AI Solutions
What is agentic AI in healthcare?
Short answer
Agentic AI is software that pursues a goal by planning steps, using tools such as record searches, forms and databases, and checking its own progress. In healthcare it can assemble prior-auth packets, draft documentation and screen research candidates, while clinical and coverage decisions stay with qualified people.
- Author
- Kumbirai Muhwati
- Published
- Last reviewed
- Reading time
- 4 min
Agentic AI, generative AI and rules-based automation
- Rules-based automation follows fixed if-then logic written in advance. It is predictable, but it breaks when inputs vary, such as a fax in an unexpected layout.
- Generative AI produces content in response to a prompt: a summary, a letter or an answer. One request, one output.
- Agentic AI uses a model to decide what to do next. It breaks a goal into steps, calls tools, reads the results and adjusts, until the task is done or it needs a person.
How an agent works
- Goal: the task it is given, such as “prepare the prior authorization request for this MRI order”.
- Plan: the steps it decides on, such as finding the order, the relevant notes and the payer’s criteria.
- Tools: the systems it may use, such as an EHR search, a payer criteria lookup or a document generator.
- Memory: what it keeps between steps, such as the documents it has already read.
- Permissions: exactly which tools and data it may touch, and which actions need a person’s approval.
Tools are increasingly connected through the Model Context Protocol (MCP), an open standard that lets an AI application discover and call tools and data sources through one common interface. In plain terms, MCP is a universal adapter between an agent and the systems it may use. It makes connecting tools easier, which makes permissions more important: an agent can do whatever its connected tools allow.
Where agents help today
- Prior authorization: gathering the relevant notes, labs and imaging reports, checking them against the payer’s published criteria and drafting the request for staff to review and submit.
- Clinical documentation: drafting discharge summaries, referral letters and appeal letters for a clinician to edit and sign.
- Care-gap outreach: finding patients overdue for screenings and drafting outreach for the care team to approve.
- Cohort screening: checking candidates against trial eligibility criteria and explaining why each one meets or fails them.
- Literature review: searching and summarizing published evidence, with citations a researcher can verify.
Where people must stay in charge
Agents should not make clinical or coverage determinations on their own, and regulators agree. In a February 2024 FAQ on its 2023 Medicare Advantage rule (CMS-4201-F), CMS said plans may use algorithms and AI to help make coverage decisions, but each decision must be based on the individual patient’s circumstances, such as their medical history, the physician’s recommendations and clinical notes, and not on a larger data set alone. The tool also may not change the plan’s publicly posted coverage criteria.
The practical rule: agents gather, check, draft and route. Licensed clinicians and reviewers decide.
Guardrails that make agents safe to deploy
- Tool allowlists: each agent gets only the tools and data scopes its task needs, read-only by default.
- Human sign-off: high-impact actions, such as submitting a request, writing to the chart or contacting a patient, need a person to confirm them.
- Audit logs: every step and tool call is recorded with its inputs and outputs, so any result can be traced and reviewed.
- Evaluation on your own cases: test on real examples from your workflow, including edge cases, failure modes and performance across patient groups, before go-live.
- Monitoring in production: track overrides, errors and incidents, and feed them back into evaluation.
- PHI protection: run in a HIPAA-compliant environment, with de-identified data used for any training by default.
Regulation to watch
- FDA: AI that meets the definition of a medical device is regulated as device software. FDA finalized guidance on predetermined change control plans for AI-enabled device software functions in December 2024, and issued draft guidance on lifecycle management and marketing submissions in January 2025. Administrative agents are generally not devices; clinical decision tools can be. Intended use decides.
- CMS: the Medicare Advantage requirements above, and the prior authorization decision timeframes and APIs in CMS-0057-F.
- EU AI Act: AI in medical devices and some health-related uses is high-risk. After the Digital Omnibus on AI, those obligations apply from 2 December 2027 for stand-alone (Annex III) systems and 2 August 2028 for AI in regulated products such as medical devices.
Frequently asked questions
- Is agentic AI the same as a chatbot?
- No. A chatbot answers in a conversation. An agent takes steps in other systems, such as searching records or filling in forms, to complete a task.
- Can an AI agent deny a prior authorization?
- It shouldn’t. For Medicare Advantage, CMS requires coverage decisions to be based on the individual patient’s circumstances. Good practice is for agents to gather and draft, and for licensed reviewers to decide.
- What is MCP?
- The Model Context Protocol is an open standard for connecting AI applications to tools and data sources through a common interface. It makes tools easy to connect, so tool permissions need careful design.
- Does the FDA regulate agentic AI?
- Only when the software meets the definition of a medical device, for example when it is intended to diagnose or treat. Administrative agents generally are not devices, but the intended use decides.
Sources
- 42 CFR 422.101, Medicare Advantage coverage requirements (eCFR)
- AAMC summary of the CMS February 2024 FAQ on AI in Medicare Advantage
- FDA draft guidance: AI-enabled device software functions (January 2025)
- FDA guidance: predetermined change control plans for AI-enabled devices (December 2024)
- Model Context Protocol
- Regulation (EU) 2024/1689, the AI Act (EUR-Lex)
General information, not legal or medical advice.