How to start using AI in a medical practice: five steps, biggest drain first

Straight answer: start with an ambient scribe. An ambient scribe is software that listens to the visit and drafts the note so you review and sign instead of type. It hits the drain most doctors feel first: charting after hours. Heidi Health has a free plan and Freed publishes pricing from $39/mo. Before any patient data goes in, get a signed BAA. The rest of this page is the five-step plan.

Step 1. Pick the biggest pain

Where does your practice actually lose time? Be honest and pick one thing, not a wish list. For most solo doctors it is the note finished at 9 p.m. instead of at dinner. For others it is a phone that rings during a visit, prior authorizations that take a staffer half a day, or no-shows that leave empty slots. Prior authorization, or prior auth, is the approval an insurer requires before certain care or medications.

Put a rough number on it. Minutes charting per day. Calls that go to voicemail per week. Hours spent on one prior auth. The drain you pick decides the tool. A demo should never decide it for you.

Step 2. Pick one tool, and get a signed BAA first

Match one tool to that one drain. If the drain is documentation, start with an ambient scribe. The lowest-cost, lowest-risk options are Freed, which publishes pricing from $39/mo, and Heidi Health, which has a free plan you can test before you pay. The bigger scribing names, Abridge, Microsoft Dragon Copilot (formerly Nuance DAX Copilot), and Suki AI, are quote-based, so you get a number after talking with the vendor.

If the drain is the phone or the front desk, that is a later step, not day one. When you get there, Klara (now sold as ModMed Patient Engagement, powered by Klara) handles secure two-way patient messaging, Notable automates intake and prior auth, and Hyro is an AI voice agent that answers routine calls. Those are quote-based and aimed at larger practices. CodaMetrix automates coding for high-volume groups.

This is the step you cannot skip: get the signed BAA in writing and have your compliance or IT lead review the vendor's data-handling and model-training terms before any patient information reaches the tool. Purpose-built medical tools sign a BAA and state they do not train their models on your patient data. If a vendor will not sign one, do not use it with patient data, and never put patient information into a general consumer AI that has not signed a BAA.

Step 3. Run a 30 to 60 day pilot

Turn the tool on for a subset of visits or patients and keep your old process running alongside for 30 to 60 days. That is long enough to see a real change without betting the whole practice on an unproven tool. Track the one number you named in Step 1: charting minutes per day, no-show rate, calls answered, or prior auths cleared.

Keep a clinician on every clinical output during the pilot. The AI drafts the note; you read it, fix it, and sign it. That review is also what makes the pilot honest, because you see exactly what the tool got right and where it needed a hand.

Step 4. Train the team

A short, hands-on start beats a long manual. Show the physician and any staff how to edit a drafted note, how to correct the AI so it learns your style, and exactly where the human check sits. Tune one note template to your specialty so the drafts start closer to how you already write.

Write down one rule and post it: no AI note, code, or patient message leaves the practice without a person checking it first. That single line keeps the tool a helper, not a decision-maker, and keeps responsibility where it belongs.

Step 5. Measure, then expand or switch

At the end of the pilot, look at the number. If your after-hours charting time dropped, keep the scribe and pick the next drain, such as patient messaging or scheduling and reminders to cut no-shows. Add one job at a time so each change stays measurable. If the number did not move, change the tool rather than give up on the idea. You spent a month and a small subscription learning something true, which is cheap.

What AI can and cannot take off your plate

AI handles drafts and admin work, not diagnosis. An ambient scribe writes the first version of the note. A messaging tool answers routine refill or directions questions so the phone rings less. A prior-auth tool assembles the request from the chart for a person to check. A voice agent books a routine visit at 8 p.m. and routes an urgent caller to a human.

What AI does not do: it does not decide the diagnosis, choose the treatment, or sign the note for you. The physician stays responsible for every clinical output. The value is time back, not judgment replaced. For a fuller list of the tools and where each one fits, see the AI Help for Doctors home page, or the broader directory entry for AI for medical practices.