AI 101 for Physicians, From Chatbots to Agents
Most of us are already using AI, whether we planned to or not. In the AMA's 2026 survey, 81% of physicians said they use AI in their practices, up from 38% in 2023, mostly to summarize research and help with documentation. The vocabulary has moved fast, though. Here is a plain-language guide to what the terms mean, what the tools do, and what they cost.
What We Mean by AI
When people say AI today, they usually mean a large language model. These systems are trained on enormous amounts of text and learn to predict what words should come next. That simple idea turns out to be powerful. A good model can summarize a 40-page discharge packet, draft a patient letter, or explain a new guideline in a few seconds.
It can also be confidently wrong. Models don't look facts up the way a search engine does unless they're built to, and they can produce citations, doses or details that sound right and aren't. Dr. Seidelman's vitamin D example in this issue is a good illustration. The output is a draft from a fast, well-read assistant, not a second opinion.
The Models You'll Hear About
Most general-purpose tools come from a handful of companies. ChatGPT is made by OpenAI, Claude by Anthropic, Gemini by Google, and Copilot by Microsoft, which is built into Microsoft 365 and Outlook. Each company offers several versions. The lighter, faster ones handle quick questions and everyday writing. The more advanced ones take longer and do better at multi-step reasoning, long documents and data analysis. The specific version names change every few months, so the company matters more than the number after it.
Then there are tools built for medicine. OpenEvidence, for example, answers clinical questions from the medical literature and cites its sources. AI is also increasingly built into the software you already use, including EHRs, imaging platforms and billing systems, often without a separate login.
From Chatbots to Agents
A chatbot answers when you ask. An agent goes a step further and takes actions on its own to finish a task. It might read an incoming fax, pull the right chart, fill out a form, place a phone call and report back when it's done. The model is the engine. The agent is the engine connected to your systems and given a job.
In physician practices, agents are showing up in a few places:
- Ambient scribes listen to the visit, with patient consent, and draft the note for you to review and sign.
- Voice agents answer calls and book appointments, including after hours, and send reminders.
- Revenue cycle agents verify insurance eligibility and call payers to check on stalled prior authorizations.
- Inbox tools draft replies to patient portal messages for a clinician to edit.
Insurers use AI too. Since October 2025, Maryland law has limited how carriers can use AI in utilization review, requiring decisions to rest on the individual patient's clinical information and keeping those systems open to state audit.
What It Costs
General-purpose chatbots have free versions, and the standard paid plans for ChatGPT, Claude and Gemini run about $20 a month. Business plans, which add administrative controls, are about $20 to $25 per user per month, with premium tiers reaching $100 to $200.
Clinical tools cost more. Ambient scribes aimed at small practices start around $40 to $120 per clinician per month, while enterprise products like Abridge and Nuance DAX Copilot can run several hundred dollars per provider per month and often require an annual contract. Voice and revenue cycle agents are usually priced per call, per minute or by contract, so ask for a quote based on your actual call volume.
Protecting Patient Information
This is where most of the risk sits. Any tool that touches protected health information needs a signed business associate agreement. Consumer chatbot accounts, including the free and $20 plans, generally don't come with one, and the law firm Holland & Hart notes that without a BAA, disclosing PHI to an AI vendor is generally impermissible. Keep patient details out of personal accounts, and confirm that any clinical product you buy will sign a BAA for the specific plan you're on.
Where to Start
Pick one real problem, like after-hours calls or charting time, rather than shopping for AI in general. Pilot one tool with a few clinicians, check its output closely, and read the contract's liability terms before you sign. The AMA's new guidance on liability for clinical AI, covered in this issue, is a helpful lens, and PEAKE's November 18 webinar on evaluating AI tools walks through the vendor process step by step.
MCMS will keep following how AI is changing the practice of medicine in Montgomery County. If you're using a tool that's working well, or one that isn't, we'd like to hear about it. Contact Marissa Valeri at mvaleri@montgomerymedicine.org.
