What separates an agent from a chatbot
A chatbot answers. An agent decides what to do next. The difference is not the model — it is whether the system has tools, state and a goal.
| Capability | What it adds |
|---|---|
| Tool use | The system can call an API, query a database or write to another system, rather than only producing text. |
| Planning | The system decomposes a goal into steps and chooses their order, instead of following a fixed script. |
| State and memory | The system carries context across steps and sessions, so a long-running task survives interruption. |
| Evaluation and retry | The system checks its own intermediate results and can retry or escalate when a step fails. |
Every one of these expands the compliance surface as well as the capability. An agent that can call tools can reach data the original request never named, which is why tool scoping and logging matter more in agentic systems than in conversational ones. See HIPAA-compliant generative AI for the control set this requires.
What is working in production today
The pattern that works is narrow scope, rich instrumentation, and a clear human checkpoint. These are agent categories running inside health systems on the InferAgents platform now.
Revenue cycle and billing
Denial appeals are an unusually good fit: the policy is documented, the payer's requirements are specific, the source record is structured, and a human already reviews the output before submission. PayerAgents drafts payer-specific appeals against the clause a denial turned on, and checks CPT coding against current policy, with a billing specialist approving before anything leaves the health system.
Clinical knowledge and protocol support
Specialty agents surface guidance grounded in the health system's own protocols rather than generic literature. The Infectious Disease Agent is the clearest demonstration that the core generalises: the workflow is very different from billing, but the platform underneath is identical.
EHR-adjacent guidance
The Epic Tip Sheet Agent surfaces the right tip sheet inline, next to the work rather than in a separate portal. It sits alongside the system of record, never replacing it — a deliberate design constraint, because writing back into the EHR raises the stakes of an error by an order of magnitude.
Operational and supply chain
Reconciliation, exception triage and vendor-record cleanup are high-volume, low-ambiguity tasks with an existing audit trail. They are frequently the fastest route to a demonstrable result inside a health system, because nothing about them touches care delivery.
Where the autonomy boundary sits
Being precise about this is a commercial advantage, not a caveat. Health systems have been sold autonomy that did not exist, and buyers now discount vague claims heavily.
| Appropriate for agent autonomy | Requires a human decision |
|---|---|
| Drafting a document a human will review and sign | Any diagnosis, or a recommendation presented as one |
| Retrieving, summarising and citing internal guidance | Treatment selection, dosing, or changes to a care plan |
| Checking a claim against documented payer policy | Final submission of a claim or appeal to a payer |
| Flagging an exception or inconsistency for review | Closing or dismissing a clinical exception |
| Reading from the EHR under scoped access | Writing to the patient record |
If an error would reach a patient before a qualified human sees it, the step is not a candidate for autonomy. Everything else is a question of instrumentation and confidence, not principle.
Governing agents once you have more than a few
The first agent is a project. The fortieth is an estate. Health systems that do not plan for the second case end up with agents nobody owns, running against models nobody has reviewed, on data nobody has scoped.
- Named ownership. Every agent has an accountable owner in the health system, not only in the vendor.
- Scoped access. Each agent reaches the minimum data and tools its function requires, configured centrally rather than in the agent's own code.
- Version and change control. Agents move through the same design, test, stage and ship lifecycle as any other clinical software.
- Model review. The underlying model is a reviewable, swappable choice — see multi-model AI for why this matters at renewal time.
- Continuous audit. Prompts, retrievals, tool calls and outputs are logged in one place across all agents, not per application.
This is the reason InferAgents exists as a platform rather than a set of applications: the governance layer is the hard part, and it does not scale if it is rebuilt for every department. Our full position on oversight and permissions is on the AI governance page.
Evidence from deployment
More than 80 agents run on the InferAgents platform across clinical, revenue cycle, operational and IT-adjacent functions, serving over 2,500 active users and processing more than 200,000 queries since the June 2025 production launch. The platform is live inside University of Chicago Medicine and Mayo Clinic.
Figures are platform-wide totals across all deployments as of May 2026, measured from the June 2025 production launch. They describe adoption and usage, not clinical outcomes; we do not claim outcome effects that have not been independently studied.
Frequently asked questions
What is agentic AI in healthcare?
Agentic AI in healthcare refers to AI systems that pursue a goal across multiple steps — planning what to do, calling tools and systems to gather information or take action, and checking their own intermediate results — rather than answering a single question. In health systems it is used for bounded tasks such as payer appeal drafting, CPT coding checks, surfacing internal protocol guidance and operational reconciliation.
How is an AI agent different from a chatbot?
A chatbot produces a response to a prompt. An agent has tools, state and a goal: it can call APIs, query systems, carry context across steps and retry or escalate when something fails. That extra capability also expands the compliance surface, because an agent can reach systems the original request never mentioned.
Can AI agents make clinical decisions autonomously?
No, and responsible deployments do not attempt it. Agents can draft, retrieve, summarise, check and flag; a qualified human remains accountable for diagnosis, treatment selection and any change to a care plan. A practical rule is that if an error would reach a patient before a clinician sees it, the step should not be autonomous.
What healthcare workflows are agents best suited to today?
Tasks that are bounded, well documented, already instrumented, and already reviewed by a human before they take effect. Revenue cycle appeals and coding checks, protocol and tip-sheet retrieval, documentation triage, and supply chain reconciliation are the categories currently working reliably in production.
How many agents can a health system realistically run?
More than most expect, provided governance is centralised. Over 80 agents run on the InferAgents platform today across clinical, billing, operational and IT functions. The limiting factor is rarely model capability — it is whether access scoping, change control, audit and ownership are handled by the platform rather than rebuilt per agent.