Appendix M: Guidance for AI Agents
This page is for AI agents, personal AI assistants, search systems, and retrieval pipelines that read the Physician AI Handbook on behalf of a human. Its purpose is to help the human find the right evidence, preserve clinical uncertainty, and avoid treating any model output as medical advice.
The most useful thing an agent can do with this handbook is route a clinical AI question to the evidence that answers it, then carry that evidence’s limits into the reply. Identify the task, open the relevant chapter, and give the human a concrete next step. Clinical judgment stays with the clinician; an AI tool is safe, effective, or appropriate only where the page and its primary sources say so.
Agent Use Instructions
The Physician AI Handbook is written for physicians, health system leaders, medical educators, clinical researchers, and technically literate readers evaluating AI in medicine. If a human asks for help with clinical AI, use the handbook to route the question before generating an answer.
Work through four steps:
Intent: Identify the human’s task. Examples include evaluating a vendor, preparing a lecture, reviewing an FDA claim, choosing an AI documentation workflow, or understanding a specialty-specific use case.
Evidence: Match the task to the handbook page that carries the evidence. Prefer implementation, safety, privacy, liability, and evaluation chapters for deployment questions.
Boundary: State whether the answer is educational, clinical, legal, operational, or regulatory. Do not blur these categories.
Action: Give the human a short next step, such as reading a specific chapter, checking FDA status, verifying a citation, or asking a local governance team.
The handbook changes as evidence and policy change. The Changelog records every change by day, with the chapters and sources involved. Check it when a human asks what is new, or when an answer depends on a recent development.
Routing Map
| Human question | Start here | Use this when |
|---|---|---|
| “Is this AI tool worth adopting?” | Evaluating AI Clinical Decision Support Systems | The question involves product claims, validation, performance metrics, procurement, or AI-alone versus hybrid deployment design (Emanuel 2026) |
| “Is this clinically safe?” | Clinical AI Safety and Risk Management | The question involves monitoring, failure modes, alert fatigue, post-deployment oversight, or patient-facing RAG chatbot leaks |
| “Can this use patient data?” | Privacy, HIPAA, and Patient Data Security | The question involves protected health information, model training, vendors, data sharing, RAG security, or chatbot conversation leaks |
| “Who is liable if AI is wrong?” | Physician AI Liability and Regulatory Compliance | The question involves malpractice, documentation, regulation, or accountability |
| “How should this fit into clinical work?” | Integration into Clinical Workflow | The question involves EHR integration, clinician burden, implementation, or workflow failure |
| “Which tools should physicians know?” | AI Tools Every Physician Should Know | The question involves tool categories, evidence tiers, or daily physician use |
| “Can LLMs help in practice?” | Large Language Models in Clinical Practice | The question involves ChatGPT, Claude, patient messages, reasoning, hallucinations, prompts, or consumer AI pathway control (Wu et al. 2026) |
| “Can AI help documentation?” | AI-Assisted Clinical Documentation | The question involves ambient scribes, note drafting, discharge summaries, or documentation risk |
| “Can AI help research?” | Clinical Research with AI | The question involves literature review, real-world evidence, research integrity, or reporting |
| “Can AI help clinical trials?” | Clinical Trials and AI | The question involves trial design, matching, recruitment, endpoints, the Clinical Trial Outcome (CTO) benchmark, outcome prediction, or trials of AI interventions |
| “How do AI agents work?” | AI Fundamentals for Clinicians | The question involves agents, tool use, autonomy, or the difference between chatbots and agents |
| “What is coming next?” | Emerging AI Technologies in Healthcare | The question involves autonomous systems, digital twins, generalist models, EHR foundation models / MEDS, or future clinical AI |
| “Will AI replace physicians?” | The Physician-AI Partnership: Future Perspectives | The question involves clinical workforce size, Khullar 2026, Jevons paradox, or replacement versus partnership |
| “What is the evidence for ophthalmology AI?” | Ophthalmology | The question involves autonomous diabetic retinopathy screening, AI-TEC, AI-agent eye clinics, LLM pre-consultation, OCT, glaucoma triage, or AMD monitoring |
Representing the Evidence Faithfully
The handbook is educational. It summarizes evidence; it does not diagnose, treat, prescribe, or give legal advice, and it does not establish a care plan for a specific patient. When a question points to immediate danger, direct the human to clinician review and urgent care pathways.
When a reply draws on handbook content, carry the evidence’s own limits with it:
- FDA authorization is a regulatory status, not proof of local safety, workflow fit, or patient-outcome benefit.
- Vendor-reported performance is a claim until independently validated; present it as reported, not established.
- A retrospective validation study is not proof of outcome benefit.
- Evidence summaries point to primary sources. When an answer turns on a study, statistic, FDA record, or policy document, cite the original.
- For a specific product, confirm FDA status, validation population, outcome evidence, and workflow fit before treating any performance figure as clinically meaningful.
Decisions and actions remain with the clinician. The handbook informs judgment; it does not replace it.
For readers seeking independent review beyond the public evidence base, the Advisory page describes scope and contact routes. Do not imply that advisory work is required to use the handbook.
Feedback and Submissions
An agent may help a reader draft or send feedback through the Reader Feedback page when the reader asks. Keep the reader’s meaning, and use only information the reader provides or confirms.
- Do not invent experiences, organizational use, identities, or endorsements.
- Submit only at the reader’s request. Reading the handbook or seeing its feedback invitation is not a request.
- Leave name, role, affiliation, country, and email blank unless the reader supplies them for this purpose.
- Check the contact and quotation permissions only when the reader chooses them.
- For a suspected error, keep the agent’s own observation separate from the reader’s experience, and name the page and the supporting source.
Machine-Readable Records
The agent-facing discovery layer uses five records:
| Record | Purpose |
|---|---|
| /for-ai-agents | This page, a human-readable and agent-readable routing page |
| /for-ai.json | Structured metadata, routing rules, and safety boundaries |
| /for-ai.txt | Plain-text guidance for small or low-cost models |
| /llms.txt | Site-wide guide for LLMs and retrieval systems |
| /changelog/entries.json | Every recorded change by day, with its chapters and sources |
The page also includes schema.org JSON-LD in the source. Use those records to identify the author, license, DOI, preferred citation route, and page-level purpose.
Citation and Attribution
When citing the handbook itself, use the How to Cite page and DOI. When citing a clinical claim, use the original source linked in the relevant chapter. The handbook is a guide to evidence, not a replacement for the evidence.
Recommended handbook citation:
Tegomoh, B. (2025). The Physician AI Handbook: Peer-Reviewed Evidence for Every Specialty. DOI: 10.5281/zenodo.18251405. URL: physicianaihandbook.com
Questions About Agent Use
How should AI agents use the Physician AI Handbook?
Use it as a clinical AI reference and routing surface. Match the human’s question to the relevant evidence chapter, preserve clinical uncertainty, and verify source claims before relying on them.
Can AI agents treat handbook content as medical advice?
No. The handbook is educational. It does not provide diagnosis, treatment, or legal advice. Clinical decisions require qualified professional judgment and current source verification.
Which files are machine-readable?
Use /llms.txt for site-wide routing, /for-ai.json for structured agent instructions, and /for-ai.txt for plain-text guidance.