AI for Doctors.
AI will not replace your clinical judgment, your examination, or your accountability. It is already replacing the two hours of documentation after clinic, the fourth rewrite of a prior-auth letter, and the discharge instructions nobody has time to translate into plain language. This handbook is the practical middle ground: where AI genuinely helps clinical work, the three rules that keep patients safe, and five workflows you can use this week — no technical background required.
Where AI actually helps — and where it doesn't
The useful mental model: treat AI like a fast, tireless medical scribe crossed with a well-read student — excellent at producing drafts and summaries from what you give it, confidently wrong just often enough that nothing it says about a patient goes unverified. You'd gladly let it draft your note; you would never let it decide a dose.
| Task | Verdict | Why |
|---|---|---|
| Drafting clinical notes from your dictation or a visit transcript | Excellent | The most mature clinical AI use case — ambient scribes are deployed at scale |
| Patient-friendly explanations & discharge instructions | Excellent | Translation of your verified plan into plain language |
| Administrative drafting — prior auths, referrals, appeals | Excellent | Formulaic prose nobody enjoys writing; you review and sign |
| Summarizing papers or guidelines you provide | Strong | Grounded in text you gave it — far less room to invent |
| Differential brainstorm — "what am I missing?" | Useful, verify | A hypothesis generator to check against real references — never a diagnostician |
| Drug doses, interactions, contraindications — unverified | Never | Models state wrong doses with total confidence — see Rule 2 |
| Patient-identifiable data in consumer tools | Never | PHI requires a BAA-covered tool — see Rule 1 |
The three non-negotiable rules
Rule 1 — Patient-identifiable information never goes into a consumer AI tool.
Protected health information handled by a third-party tool requires a business associate agreement (BAA) under HIPAA — and consumer chatbots don't sign them. Names, dates of birth, MRNs, and rare-condition-plus-demographics combinations that identify by inference: none of it. The discipline: de-identify fully before using general tools ("a 60-year-old with atrial fibrillation on apixaban," not the patient), or use healthcare tools your institution has deployed under a BAA. When in doubt, your compliance office exists for exactly this question.
Rule 2 — Anything clinical gets verified against a real reference. Anything.
Language models generate plausible text, not verified medicine. They will state a wrong dose in a perfect sentence, invent a contraindication, or cite a study that doesn't exist — with identical confidence to when they're right. The discipline: any dose, interaction, diagnostic criterion, or citation that originated from an AI gets checked in UpToDate, the package insert, PubMed, or your institutional references before it touches a decision. The AI's role is to draft and to remind — the reference's role is to be right.
Rule 3 — The chart is yours. The advice is yours.
When you sign a note, you attest to it — "the AI drafted it" changes nothing about that. AI-drafted documentation gets the same review you'd give a human scribe's draft: read it fully, correct what's wrong, delete what didn't happen. The time savings are real and large; they come from skipping composition, not skipping review. An unread AI note with a hallucinated exam finding in it is your hallucinated exam finding.
Five workflows you can use this week
Each recipe: what to give the AI, what to ask, and what you must verify. Prompts are starting points — keep your de-identification discipline throughout.
1 · Visit note from your dictation
- Dictate or type the raw facts — unstructured is fine, that's the point.
- Ask for structure:
2 · Discharge instructions a patient can actually follow
- Start from your verified plan, then translate:
3 · Prior authorization & appeal letters
- Give the clinical justification and the denial reason:
4 · The "what am I missing?" differential check
- Use it as a checklist generator after you've formed your own impression:
5 · Literature summary — of papers you provide
- Give it the actual paper or abstract, not a topic:
The judgment exercise: spot the danger
Three scenarios from real practice patterns. Pick what you'd do.
1. Clinic ran long. You want to paste this afternoon's charts — names, DOBs, MRNs included — into a free chatbot to draft your notes. It would save an hour. Do you?
Rule 1. A consumer tool without a BAA handling names + MRNs + clinical detail is a HIPAA problem regardless of your account settings, and deleting the chat doesn't undo provider-side retention. This exact convenience shortcut is why institutions deploy BAA-covered scribe tools — use those, or strip identifiers completely.
2. Drafting instructions for a pediatric patient, the AI writes "amoxicillin 90 mg/kg/day divided BID." It sounds like the high-dose regimen you vaguely remember. Ship it?
Rule 2. The dose may well be right — high-dose amoxicillin is a real regimen — but "sounds right + AI said it" is exactly how a wrong dose gets co-signed by two unreliable sources. And asking the model to check itself just samples the same distribution again. Thirty seconds in the reference settles it with authority.
3. The ambient scribe drafted a beautiful note for a 10-minute visit — including "cardiovascular: regular rate and rhythm, no murmurs." You never auscultated. Sign it?
Rule 3. Signing attests you did what the note says. A documented exam that never happened is a falsified record — "the AI added it" is not a defense, and in an audit or a lawsuit that line is indistinguishable from your claim. The scribe saves you composition time; review time is the part you keep.
Evaluating tools: the questions that matter
You don't need to understand transformers to procure clinical AI well. You need written answers to six questions:
| Question | Answer you want |
|---|---|
| Will you sign a BAA? | Yes — non-negotiable for anything touching PHI |
| Is our data used to train your models? | No, contractually |
| Retention and deletion policy? | Defined, short, deletable on request |
| For clinical-decision features: regulatory status? | Clear answer on FDA clearance or why it's out of scope |
| EHR integration and audit logging? | Yes — you want a record of what was generated and edited |
| Published accuracy data for our specialty? | Real evaluations, not marketing claims |
Ambient scribes are the most proven category and the sanest place for a practice to start — the workflow (draft → physician review → sign) has the safety review built in. Patient-facing chatbots and diagnostic aids carry far higher stakes and deserve proportionally harder questions. And if your institution already has approved tools, that list beats anything an article can tell you.
Quick answers
Is using AI in clinical practice even allowed?
Broadly yes, within HIPAA and institutional policy — ambient scribes are in routine use at major health systems. The bright lines are patient data handling (BAA), verification of clinical content, and your attestation on the record.
AI passes medical licensing exams now — doesn't that make it reliable?
Exam benchmarks measure recall on well-posed questions, not messy real patients, incomplete histories, or accountability. Impressive scores coexist with confident dosing errors. The exam result is why it's a useful drafting partner; the error mode is why Rule 2 exists.
Will AI replace physicians?
It's replacing tasks — documentation, translation, drafting, triage support — not examination, judgment, procedures, or responsibility. The realistic shift: clinicians who use it well recover hours per week, and the human parts of the job concentrate.
Where should a skeptical physician start?
Recipe 2 — patient-friendly translation of a plan you've already verified. Zero clinical risk from the model (the medicine is yours), zero PHI risk if de-identified, and the quality of the translation is genuinely startling the first time.
Related on Vibe Engines
Explore the topic
See this alongside everything else on the same subject — handbooks, system designs, challenges and tools, in one place.
More Handbooks
- The Prompting HandbookA friendly, hands-on field guide for everyday humans — learn the CRISP framework, spot bad prompts, practice with real recipes, play a drag-and-drop game, and test yourself with a quiz. No code required.Read →
- The Agentic AI Interview HandbookTwenty topics every senior AI engineer should be able to reason about live — from eval pipelines to reliability patterns for generative systems.Read →
- The Senior AI Engineer Interview Handbook60 questions across architecture, production incidents, agentic systems, RAG, evals, cost, safety, and leadership — what staff-level AI interviewers actually probe for.Read →
- 50 Angular Interview QuestionsA visual handbook covering components, change detection, RxJS, signals, routing, forms, performance, and testing — what interviewers actually probe for in senior Angular roles.Read →
- 50 Python Interview QuestionsFundamentals to advanced: data structures, OOP, iterators & generators, the GIL, asyncio, memory, testing, and the standard library — a visual walk through everything a Python interview touches.Read →
- 51 LLM Evals Interview QuestionsGolden sets, LLM-as-judge, regression testing, offline vs online evals, RAG evals, agent evals, red-teaming, and observability — demystified for interviews and production.Read →
Explore more from Vibe Engines
Get the next one in your inbox.
New handbooks, system-design walkthroughs, and tools — straight to your inbox. No spam, unsubscribe anytime.