The Map · safe use of AI in clinical practice

Can it be used for everything?

A chat open in a tab. A scribe listening to the visit. An assistant inside the record.

The tools are already in the exam room, and adoption has outrun our ability to evaluate them. The answer changes from task to task, and that is what this map shows.

Physicians reporting awareness or use of AI in practice · AMA, 2026181%
Best model on real clinical text tasks · BRIDGE, 2026244.8%

The same models clear 90 on knowledge exams2. An exam is not a clinic.

The five principles

Thirteen cards, five rules underneath

Every card here is one of these five principles applied to a specific task. If you read nothing else on this page, read these.

01What cannot be verified is not delegated.Delegability is not model capability. It is capability, plus the consequence of the error, plus your real ability to check it.
02An invisible gap becomes an explicit one."Not documented," "not assessed," "no prior value," "[confirm]." What is missing gets declared, never filled in silently.
03The tool receives context, not authority.The clinical question, the real patient, and the declared guideline go into the prompt. The decision does not.
04Verification follows risk, not perceived confidence.Self-reported confidence is not a reliable proxy for accuracy: in one radiology study, it did not correlate significantly with correctness. You check what is expensive to get wrong.
05Transformation can be delegated; judgment keeps an owner.Summarizing, structuring, and drafting transfer with review. Interpreting in context and deciding do not.

Safe use is not trusting everything or doubting everything. It is knowing what to check.

The clinical principles in this guide are jurisdiction-agnostic. Privacy, documentation, consent, and professional requirements vary by location and institution; where it matters, the regional specifics appear separately.

The map

One task, one safe role for AI

Each card opens its own page. Sort by delegation and the answer to the question above assembles itself; sort by moment of the day and notice how thin "during the visit" is.

foundation delegable draft · review required copilot · directed verification decision · not delegable outside the spectrumFinal authority: always yours.
01

Setting up your tool

Ten minutes that protect every task that follows.The riskWrong environment, and the rest of the map does not hold: the contract decides where your data goes.foundation02

Chart summarization

Compresses months of record into minutes of reading.The riskOmission is invisible: you do not miss what you never saw.copilot03

Interpreting test results

Organizes the panel and answers the question that was asked.The riskWithout the clinical question it weaves everything into one story and turns a visit into a screening program.copilot04

Ambient visit transcription

Listens to the visit and gives you your eyes back.The riskDocuments what nobody said and an exam that never happened. The signature is yours.delegable draft05

Diagnosis and differential

Broadens the differential and tests your hypothesis.The riskAgrees with your anchor instead of challenging it, and folds under pressure.copilot06

Clinical scores and calculations

Builds the score and interprets the result.The riskThe error enters before the arithmetic, in the criterion filled in by inference.copilot07

Guideline retrieval

Finds the recommendation and the passage in seconds.The riskReturns what talks about the guideline with the authority of the guideline.copilot08

Treatment planning

Lays out options and trade-offs for your decision.The riskCalibrated for the typical patient in the literature, blind to cost, access, and routine.decision09

Medication reconciliation

Cross-references lists and flags interactions for you to check.The riskA plausible wrong dose does not announce itself.copilot10

Referral letters and replies

Structures the draft that travels to a colleague.The riskIt slips through easily inside well-written formal prose.delegable draft11

Patient education and plain-language translation

Translates your plan into the patient's language.The riskText that reassures optimizes tone and eats the warning sign.delegable draft12

Sick notes, certificates, and formal reports

Drafts the document that leaves confidentiality.The riskWrites for whoever asked, not for whoever will read it.delegable draft13

Limits: images, ECGs, and tracings

The endpoint of the spectrum, and what makes it credible.The riskWhat cannot be verified is not delegated. It is referred.outside the spectrum
In construction

Four more on the way

These follow the same anatomy, and they publish as they are finished.

Patient messages

Drafting replies with the right tone, and catching the clinical advice that slips into an administrative answer.

Prior authorizations and appeals

Persuasive writing where every fact still has to be exactly right, in whichever payer system you work under.

Critical appraisal of literature

Reading a paper with the tool without letting it summarize away the thing that makes the paper weak.

Teaching and journal club

Preparing a session, and using the tool to build the questions rather than the answers.

What already exists

This map did not start in a vacuum

The neighbors, credited

Institutional governance toolkits, academic prompting tutorials for clinicians, safe-use recommendations for diagnostic AI, and courses focused on documentation address adjacent layers of the same problem.

What this page adds

It organizes the clinician's day by task, anchors every rule to a documented failure mode, distributes the safe role of AI across subtasks rather than across products, and keeps the endpoint visible: the one card where the answer is no.

Sources on this page
  1. [report] American Medical Association. Augmented Intelligence Research survey, 2026.
  2. [article] BRIDGE: benchmarking large language models for understanding real-world clinical practice texts. Nature Biomedical Engineering, 2026. doi.org/10.1038/s41551-026-01719-2
  3. [article] Huppertz MS, et al. Revolution or risk? Assessing the potential and challenges of GPT-4V in radiologic image interpretation. European Radiology, 2025;35:1111-1121. doi.org/10.1007/s00330-024-11115-6

Further reading · Menlo Ventures, The State of AI in Healthcare (2025): ambient scribe adoption estimated around 35% in large US systems. An industry report, useful for sizing adoption; it does not support a clinical claim.

Patients should have a right to an unhurried visit, thirty minutes or an hour when the case is complex. Clinicians should have a right to enough time to build the reasoning that produces good care. The tools on this map exist to give minutes back to that time. Not to compress it further.