Applied AI in medicine: what it is, how it is used, and how to evaluate it
A clinician's guide to applied AI: the types used in clinical practice, why evidence grading and citation checks decide whether an AI answer can be trusted, where the safety risks are, what regulators look at, and the questions to ask before relying on any tool.
Published by Hormonaly · Last updated: · Clinical leadership
For licensed clinicians: Educational information for licensed clinicians, not medical advice for patients. Any dosing requires prescriber supervision and laboratory monitoring.
What is applied AI in medicine?
Applied AI in medicine is artificial intelligence put to work on a defined clinical or operational task, with known inputs, a reviewable output and a person accountable for the result. The definition matters because it separates tools you can evaluate from tools you cannot. A general-purpose chatbot asked an open question can produce fluent text about almost anything; an applied system is built for one job, states what it used, and hands its output to a clinician who decides what to do with it.
In practice that means software that does something specific: retrieves and grades the evidence behind a clinical question, drafts a note, prepares a prescription for a prescriber to review, screens a regimen for interactions, or builds training material that faculty approve. Each of these can be tested against a standard, and each leaves a record of what it did.
Types of applied AI in clinical practice
Applied AI in clinical settings falls into a small number of task families. They differ in what goes wrong when they fail, and so in how much review they need.
- Evidence retrieval and synthesis: answering a clinical question from the published literature, with the sources shown and the certainty of the evidence stated.
- Clinical documentation: drafting notes and summaries from an encounter or a record for the clinician to edit and sign.
- Prescribing support: preparing a prescription with dose, interaction screening and regulatory status attached, for a prescriber to review, sign and send.
- Safety screening: checking a regimen for interactions, contraindications and dose ranges before it reaches a patient.
- Medical education: building courses, cases and assessments from an institution's own material, approved by faculty before release.
- Interpretation of images and signals: reading imaging, ECGs or other signals. Hormonaly does not build tools in this family; it is listed so the map is complete.
- Operations: scheduling, intake, coding and other administrative work that sits around clinical care.
Helix API: evidence retrieval and synthesis · E-Prescribing Agents
Agentic AI and chatbots: what is the difference?
A chatbot answers in one pass from what its model already contains. An agentic system breaks a task into steps and gives each step to a specialized agent: one retrieves literature, another checks that each citation exists and says what it is quoted for, another screens doses and interactions, and the result is assembled only after those checks pass. The steps can be logged, tested and audited one by one.
Agentic design does not ensure quality on its own. What makes it useful in medicine is that each step has a defined input, a check and an owner, and that anything which would change clinical guidance is held for a clinician's sign-off.
Evidence quality: why grading matters
An answer is only as good as the evidence under it, and clinical evidence varies enormously in strength. The GRADE framework (Grading of Recommendations, Assessment, Development and Evaluation) is the method major clinical guidelines use to rate the certainty of evidence; the Endocrine Society guideline on testosterone therapy and the American College of Physicians guideline on age-related low testosterone, for example, were both developed with it.
Hormonaly grades every answer and protocol on the same principle, from A (high certainty) to D (very low certainty), and grades evidence limited to animal or laboratory studies as D. A grade does not make a claim true or false; it tells the clinician how much confidence the evidence can bear before acting on it.
Citations are the other half. Hormonaly verifies citations twice, when evidence is retrieved and again after the answer is written, and does not list retracted papers as references. Our published benchmark measures how often clinical language models cite a real paper for a claim that paper does not support, because a reference that exists is not the same as a reference that is right.
Safety: where AI in medicine goes wrong
The failures that matter most in clinical AI are specific and predictable, which is what makes them preventable.
- Unsupported or misattributed citations: a real paper cited for something it does not say.
- Stale evidence: an answer that was right before a new trial, safety communication or regulatory action.
- Scope creep: a tool built to summarize evidence that starts to sound like it is making a treatment decision.
- Dose and interaction errors: a plausible number that is wrong for the formulation, route or patient.
- Overconfidence: weak or animal-only evidence presented in the same voice as a guideline recommendation.
How Hormonaly guards against those failures
Every product runs on one governed core. Before anything is released it passes safety checkpoints for doses, citations, interactions and claim scope. New research, FDA actions and trial updates are scanned nightly, every protocol is re-audited at least weekly, and anything that would change clinical guidance waits for a physician's sign-off. E-Prescribing Agents prepare prescriptions; the prescriber reviews, signs and sends them. Enterprise workloads run on HIPAA and SOC 2 compliant infrastructure, with PHI under a signed BAA.
Regulatory notes, at a high level
Whether a piece of clinical software is regulated, and how, depends on what it is intended to do and where it is used. Software intended to diagnose, treat or make a decision about an individual patient is treated differently from software that helps a clinician find and weigh evidence while leaving the decision to them. Ask any vendor what its intended use is, and check the position of your own regulator.
Hormonaly is an evidence synthesis platform for licensed professionals. It is not a medical device, and it does not diagnose, prescribe or make treatment decisions. Its protocol pages state each compound's regulatory status, including when a use is off-label or a compound is not approved, because that status belongs in front of the clinician alongside the evidence.
How clinicians should evaluate an AI tool
Before relying on any AI tool in practice, ask it, or its vendor, these questions.
- Intended use: what task is it built for, and what does it say it does not do?
- Sources: where does the evidence come from, and can you open every reference it cites?
- Grading: does it tell you how certain the evidence is, and does it grade animal-only evidence differently from trials in people?
- Currency: how often is the evidence refreshed, and how are retractions and safety communications handled?
- Human review: who signs off on changes to clinical content, and who signs the prescription?
- Failure behavior: when the evidence is insufficient, does it say so rather than filling the gap?
- Data protection: where is patient data processed, is there a BAA, and is your data used to train a shared model?
- Auditability: can you see what it did, step by step, after the fact?
Frequently asked questions
What is the difference between applied AI and general AI chatbots in medicine?
Applied AI is built for one defined clinical task, shows what it used and hands its output to an accountable clinician. A general chatbot answers open questions from what its model contains, without a defined task, a stated evidence base or a review step.
Is AI clinical decision support regulated?
It depends on the software's intended use and the jurisdiction. Software intended to diagnose or treat an individual patient is treated differently from software that helps clinicians find and weigh evidence. Hormonaly is an evidence synthesis platform for licensed professionals, not a medical device.
How can I tell whether an AI tool's citations are reliable?
Open them. Check that each reference exists, that it says what the tool claims it says, and that it has not been retracted. Prefer tools that grade the certainty of evidence and verify citations after the answer is written, not only when evidence is retrieved. How Hormonaly grades evidence
Does AI replace clinical judgment?
No. Applied AI can retrieve, grade and organize evidence and draft documents, but the licensed clinician makes the decision. At Hormonaly, anything that would change clinical guidance waits for a physician's sign-off, and the prescriber reviews and signs every prescription.
What does GRADE mean in an AI answer?
GRADE rates how certain the evidence behind a statement is. Hormonaly shows it from A (high certainty) to D (very low certainty) and grades evidence limited to animal or laboratory studies as D. Research and methodology
References
- Testosterone Therapy in Men With Hypogonadism: An Endocrine Society Clinical Practice Guideline. J Clin Endocrinol Metab. 2018. PMID: 29562364
- Testosterone Treatment in Adult Men With Age-Related Low Testosterone: A Clinical Guideline From the American College of Physicians. Ann Intern Med. 2020. PMID: 31905405