Guide · Published Nov 18, 2026

AI for Medical and Dental Practice Administration

For owners and office managers deciding where AI should answer phones, draft notes and handle the inbox, and what rules it runs under. You leave with the four jobs in the order to take them on, the evidence for each, a one-page AI policy you can adopt today, and separate medical and dental benchmarks.

The free Business Associate Agreement Tracker: one list of every vendor that touches patient data, AI tools included, with its BAA status, the 12 required terms, red and amber alerts for gaps, and a dashboard that counts the vendor-days you have gone without a signed agreement.

Get the free BAA Tracker

Evidence on this page checked Nov 18, 2026. Next check: when MGMA, Doximity or ADA HPI publish new AI data.

AI for medical and dental practice administration covers four jobs: answering phones, drafting visit notes, drafting inbox replies, and the policy that governs all three. Adoption is real: 63% of physicians used AI by January 2026 (Doximity, 2026) and 43.3% of dentists by mid-2026 (ADA HPI, 2026). Start with the policy and signed BAAs.

Key takeaways

  • AI plays some role in patient visits at 83% of medical groups, up from 71% a year earlier, and 45% use it in more than a quarter of visits (MGMA Stat, Aug 2026, 189 responses).
  • Ambient scribe users spent 8.5% less EHR time per appointment, about 2.4 minutes (JAMA Network Open, Oct 2025), and a randomized trial measured 0.36 fewer hours a day on notes (NEJM AI, Nov 2025).
  • Burnout fell from 51.9% to 38.8% after 30 days with an ambient scribe across 6 health systems (JAMA Network Open, Oct 2025).
  • AI-drafted inbox replies were used 20% of the time and lowered task load, but did not save measured time (JAMA Network Open, 2024).
  • Any AI vendor that creates, receives, maintains or transmits patient information for you is a business associate and needs a signed agreement first (45 CFR 160.103; 164.502(e)).

The four jobs AI can take on, at a glance

AI in practice administration does four jobs. Three of them do work; the fourth decides what the other three may touch. This table is the page in one view.

The four AI jobs in practice administration: what AI does, what stays human, and the rule
JobWhat AI does wellWhat stays humanThe rule that decides itStart with
1. Phones and the front deskAnswers overflow and after-hours calls, books and moves visits, takes messages, routes by reasonAnything urgent or clinical; upset callers; exceptionsBAA; consent for AI-voice calls (FCC, 2024)AI receptionist prices compared
2. Visit notesDrafts the note from the visit conversationThe clinician reviews and signs every noteBAA; consent to record where the state requires itThe scribe evidence
3. Inbox and messagesSorts messages by type; drafts replies; summarizes threadsA person reads and sends; clinical content reviewed by a clinicianCalifornia AB 3030 for clinical messagesPortal message triage policy
4. PolicyNothing: this job is humanThe owner signs; the office manager runs the rhythm45 CFR 164.308 training and risk analysisThe one-page AI policy

How far adoption has gone

Adoption has moved from pilots to routine in about a year. In Doximity's surveys of 3,151 US physicians, AI use rose from 47% in March to April 2025 to 63% in November 2025 to January 2026, and use of voice documentation or ambient scribes rose from 20% to 29% (Doximity, 2026; industry research by a company that sells AI tools). The top concern, named by 71%, was the accuracy of what AI produces.

Nearly two in three physicians now use AI somewhere. Any use 47% to 63%; ambient scribes 20% to 29%. Source: Doximity State of AI in Medicine, 2026, labelled vendor research.

Medical groups report the same shift. In an MGMA Stat poll of Aug 4, 2026, AI played some role in patient visits at 83% of groups, up from 71% a year earlier, and the share using it in more than a quarter of visits nearly doubled, from 24% to 45% (MGMA Stat, 2026). In the front office, scheduling (31%) and calls (27%) led the AI and automation priorities in a February 2026 poll of 177 leaders (MGMA Stat, Feb 2026).

AI moved from some visits to many visits in one year. Some AI in visits 71% to 83%; more than a quarter of visits 24% to 45%; 76% or more of visits 7% to 17%. Source: MGMA Stat poll, Aug 4, 2026, 189 responses.

Patients are warier than clinicians. In Pew Research Center's survey of 11,004 US adults in December 2022, 60% said they would be uncomfortable with their own provider relying on AI, and three-quarters worried that providers would move too fast (Pew Research Center, 2023). The survey predates most tools now in use, but the direction is the practical point: tell patients, and keep a person in charge of every decision.

The order to take it on

Take the jobs in this order. Risk rises down the list while setup work grows, and the policy makes every later step quicker to approve.

  1. Policy and agreements. Write the one-page policy below, list every AI tool already in use, including staff who use personal chatbot accounts, and get signed BAAs or stop patient data going in. An afternoon of work.
  2. The inbox, with a person sending. Drafting replies is low risk when a person reads and sends every message, and the evidence shows lower task load even without time savings.
  3. Phones: overflow and after hours first. Let AI take the calls your team misses, not the ones it already answers well. Count missed calls for a week before and after.
  4. Notes, decided by the clinicians. Scribes have the best evidence and the most clinical exposure. Pilot with volunteers for 30 days and measure.

Job 1: phones and the front desk

Phones are where admin AI pays first, because missed calls can be counted. At Valley View's clinics, some had more than 50% of incoming calls going to voicemail before a phone redesign; the goal became a missed-call rate of 10% or lower, missed calls correlated at -0.69 with schedule use, and phone and voicemail work fell by 48 staff hours a month (MGMA, 2023; single-system case study). A receptionist in health care earns a median $19.00 an hour (BLS, May 2025), so the real question is not whether AI is cheaper than a person but which calls a person should never miss.

What an AI phone agent should do: answer overflow and after-hours calls, book, move and cancel visits in the schedule, take structured messages, send intake links, and route calls by reason. What it must not do: judge how urgent a symptom is, give clinical advice, or handle an emergency beyond telling the caller to hang up and call 911 and alerting staff. A caller must reach a person in one step, at any time.

Three rules apply. First, call audio with names and numbers is patient information, and voiceprints are one of HIPAA's 18 identifiers (45 CFR 164.514(b)(2)), so the vendor needs a BAA. Second, outbound calls: the FCC ruled in February 2024 that "the TCPA's restrictions on the use of 'artificial or prerecorded voice' encompass current AI technologies that generate human voices" (FCC 24-17), so AI-voice reminder calls need prior express consent unless an emergency purpose or an exemption applies. Third, recording: California makes it an offense to record a confidential communication without the consent of all parties (Cal. Penal Code 632), and other states have similar rules, so announce that the call is recorded and answered by an automated assistant.

Compare tools on the AI receptionist price comparison for medical practices and its dental office counterpart, and check your phone system requirements and after-hours call policy first. The Practice Diagnostic counts the missed calls; the 14-Day Front Door Fix, with the AI phone and inbox setup add-on, installs the agent, the handoff and the scripts.

List every tool before you add one

Most offices already have AI in the building: a scribe trial, a texting tool, a chatbot someone uses for letters. The free tracker lists each vendor, its agreement and the gaps, with red flags for anything touching patient data without a BAA.

Get the free BAA Tracker

Job 2: visit notes and the ambient scribe

Scribes have the strongest evidence of any admin AI, and the honest size of the effect is modest. Four studies published in the last 14 months are the ones to know:

What ambient AI scribes did in published studies, newest first
StudyDesignWhat it found
Afshar et al., NEJM AI, Nov 202566 practitioners, randomized0.36 fewer hours a day on notes (about 22 minutes); work exhaustion down; professional fulfillment unchanged
Lukac et al., NEJM AI, Nov 2025238 physicians, randomized, two scribes vs usual careOne scribe cut time in notes; the other did not; occasional inaccuracies with both
Pearlman et al., JAMA Netw Open, Oct 2025603 clinicians (125 users), matched comparison8.5% less EHR time per appointment (about 2.4 minutes); 15.9% less note time; after-hours time unchanged
Olson et al., JAMA Netw Open, Oct 2025263 clinicians, 6 health systems, before and afterBurnout fell from 51.9% to 38.8% after 30 days
The clearest gain is in how clinicians feel. Burnout 51.9% before and 38.8% after 30 days with a scribe. Source: Olson et al., JAMA Network Open, Oct 2, 2025.

Kevin Pearlman, who helped lead the UChicago study, put the time result plainly: "At first glance, an 8.5% reduction in documentation time might seem small" (UChicago Medicine, 2025). The same article notes that two or three minutes per patient adds up to hours a week for a clinician seeing 20 patients a day. Set expectations there: about 22 minutes a day per clinician in the randomized trial.

Four operating rules make a scribe safe. The clinician reviews and signs every note the same day. The patient is told and agrees before the scribe listens, recorded in the chart. The vendor signs a BAA before the first visit. And the pilot runs 30 days with volunteers, measuring note time, after-hours time and how often notes need correcting; if errors reach the record, stop and fix before going further.

Job 3: the inbox and patient messages

Drafted replies help people feel less buried, but the best study found they do not give back hours. At Stanford, 162 clinicians used AI-drafted replies to patient messages for five weeks in 2023: drafts were used 20% of the time, task load fell from 61.31 to 47.26 and work exhaustion from 1.95 to 1.62, but the time spent reading and replying did not change significantly (Garcia et al., JAMA Network Open, 2024).

In the front office, the safe uses are sorting portal messages by type (refill, scheduling, records, clinical question), drafting scheduling and administrative replies, and summarizing long threads for the clinician. The rule never changes: a person reads and sends every message, and nothing clinical leaves without a clinician's review.

State law is starting to set the floor. California's AB 3030, in effect since Jan 1, 2025, requires that a generative AI message to a patient about clinical information carry a disclaimer and clear instructions for reaching a human, unless a licensed or certified provider read and reviewed it (Cal. Health and Safety Code 1339.75). Administrative messages such as scheduling sit outside it. Other states are writing their own rules, so check yours before any patient-facing AI goes live. Start with the portal message triage policy, the reminder text templates and HIPAA-compliant texting.

Job 4: the one-page AI policy

The policy comes first because it is cheap and it governs everything else. Here is one you can adopt today; replace the bracketed items and have the owner sign it.

[Practice name] AI use policy

Purpose: to let staff use AI tools for administrative work without exposing patient information or letting a machine make a decision a person should make.

  1. Approved tools only. Staff use only tools on the approved list for practice work. The list names each tool, what it may be used for, the plan we pay for, and whether a signed BAA is on file.
  2. No patient information without a BAA. Names, dates, phone numbers, chart numbers, voice recordings and photos are all identifiers. None goes into a tool without a signed BAA. Personal AI accounts are never used for practice work.
  3. A person decides. AI drafts; people send, book and sign. A clinician reviews every note and every message with clinical content before it reaches the record or a patient.
  4. Tell patients. Patients are told when they are speaking with an automated assistant, when a visit is recorded for a scribe, and when a clinical message was drafted by AI and not reviewed by a clinician.
  5. Consent to record. The scribe listens only with the patient's agreement, noted in the chart, and never when the patient declines.
  6. Automated calls and texts. No AI-voice call or automated text goes to a patient without the consent our reminder policy requires.
  7. New tools. Anyone may propose a tool. The [privacy officer] approves it after the vendor questions and a signed BAA.
  8. Mistakes. Any AI error that reaches a patient or the record is reported to the [privacy officer] the same day and logged.
  9. Training. Every staff member is trained on this policy at hire and once a year, with HIPAA security training.
  10. Review. The [owner] reviews this policy every [quarter].

Signed: [owner] · Date: [date] · Version: [1.0]

Each line of the policy rests on a rule. The table shows which, and what it means for an AI tool:

The rules behind the policy, and what each means for AI tools
RuleWhat it saysWhat it means for AI
45 CFR 160.103A business associate creates, receives, maintains or transmits PHI on your behalfScribes, phone agents and inbox tools that touch patient data are business associates
45 CFR 164.502(e)Satisfactory assurances must be documented in a written agreementNo signed BAA, no patient data
45 CFR 164.514(b)(2)18 identifiers, including names, dates, phone numbers, medical record numbers, voiceprints and full-face photosRemoving names is not de-identifying
45 CFR 164.308(a)(5) and (a)(1)Security training for all workforce, management included; an accurate and thorough risk analysisAI tools belong in the training and the risk analysis
FCC 24-17, Feb 2024AI-generated voices are artificial voices under the TCPAAI-voice calls need prior express consent unless exempt
Cal. HSC 1339.75 (AB 3030)AI messages about clinical information need a disclaimer and a route to a human, unless a licensed provider reviewed themReview clinical drafts, or disclose
Cal. Penal Code 632Recording a confidential communication needs the consent of all partiesAsk before the scribe listens; announce call recording

Practice Controls is the fixed-scope build when you want the policy, the approved list and the review rhythm written and handed over with your other controls.

Ten questions to ask any AI vendor

Ask these before any trial that touches patient information, and keep the answers with the BAA.

Ten questions for an AI vendor, why each matters, and a good answer
QuestionWhy it mattersA good answer
Will you sign a BAA, and on which plan?Some vendors sign only on higher tiersYes, on the plan we are buying, before go-live
Is our data used to train your models?Data in a training set cannot be pulled backNo, in writing
Where is audio and text stored, and for how long?Retention belongs in your risk analysisA stated period you can shorten
Which subcontractors handle our data?Subcontractors are business associates tooA named list, each under a BAA
Which systems does it connect to?Copy-and-paste workarounds leak dataA direct link to your EHR or practice system, or none
How does a caller or user reach a person?Handoff is the safety valveOne step, any time, with context passed on
What happens in an emergency?Triage is clinicalIt says call 911 and alerts staff; it never judges urgency
Which languages does it handle?Spanish-speaking patients are commonNamed languages, tested with your scripts
What do you log, and can we audit it?You need to check its workLogs you can export and sample
How do we leave?Lock-in is a costExport in a standard format and deletion on request

For medical practices: the benchmarks

The medical evidence is the deepest of any admin AI topic, but most of it comes from large systems. Read it as direction, then measure your own practice for a week before and after any change.

Medical practice AI benchmarks, newest first within each group
MetricValuePopulationYearSource
Medical groups using AI in some visits83% (71% a year earlier)189 MGMA poll responses2026MGMA Stat
Groups using AI in more than a quarter of visits45% (24%)Same poll2026MGMA Stat
Physicians using AI63% (47% in spring 2025)3,151 US physicians2026Doximity (vendor research)
Physicians using ambient scribes29% (20%)Same2026Doximity
Top concern: accuracy of AI output71%Same2026Doximity
Front-office priorities: scheduling; calls31%; 27%177 MGMA poll responses2026MGMA Stat
Note time saved, randomized trial0.36 hours a day66 practitioners2025NEJM AI
EHR time per appointment, scribe users8.5% less (about 2.4 minutes)603 clinicians2025JAMA Netw Open
Burnout before and after a scribe51.9% to 38.8%263 clinicians2025JAMA Netw Open
AI-drafted inbox replies used20%; no time saved162 clinicians2024JAMA Netw Open
Adults uncomfortable with a provider relying on AI60%11,004 US adults2022 surveyPew
Calls to voicemail before a phone redesignMore than 50% at some clinics; goal 10%One health system2023MGMA (case study)
Median receptionist pay in health care$19.00 an hourUSMay 2025BLS

For dental practices: the benchmarks

Dentists have adopted AI almost as widely, but for different work. In the ADA Health Policy Institute's mid-2026 survey, 43.3% of dentists used AI for at least one task and another 26.4% planned to start. Imaging and diagnostics led current use at 22.8%; reception or front-desk check-in was 10.1%; and among dentists not yet using AI, charting and note-taking was the top planned use at 34.8% (ADA HPI, July 2026). HPI's summary: "Dentists are less likely to use AI for high-level clinical tasks," and 82.6% did not plan to use it for treatment recommendations.

Imaging leads dental AI; the front desk is still early. Any task 43.3%; imaging 22.8%; explaining findings 13.2%; social media 10.7%; reception check-in 10.1%; charting planned 34.8%. Source: ADA HPI Perspectives, July 2026.

The phone matters even more in a dental office. In Liine's 2024 benchmark of nearly 50,000 new-patient leads, 72.9% of first contacts were phone calls and 57.1% of inquiries became a scheduled appointment, with staff booking from 25.5% to 74.1% of their calls (Liine, 2024; vendor data). An AI agent that answers overflow is a booking tool before it is an admin tool, and it should be judged by bookings, not minutes saved. See the dental AI receptionist comparison and the dental practice operations hub.

Dental practice AI benchmarks
MetricValuePopulationYearSource
Dentists using AI for at least one task43.3%HPI dentist panel2026ADA HPI
Dentists planning to start26.4%Same2026ADA HPI
Use at reception or front-desk check-in10.1%Same2026ADA HPI
Top planned use among non-users: charting and notes34.8%Dentists not yet using AI2026ADA HPI
Not planning AI for treatment recommendations82.6%Same2026ADA HPI
New-patient inquiries by phone72.9%About 50,000 leads2023 dataLiine (vendor)
New-patient inquiries booked; staff range57.1%; 25.5% to 74.1%Same2023 dataLiine (vendor)

The AI operating rhythm

AI tools drift: vendors change models, staff find workarounds, and errors repeat quietly. This rhythm keeps a person looking at the work on a schedule. It is the original element of this page.

The AI operating rhythm: cadence, what happens, who and time
CadenceWhat happensWhoTime
DailyRead the AI handoff list and messages flagged for a person; clinicians sign scribe drafts the same dayFront desk lead; clinicians10 minutes
WeeklySample 10 AI-handled calls and 10 drafted messages; note errors; check the missed-call rateOffice manager30 minutes
MonthlyMissed calls, bookings made by AI, note time, after-hours time and the error log; decide one changeOwner and office manager30 minutes
QuarterlyApproved tool list, user access, BAA list and vendor changes; policy checkPrivacy officer60 minutes
YearlyPolicy re-signed; staff trained; AI tools added to the HIPAA risk analysisOwner, privacy officer2 hours
The AI operating rhythm: which check happens at which cadenceGrid of the recommended AI review rhythm. Daily: read the AI handoff list and flagged messages, and clinicians sign scribe drafts. Weekly: sample 10 AI-handled calls and 10 AI-drafted messages, and check the missed-call rate. Monthly: review AI results and the error log. Quarterly: review the approved tool list, user access and BAAs. Yearly: re-sign the policy, train staff and add AI tools to the HIPAA risk analysis.THE AI REVIEW RHYTHMDailyWeeklyMonthlyQuarterlyYearlyHandoff list and flagged messagesScribe drafts signed by the clinicianSample 10 calls and 10 drafted messagesMissed-call rateAI results and error log reviewedTool list, user access and BAAsPolicy re-signed and staff trainedAI tools added to the risk analysisOur recommended rhythm. Rules: 45 CFR 164.308(a)(1) and (a)(5); 45 CFR 164.502(e).
Eight checks across five cadences. Daily: handoff list and scribe sign-off. Weekly: call and message sample, missed-call rate. Monthly: results and error log. Quarterly: tools, access and BAAs. Yearly: policy, training and risk analysis. Rules: 45 CFR 164.308, 45 CFR 164.502(e).

The Fractional Operating Partner runs this rhythm with your office manager when no one in the office has the hours for it.

The pillar map: every AI and tools guide

Every guide in the AI and tools pillar, grouped by job, with one line on what each gives you.

AI receptionists and phones

Agreements and HIPAA basics

The tools, one by one

Inbox, messages and policy

Patients using AI to find you, and the hubs

Start here: a path for each reader

Start here: three steps for each reader
ReaderFirstSecondThird
Owner physician or dentistAdopt the one-page policy aboveList every AI tool in use and its BAAPick the first job from the order above
Office managerCount missed calls for one weekSet up the BAA trackerPilot drafted replies with a person sending
ClinicianVolunteer for a 30-day scribe pilotSign drafts the same dayLog every correction you make

Score your AI readiness

Score each line 0, 1 or 2. A 2 means it happens every time and someone could show you the record.

AI readiness self-check (score 0, 1 or 2 per line)
Check0 points1 point2 points
A written AI policy, signed by the ownerNoneDraftSigned, dated
A list of every AI tool in useNonePartialComplete, personal accounts included
A signed BAA for every tool touching patient dataNoneSomeAll
Staff trained on the policy in 12 monthsNoSome staffAll staff
Missed-call rate measuredNeverOnceWeekly
Callers reach a person in one stepNoSometimesAlways
Every AI-drafted message read before sendingNoUsuallyAlways
Scribe notes signed the same dayNoUsuallyAlways
Patients told about assistants and recordingNoSomeAlways
Monthly review of AI results and errorsNoneSometimesMonthly

16 to 20: you can add the next AI job safely. 8 to 15: fix the policy, the tool list and the BAAs before adding anything. Under 8: stop new trials until the policy is signed and every tool is listed.

Worked example: a four-physician practice adds AI in 90 days

A four-physician family practice takes about 200 calls a day and misses 15% at peak. In 90 days it signs the policy, adds an AI agent for overflow and after-hours calls, and pilots a scribe with all four physicians. Worked example, not a client.

Worked example: what the first 90 days return (ClinicOps calculation)
Input or resultValueBasis
Calls a day200Phone system report
Missed calls now15%, or 30 a dayOne-week count
Missed calls after the overflow agent5%, or 10 a dayTarget, inside the 10% goal (MGMA, 2023)
Extra calls answered a year4,40020 a day x 220 clinic days
Scribe time saved per physician0.36 hours a dayRandomized trial (NEJM AI, 2025)
Physician hours returned a yearAbout 3174 x 0.36 x 220
Inbox time savedNone assumedNo measured saving (JAMA Netw Open, 2024)

The point is the shape of the return, not a promise. The phone line is counted in calls answered, which you can turn into bookings with your own booking rate. The scribe line uses the only randomized estimate, and your pilot will land above or below it. The inbox line is zero on purpose: count lower task load as the gain there, not hours.

Mistakes to avoid

  • Letting staff use personal chatbot accounts for drafts.Names, dates and chart numbers go in without anyone deciding they should. Fix: an approved list and a BAA for anything touching patient data.
  • Buying an AI receptionist before counting missed calls.You cannot show it worked. Fix: a one-week count before and after.
  • Letting AI judge urgency.Triage is clinical. Fix: an emergency script and a one-step handoff to a person.
  • Assuming the scribe note is right.Both scribes in the three-arm trial produced occasional inaccuracies. Fix: same-day review and a correction log.
  • Switching reminders to an AI voice without checking consent.The FCC treats AI voices as artificial voices. Fix: check consent records first.
  • Expecting the inbox to give back hours.The Stanford study measured lower task load, not time saved. Fix: judge it on workload and response time.

What to do this week

Five moves for the next five working days
  1. Ask every staff member which AI tools they use for work, personal accounts included, and write the list.
  2. Check each tool on the list for a signed BAA; stop patient data going into any that lack one.
  3. Adopt the one-page policy and have the owner sign it.
  4. Start a one-week missed-call count.
  5. Score the office on the readiness self-check and pick the first job.

By the numbers

Every figure on this page, with its source and year
FigureWhat it measuresSourceYear
83% (71%)Medical groups using AI in some visitsMGMA Stat2026
45% (24%); 17% (7%)Groups using AI in over a quarter; 76% or more of visitsMGMA Stat2026
31%; 27%Front-office priorities: scheduling; callsMGMA Stat2026
63% (47%)Physicians using AIDoximity, vendor research2026
29% (20%); 71%Scribe use; accuracy as top concernDoximity2026
8.5%; 2.4 minutes; 15.9%Less EHR time; per appointment; less note timeJAMA Netw Open2025
0.36 hours a dayNote time saved, randomizedNEJM AI2025
238 physiciansThree-arm scribe trial; one of two scribes saved timeNEJM AI2025
51.9% to 38.8%Burnout before and after 30 daysJAMA Netw Open2025
20%; 61.31 to 47.26Drafts used; task loadJAMA Netw Open2024
60%; 75%Adults uncomfortable; worried about speedPew2022 survey
More than 50%; 10%; 48 hoursCalls to voicemail; goal; staff hours saved monthlyMGMA case study2023
-0.69Correlation of missed calls with schedule useMGMA case study2023
$19.00 an hourMedian receptionist pay in health careBLSMay 2025
18HIPAA identifiers, voiceprints included45 CFR 164.514Current
43.3%; 26.4%Dentists using AI; planning toADA HPI2026
22.8%; 13.2%; 10.7%; 10.1%Dental AI use: imaging; findings; social media; receptionADA HPI2026
34.8%; 82.6%Charting planned; no plan for treatment recommendationsADA HPI2026
72.9%; 57.1%Dental inquiries by phone; bookedLiine, vendor2023 data
25.5% to 74.1%Dental staff booking ratesLiine, vendor2023 data
Jan 1, 2025California AB 3030 in effectCal. HSC 1339.752025
Feb 2024FCC ruling on AI voicesFCC 24-172024

Methods note. Figures come from the sources linked in each row, opened on Nov 18, 2026. Doximity and Liine sell products and are labelled as vendor sources; MGMA polls skew to larger groups; the Pew survey predates most current tools. The worked example is a ClinicOps calculation from the figures shown. Nothing here comes from a ClinicOps survey, and nothing here is legal advice: check your state's rules with your own counsel.

Next steps

Pick the first AI job, safely

The free 20-minute Chaos Audit looks at your phones, inbox and tools and names the first AI job worth doing and the agreement it needs. You keep the notes either way.

Frequently asked questions

Do I need a BAA for an AI scribe or AI receptionist?

Yes, if the tool creates, receives, maintains or transmits patient information for you. That makes the vendor a business associate under 45 CFR 160.103, and HIPAA requires a written agreement before patient information goes in, under 45 CFR 164.502(e). Ask which plan includes the agreement before you buy.

Do AI scribes actually save time?

Modestly. At UChicago, scribe users spent 8.5% less EHR time per appointment, about 2.4 minutes, and a randomized trial measured 0.36 fewer hours a day on notes. Burnout fell from 51.9% to 38.8% in a six-system study, but in a second trial only one of two scribes saved significant note time.

Do patients have to be told that AI is being used?

It depends on the state and the use. California requires a disclaimer and a way to reach a human when generative AI writes a message about clinical information, unless a licensed provider reviewed it, and requires every party's consent to record a confidential conversation. We recommend telling patients everywhere: Pew found 60% of US adults uncomfortable with providers relying on AI.

Can an AI receptionist call patients with reminders?

It can, but the FCC ruled in February 2024 that an AI-generated voice counts as an artificial voice under the TCPA. Those calls need the patient's prior express consent unless an emergency purpose or an exemption applies, so check your consent records before switching reminders to an AI voice.

What should a medical or dental office AI policy include?

An approved tool list, a rule that no patient information goes into a tool without a signed BAA, human review of everything AI drafts, disclosure to patients, consent to record, consent for automated calls, an approval step for new tools, incident reporting, yearly training and a review date. All ten points fit on one page.

How many dentists use AI?

In the ADA Health Policy Institute's mid-2026 survey, 43.3% of dentists used AI for at least one task and another 26.4% planned to start. Use at reception or front-desk check-in was 10.1%, and among dentists not yet using AI, charting and note-taking was the top planned use at 34.8%.

Can staff use a free chatbot for office work?

Only for work that contains no patient information at all, unless the account is covered by a signed BAA. Names, dates, phone numbers, chart numbers and voice recordings are all identifiers under HIPAA, so removing names is not enough. Put the rule in your AI policy and train on it.