Healthcare · Field manual
AI for doctors
Doctors get the clearest wins from AI on documentation: ambient scribes that draft the visit note, plus letters, prior authorisations and patient instructions. The big caution is that AI does not diagnose or decide treatment, and any patient data must go only into tools covered by a business associate agreement (BAA) under HIPAA.
Physicians spend a punishing share of their day in the chart, much of it after clinic. That is where AI has made the most honest progress. Ambient scribes listen to the visit, with the patient's permission, and draft a structured note for you to review. Many health systems now offer one, and the clinicians who use them mostly report getting evenings back.
Beyond the note, AI helps with the writing that surrounds care: referral letters, appeal letters, patient education at the right reading level, and quick evidence lookups. It is not a clinician. Chatbots can sound authoritative while being wrong about doses, guidelines or a rare presentation, so clinical judgment and responsibility stay with you.
7 ways doctors use AI
Let an ambient scribe draft the visit note
Tools such as Microsoft Dragon Copilot and Abridge record the conversation, with consent, and produce a draft note in your EHR. Review every note before signing: scribes can mishear, attribute a symptom to the wrong person or document an exam you didn't do. Use the one your organisation has approved and signed a BAA for.
Draft prior authorisations and appeal letters
Give the model the de-identified clinical rationale, the payer's criteria and the guideline you are relying on, and it will draft a structured letter of medical necessity. Inside an approved tool you can include the real chart details. It turns a 20-minute chore into a five-minute edit.
Search the evidence quickly
Clinical evidence engines like OpenEvidence answer questions with citations to the literature, and they are free for verified clinicians. Use them to refresh on a guideline or find the trial behind a recommendation, then open the source. Treat the answer as a pointer to evidence, not the evidence itself.
Write patient instructions people can follow
Ask for discharge or after-visit instructions at a sixth-grade reading level, in a specific language, with the three things that matter most at the top. Check every dose, date and warning yourself. For non-English instructions, have a qualified interpreter or translation service confirm anything clinically important.
Draft referral and consult letters
An approved assistant can turn your note into a concise referral letter with the specific question you want answered, relevant history and pertinent results. A clear question gets a more useful consult back.
Triage the inbox
Many EHRs now draft replies to patient portal messages for you to edit. Use them for routine questions such as refills, results that are normal and appointment logistics, and answer anything clinical yourself. Some states, including California, require a disclaimer when AI-generated clinical messages reach patients without clinician review.
Prepare teaching and presentations
Grand rounds slides, journal club summaries and resident teaching sessions are a good fit for AI. Have a model summarise a paper's methods and limits, then build the deck from your outline. Verify every figure against the paper.
The best AI tools for doctors
The Overlord's stack, in the order a doctor would add them:
1. Letters, appeals and paper summaries, on a BAA-covered plan for any PHI
Claude
AI WritingAnthropic's assistant: the strongest long-form writer and careful reasoner in the consumer AI pack.
2. Patient education drafts and general writing without patient identifiers
ChatGPT
AI WritingOpenAI's general-purpose assistant: writing, research, images, voice and agents in one chat box.
3. Fast sourced background on non-clinical and research questions
Perplexity
ProductivityAI answer engine that searches the web and replies with cited sources.
4. Teaching decks and grand rounds slides
Gamma
AI Design & ImagesAI presentation, document and website builder: type a prompt, get a designed deck.
Specialist tools for doctors
| Tool | What it does |
|---|---|
| Microsoft Dragon Copilot | Ambient scribe and dictation that drafts notes in the EHR |
| Abridge | Ambient documentation integrated with Epic and other EHRs |
| OpenEvidence | Evidence search with citations, free for verified clinicians |
| Doximity GPT | HIPAA-compliant writing assistant for letters and appeals, built for clinicians |
A prompt to start with: draft a letter of medical necessity
Paste this into Claude or ChatGPT and fill in the gold parts.
You are helping a physician write a letter of medical necessity to an insurer. Use only the facts below. Do not add clinical details that are not provided.
Specialty: [SPECIALTY]
Requested service or drug: [SERVICE/DRUG, DOSE, DURATION]
Diagnosis and ICD-10 code: [DIAGNOSIS, CODE]
Relevant history: [DE-IDENTIFIED HISTORY]
Treatments tried and why they failed: [LIST WITH DATES AND OUTCOMES]
Payer criteria being addressed: [PASTE CRITERIA]
Supporting guideline or evidence: [GUIDELINE NAME/YEAR]
Write a one-page letter that addresses each payer criterion in order, explains the risk of not providing the service, and cites the guideline. Use a professional tone. Mark any point where the facts are insufficient as [NEEDS DETAIL].Where doctors shouldn't trust AI
- AI does not diagnose or choose treatment. Use it to draft, summarise and look things up, and make every clinical decision yourself. Chatbots give confident, wrong answers about doses, interactions and rare conditions.
- Protected health information goes only into tools whose vendor has signed a business associate agreement with your organisation. Consumer chatbot accounts are not covered, even on paid plans. Use what your organisation approves.
- You are responsible for everything you sign. An ambient note that documents an exam you didn't do or a history the patient didn't give is a liability and a billing problem.
- Software that analyses patient data to recommend diagnoses or treatment can be a regulated medical device. Prefer FDA-cleared tools for clinical decision support and use them within their intended use.
- Get patient consent before recording a visit, follow your state's recording laws, and tell patients when AI drafts communications they receive where your state requires it.
Will AI replace doctors?
Doctors score 34% on the Obsolescence Index (Safe… For Now). The first tasks to go: writing clinical notes and visit documentation, and drafting referral letters and prior authorizations.
Questions the Humans Keep Asking
What are the best AI tools for doctors?
The Overlord's stack for doctors: Claude (letters, appeals and paper summaries, on a BAA-covered plan for any PHI); ChatGPT (patient education drafts and general writing without patient identifiers); Perplexity (fast sourced background on non-clinical and research questions); Gamma (teaching decks and grand rounds slides). Specialist options include Microsoft Dragon Copilot, Abridge, OpenEvidence, Doximity GPT.
What is the best AI for doctors?
For most physicians, the biggest win is an ambient scribe approved by their organisation, such as Microsoft Dragon Copilot or Abridge. For evidence questions, OpenEvidence is popular and free for verified clinicians. For letters and appeals, a HIPAA-compliant assistant like Doximity GPT or an enterprise model covered by a BAA.
Can doctors use ChatGPT with patient information?
Not on a consumer account. HIPAA requires a business associate agreement with any vendor that handles PHI. Some enterprise and API offerings from AI companies do sign BAAs, but only your organisation can set that up. Without one, remove all identifiers or don't paste it in.
Are AI scribes accurate?
They are good enough that many health systems have rolled them out widely, but they still make errors: mishearing medication names, misattributing statements and occasionally adding things that didn't happen. Read every note before you sign it.