started.

1. Begin With Three Everyday Workflows

Three common documentation workflows in the orthodontic practice lend themselves to AI use. Consider the explanation a treatment coordinator sends after a consultation. An approved AI tool might turn an office-approved outline into plain-language instructions about appointment scheduling, appliance care, and routine next steps. The clinical reviewer must check the text against the actual plan. AI should never invent a treatment duration, guarantee an outcome, or replace an individualized informed-consent conversation.

Insurance claim narratives are another option. Billing staff routinely organize chart information into a payer’s requested format. AI can help reorganize accurate material, but a named reviewer must confirm dates, procedure descriptions, supporting documentation, and medical-necessity language. A claim that sounds plausible but lacks support could cause reimbursement problems and undermine credibility.

Follow-up messages carry different risks. A practice could draft routine missed-appointment reminders or directions for scheduling a retainer check. A complaint about unusual pain, a broken bracket, or a worsening symptom, however, should trigger a clinical escalation, not an automatically generated message. Staff must recognize when routine patient communication has turned into a request for care.

2. Keep the Review Log Small and Useful

To get started, run a pilot program. Assign an owner for each workflow before testing a tool. Record the task, the approved input source, the tool or template, the reviewer, any material corrections, and the final decision: approved, revised, escalated, or rejected. A shared spreadsheet can suffice for a pilot. The log should help the orthodontist spot recurring errors without requiring staff to document every routine office-hours reminder.

Before expanding the pilot, agree on a release rule: No AI-generated patient communication goes out automatically. Even when a software vendor offers one-click messaging, staff should see the entire final text and know who can approve it. Build a short escalation guide for unusual symptoms, questions about treatment changes, insurance disputes, and patient complaints. If the practice sends instructions in another language, arrange review by someone qualified to check their clinical meaning. A translation that reads fluently can still change what a patient is asked to do.

Match review intensity to consequences. Generic scheduling language may need a quick template check. Informed-consent material, aftercare instructions, individual insurance claims, and anything containing patient-specific clinical or financial facts require a tighter review. The reviewer should ask whether the draft matches the patient record, reflects practice policy, introduces unsupported claims, and needs a clinician’s approval.

This system works only when team members feel able to stop a flawed draft. If office managers reward speed alone, staff may fix errors quietly or overlook them. Invite assistants, treatment coordinators, and billing specialists to describe what goes wrong and revise the checklist around those mistakes.

3. Set Privacy Boundaries Before Anyone Experiments

Start with invented patient examples and approved generic templates. Do not paste patient charts into an unapproved consumer AI tool. The HHS guidance on HIPAA and cloud computing explains that covered entities generally need an appropriate business associate agreement and safeguards when a cloud vendor handles electronic protected health information on their behalf. A paid subscription alone does not establish compliance.

Specify which tools employees can use, what information they may enter, where outputs are stored, and who investigates a suspected disclosure. When a vendor claims its product is appropriate for healthcare, request the relevant documentation and assess the actual configuration and intended use.

4. Distinguish Drafting from Clinical AI

The American Dental Association’s work on AI standards addresses the need for appropriate evidence and evaluation in dental applications. That matters when vendors group administrative writing tools and clinical imaging systems under the same AI label. A system that writes a readable payment explanation has not thereby demonstrated diagnostic accuracy or clinical safety.

Ask what each product was tested to do, who must check its results, and how failures are handled. For clinical applications, scrutinize whether evaluation data resemble the practice’s patients and cases. For administrative uses, examine factual fidelity and escalation reliability. NIST’s AI Risk Management Framework provides a practical structure for defining intended uses, assessing possible harms, measuring performance, and adjusting safeguards.

5. Measure Time Saved After Human Review

Pick one task and test it on a small set of invented or appropriately de-identified examples. Record drafting time, review time, corrections, rework, and any need for follow-up clarification. Compare those totals with the current workflow. An AI draft that saves three minutes but requires six additional minutes to verify has failed the efficiency test. Repeat the test with several staff members rather than relying on the fastest early adopter. Include at least one awkward but plausible case in which information is missing, and the correct response is to ask a clinician instead of composing a reassuring answer.

A successful pilot should leave the orthodontist with better evidence, fewer avoidable errors, and a process staff can follow when the unexpected happens. Expand only after the review log shows that the team can protect accuracy and patient trust while making the work easier. OP

Adapted from: The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026)

Photo: ID 175370101 © Nd3000 | Dreamstime.com

Gleb Tsipursky, PhD, a behavioral scientist called the “Office Whisperer” by The New York Times, helps tech-forward leaders stop overpaying for AI while boosting engagement and innovation. He serves as the CEO of the AI consultancy Disaster Avoidance Experts and has written eight books, including The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026). He previously taught at Ohio State University and the University of North Carolina at Chapel Hill. He works in Columbus, Ohio, and can be reached at gleb@disasteravoidanceexperts.com.