Scenario · Health, wellness & education

AI automation for a dental practice: what a three-chair office automated, what it bought, and what stayed human

AI automation for dentists, honestly: a Denver practice bought its phone AI, built two narrow automations with human review, and handled HIPAA. Costs included.

By Akash Bansal, founding engineerUpdated 5 November 20267 min read

Illustrative scenario. The founder in this story is a composite, not a real client. The process, costs and decisions are drawn from builds I have done; names and business details are invented to show how a build like this typically goes.

"AI automation for dentists" mostly gets sold as an AI that answers your phones. That part is real and worth buying. The bigger wins in a small practice, though, are often quieter: the hours the front desk spends every week checking insurance, chasing overdue recalls and retyping intake forms.

This is a composite story: the practice and owner are invented, and the plan is typical for this kind of build. My evidence is two real systems. Aviate reads inbound résumés from Gmail, scores them and surfaces the best ones, unattended, using cursor pagination and a bounded concurrency pool to stay within API quotas. Brahmo grades long-form exam answers through a multi-step AI pipeline where a human expert can override any score. A dental office needs both patterns: unattended where the stakes are low, a human in the loop where they aren't.

The practice and the starting point

Call her Dr. Teresa Lindqvist. She owns a three-chair general dentistry practice in Denver with two hygienists, two front-desk staff and an office manager. The practice runs on Open Dental. Her pain list, in her own order:

• Missed calls during the morning rush and at lunch, many from new patients.
• The front desk spends the afternoon before each clinic day checking insurance eligibility for tomorrow's patients.
• Hundreds of patients overdue for their six-month recall, and nobody has time to work the list.
• New-patient intake forms that the dentist reads in full, just before the appointment.

Her brief was "an AI that runs the front desk". That brief would have automated the riskiest part first.

Which parts of a dental office should AI automate?

Sort each task by two questions: does it touch clinical judgement, and what happens if the AI gets it wrong? Tasks that are routine and cheap to get wrong can run unattended. Tasks that reach patients, involve money or carry clinical meaning get a human check.

Why buy the phone answering first?

Because the problem is large and the products are mature. One vendor's analysis of 4,280 calls across 26 dental practices in February 2026 found 38% went unanswered, and new-patient calls converted to bookings at about 25% (Peerlogic, vendor data, so read it as directional). Teresa's own phone log was the better source: she counted her missed calls for two weeks before choosing anything. An AI receptionist on overflow and after-hours, with a signed BAA, was live within the month for a monthly fee. No custom code.

Who built the rest, and why

For the back-office automations, Teresa looked at an all-in-one dental AI platform and at a local IT provider. The platform didn't cover the insurance prep the way her front desk worked, and the IT provider didn't build software. She chose a founding-engineer build: one senior engineer responsible for design, code, hosting and handover, on a fixed quote after scoping. On my builds the client owns the code, repositories and cloud accounts from day one, sees a working demo every week, and gets written design docs.

The deciding fact was integration. Open Dental publishes API access per location, with a free read-only tier and paid tiers of $15, $30 and $35 a month for wider access (Open Dental API permissions). If your practice software has no API, the scope changes completely, so check this first.

The build, week by week

1. Week 1: data map and HIPAA plan. Every field the automations read, where it goes and why. Which vendors touch patient data, and a BAA for each one.
2. Week 2: read-only integration. Tomorrow's schedule, patients and insurance details, pulled from the practice software on a schedule. Nothing written back yet.
3. Week 3: overnight insurance prep. A job runs each evening, gathers what the front desk needs to verify for each of tomorrow's patients, and flags gaps: missing subscriber details, plans that look expired, frequency limits likely to matter for the planned procedure. Like Aviate, it runs unattended in a bounded pool so it never floods the API.
4. Week 4: morning review screen. One page per day: green rows need nothing, amber rows need a call, with the reason in a sentence. The front desk works the amber rows and marks them done.
5. Week 5: recall drafts. Each week the system lists overdue patients and drafts a short, personal message for each one. The office manager approves, edits or skips in a batch, the Brahmo pattern: AI does the first pass, a person has the final say.
6. Week 6: logs, handover, training. An access log, a written runbook, and a session with the front desk.

The patient-data rules you can't skip

A dental practice is a HIPAA covered entity, and anything that handles patient information on its behalf is a business associate. This is general information from an engineer, not legal advice; have a qualified person review your setup.

• A BAA with every vendor in the data path. HHS guidance says a cloud provider that creates, receives, maintains or transmits protected health information for you is a business associate, even if it only stores encrypted data, and you need a BAA with it (HHS, HIPAA and cloud computing). That includes the AI model provider. If a provider won't sign one for your use, don't send it patient data.
• Minimum necessary. The insurance job needs plan details and the planned procedure, not the full chart. Send models only the fields a step needs.
• Know the safeguards. The Privacy Rule covers who may use and share patient information; the Security Rule covers how electronic records are protected: access control, audit logs, encryption.
• A breach plan. Breaches affecting 500 or more people must be reported to HHS within 60 days of discovery, and individuals must be notified too (HHS Breach Notification Rule). Know who would make that call.
• Colorado rules too. State privacy and breach laws can add requirements. Ask your adviser what applies to your practice.

What the build and running costs look like

My starting price for an AI feature or agent is $8,000, on a fixed quote after scoping. For these two automations, insurance prep and recall drafts with review screens, my rough estimate is $8,000–$14,000 over four to six weeks, depending mostly on how messy the insurance data is and whether anything is written back to the practice software. If your own list of automations looks different, the app cost calculator gives a rough range with each AI part priced separately.

The real return is staff time. Teresa measured it the plain way: front-desk hours spent on insurance prep in the two weeks before launch, and in the two weeks after.

What went wrong

• Insurance data was messier than the schedule. Subscriber names spelt differently, plans entered under an employer that had changed. The first runs flagged too many rows amber. Tightening the rules took two rounds with the front desk, who knew which mismatches actually mattered.
• Recall drafts sounded like a chatbot. The first drafts were polite and generic. The office manager wrote five example messages in her own voice; the drafts improved immediately, and the approval rate climbed.
• A patient asked not to be texted. That preference lived in a free-text note, not a field. Communication preferences became a hard filter before any draft is created.
• The overnight job ran into the morning once. A slow API night pushed it past opening time. The job now starts earlier, retries politely, and the review screen shows clearly which rows are not ready yet.

What launch looked like

The AI receptionist went live first, on after-hours calls only, then midday overflow. The insurance prep ran in parallel with the old manual process for a week, so the front desk could compare every row. Recall drafts started with 20 patients a week, all approved by hand. Teresa's numbers to watch: missed calls per week, front-desk hours on insurance prep, and recall appointments booked from drafted messages.

Checklist for a dental practice considering AI automation

1. Count your missed calls for two weeks before buying anything.
2. Buy phone answering and reminders. Build only what your tools can't do.
3. Check your practice software's API and its pricing first.
4. List every vendor that will touch patient data, and get a BAA from each.
5. Start read-only. Write back to the practice software only once you trust the output.
6. Let AI draft; let staff approve. Loosen that only with evidence.
7. Keep clinical decisions with the dentist.
8. Own the code, the cloud account and the logs yourself.

For the same unattended-pipeline pattern in another industry, see the résumé-screening automation story; for a clinic that kept its practice software and built only the gap, the physio clinic booking story. If your practice has a back-office task like these that your software can't handle, this is the kind of focused AI build I do; here's how I work.