More on bringing AI into your practice
Guides, a case study and checklists for using AI in a medical, dental or veterinary practice, safely.
AI in your practice
Where AI earns its place in a physician's practice today, ranked by how strong the evidence is and how safe the use is. Every number is linked to the study or agency it comes from.
By Sohel Siddique, who builds Lydoh and VoiceboxMD. Updated November 4, 2026.
The short answer
Start with paperwork and phones, not diagnosis. In order of evidence and safety: ambient AI scribes for visit notes, AI on the phones and front desk, draft replies for the patient inbox, help with prior authorizations and coding, and scheduling and reminders. Wait on anything that makes clinical decisions without a clinician checking it, and on any tool that touches patient information without a Business Associate Agreement.
Most physicians are already using something. In the AMA's 2026 survey, 81% of physicians said they use AI in their practices, up from 38% in 2023, and 86% called data privacy critical to wider adoption (AMA). The question for a small practice isn't whether, it's which jobs to hand over first and how to check the work.
Sources checked November 4, 2026. General information, not medical or legal advice.
Ranked by evidence and safety
What the published evidence shows for each job, what can go wrong, and our verdict for a small practice.
| Use | Verdict | What the evidence shows | What can go wrong |
|---|---|---|---|
| Visit notes (ambient AI scribes) | Start now | Two randomized trials: one scribe cut time in notes 9.5%, the other showed no significant change (UCLA, NEJM AI); about 0.36 fewer hours a day on notes and lower exhaustion (UW Health, NEJM AI). Burnout 51.9% to 38.8% in 30 days, no control group (JAMA Netw Open). | Occasional clinically significant errors in drafts. The physician reads and signs every note. |
| Phones and front desk | Start now | Little peer-reviewed research yet; most data is from vendors, including us. Our case study prints a full month of one practice's calls, including what went wrong. | Low clinical risk if it never gives medical advice, sends urgent calls to a person and keeps a human one transfer away. |
| Inbox and patient messages (draft replies) | Pilot and measure | No time saved in two 2024 studies: reply time did not change, and read time rose 21.8% in one (Stanford, UC San Diego). Task load and exhaustion scores fell at Stanford. | A draft that sounds right and isn't. Every reply is sent by a person who read it. |
| Prior authorization and coding | Pilot and measure | The burden is large: 39 requests per physician a week, 13 hours of physician and staff time (AMA). Diagnostic billing codes improved with ambient AI in one trial (UW Health). | Codes that overstate the visit are a compliance problem. A person checks every code and submission. |
| Scheduling and reminders | Start now | Mature, rule-based tools; the AI part is understanding a patient's reply. Low stakes, easy to measure in no-shows and open slots. | Texts carry patient information. Keep them to the minimum, under a BAA. |
| Diagnosis or treatment decisions without a clinician check | Wait | Giving physicians a language model did not significantly improve diagnostic reasoning on vignettes, 76% vs 74% (JAMA Netw Open). | Wrong answers delivered with confidence, liability, and federal nondiscrimination duties for decision support tools. |
| Any tool with patient information and no BAA | Wait | Not a question of evidence. A vendor that handles patient information for you needs a Business Associate Agreement first; see which AI tools sign one. | A HIPAA disclosure the moment someone pastes a note into a personal account. |
1. Documentation
An ambient scribe listens to the visit, with the patient's permission, and drafts the note for the physician to edit and sign. It is the AI use with the most rigorous evidence behind it.
What it means for you:expect a scribe to make the day feel lighter before it makes it shorter. Use one in a version with a BAA, tell patients, and read every draft before you sign, especially medication names, doses and anything you didn't say out loud.
Disclosure:we make VoiceboxMD, an AI scribe and dictation tool and Lydoh's sister product (voiceboxmd.com); it includes a BAA with every subscription. Hold it to the same questions as any other scribe below. None of the studies on this page tested VoiceboxMD.
2. Phones and front desk
An AI receptionist answers the practice line, takes appointment and refill requests for staff, and sends urgent calls to a person. The published research here is thin; most of what exists is vendor data, including ours, so ask any vendor for real numbers and what went wrong.
Here is ours. In a one-doctor practice's first month with Lydoh, it answered 2,151 calls, about 470 a week. 18% came outside office hours, when nobody was in the office. A third were passed to the front desk on purpose, and staff marked 89% of the 843 requests Lydoh wrote up for them as handled. On the first full day, a fraud safety cap turned away 132 patient calls before we fixed it; the case study explains what happened and what changed.
Disclosure: we make Lydoh. Whoever you choose, start with after-hours and overflow calls, check what the AI hands to staff every day for the first weeks, and make sure a caller can always reach a person. Comparing options? See AI receptionists for medical practices.
3 to 5. Inbox, prior auth, scheduling
Patient messages. AI-drafted replies have not saved time in the studies so far. At Stanford, clinicians used the drafts for 20% of replies, with no change in reply, write or read time, but lower task load and exhaustion scores (Garcia et al., 2024). At UC San Diego, read time rose 21.8%, reply time did not significantly change, and replies got longer (Tai-Seale et al., 2024). Pilot it if inbox fatigue is the problem; don't buy it to save minutes.
Prior authorization and coding. Practices complete 39 prior authorization requests per physician per week, and physicians and staff spend 13 hours on them; 40% of physicians have staff who work only on prior authorizations (AMA, late-2024 survey). AI is good at drafting the letter and the appeal from a template and the payer's policy. The payer side is changing too: from January 1, 2026, Medicare Advantage, Medicaid, CHIP and marketplace plans must decide expedited requests within 72 hours and standard ones within 7 calendar days, with a prior authorization API required from January 1, 2027 (CMS). For coding, treat AI suggestions as suggestions: a person confirms every code.
Scheduling and reminders. The least glamorous and most dependable item on the list. Automated reminder texts and online rescheduling are mature; AI adds the ability to read a reply like "can I come Thursday instead". Measure no-shows and unfilled slots before and after. More on appointment reminders.
Where to wait
Diagnosis and treatment decisions. In a randomized trial, physicians given a large language model scored 76% on diagnostic reasoning against 74% with conventional resources, a difference that was not significant. The model alone scored 16 points higher than the physicians with conventional resources, which says more about how hard it is to use these tools well than about replacing anyone (Goh et al., JAMA Network Open, 2024). Use AI to look things up and summarize literature; keep the decision, and the responsibility, with the clinician. If your practice is covered by Section 1557 of the Affordable Care Act, federal rules also require reasonable efforts to find and reduce discrimination risk in patient care decision support tools (45 CFR 92.210).
Anything without a BAA. Consumer ChatGPT, Gemini and Claude on personal accounts don't sign one, so patient information can't go in. OpenAI's ChatGPT for Clinicians is free for verified US physicians, NPs, PAs and pharmacists and comes with an individual BAA, but it is designed for one person, not your staff, and OpenAI says it is "intended to support, not replace, professional judgment" (OpenAI). The full plan-by-plan answer is in Is ChatGPT HIPAA compliant?
Our take
In 2026, the AI worth paying for in a small physician practice is the AI that does paperwork and phone work, not the AI that thinks about the patient. We would change our minds if a randomized trial in real clinics, not vignettes, showed physicians using an AI diagnostic tool making fewer diagnostic errors than physicians without one. That trial hasn't been published yet.
And a test for any tool you try: if it doesn't move the one number you bought it for within 30 days of real use, it won't move it at 90. The six-system scribe study measured its burnout drop at 30 days (Olson et al.). Give a tool a month, then keep it or cancel it.
Adoption plan
One problem, one tool, one number. Written for a practice with one to five physicians and no IT department.
Days 1 to 14
Days 15 to 45
Days 46 to 90
Before you sign
The full checklist, with the HIPAA rule behind each item, is in our AI policy template and vendor checklist. For general chat tools, see which ChatGPT, Copilot, Gemini and Claude plans sign a BAA. Ask us the same questions.
Running a different kind of practice? Read AI for dentists, AI in veterinary practice, or practical AI in healthcare examples.
FAQ
Documentation. Ambient AI scribes, which listen to the visit and draft the note, have the strongest evidence of any AI a physician can buy today: two randomized trials published in NEJM AI in 2025 found less time on notes or less work exhaustion, and a six-system study in JAMA Network Open found burnout fell from 51.9% to 38.8% after 30 days. The gains are real but modest, and the physician still has to read and sign every note.
Some do, by less than the marketing suggests. In UCLA's randomized trial of 238 physicians, one scribe (Nabla) cut time in notes by 9.5% compared with usual care, and the other (Microsoft DAX Copilot) showed no significant change. In UW Health's trial, time on notes fell by about 0.36 hours a day. The more consistent finding across studies is lower exhaustion and task load.
Yes, for anything without patient information, and with patient information only in a version that comes with a Business Associate Agreement. OpenAI offers ChatGPT for Clinicians free to verified US physicians, NPs, PAs and pharmacists, with an individual BAA; it does not cover your staff. Consumer ChatGPT has no BAA. OpenAI itself says the clinician version is meant to support, not replace, professional judgment. Our ChatGPT HIPAA guide covers every plan.
The phrase covers different tools: an ambient scribe that drafts visit notes, an AI receptionist that answers the phone and takes requests, an inbox tool that drafts replies to patient messages, or a chat tool for clinical questions. Ask which job it does, whether it signs a BAA, and who checks its work before it reaches a chart or a patient. Those three answers matter more than the label.
Not without a clinician checking every answer. In a randomized trial in JAMA Network Open, giving physicians a large language model did not significantly improve their diagnostic reasoning on case vignettes (76% vs 74%). Use AI to look things up and summarize literature, and keep the decision with the physician.
It can draft the letters and appeals, which is where much of the staff time goes; the AMA's late-2024 survey found 39 prior authorization requests per physician per week and 13 hours of physician and staff time. Paste payer policies and templates freely, but send patient details only through a tool with a BAA, and have staff check every submission.
Tell every patient, and ask. An ambient scribe records the conversation, and some states require every party's consent to record. Check your state's law with your counsel, put a line in your intake paperwork, and let patients say no without friction. This page is general information, not legal advice.
It varies by tool and changes often, so get a written quote. More useful than the price: ask whether the BAA is included at every tier, whether there is a contract term, and whether you can export your data and leave. Then measure the one number the tool is supposed to move for 30 days before you commit.
Keep reading
Guides, a case study and checklists for using AI in a medical, dental or veterinary practice, safely.