AI agents for clinics and healthcare practices
Clinics use AI agents on administrative load rather than clinical decisions: patient intake, insurance eligibility checks, prior authorisation packets, referral coordination and claim denial triage. Every one of these touches protected health information, so the architecture has to establish a BAA-covered processing path before anything ships. Clinical judgement stays with clinicians.
Start with the paperwork, not the medicine
The administrative burden in a small practice is the problem worth solving. Prior authorisation alone consumes a substantial share of clinical staff time, and it is almost entirely form-filling, document assembly and telephone follow-up — repetitive, high-volume, and requiring context from several systems at once. That is a precise description of what agents are good at.
Clinical decision support is a different product with a different regulatory posture. An agent that drafts a prior auth packet is an administrative tool. An agent that suggests a diagnosis may be a regulated medical device. Keep the line bright, and build on the administrative side of it.
HIPAA comes before the workflow design
Any agent touching patient data is handling PHI, which means the model provider is a business associate and you need a BAA in place with them. Both major providers offer this on their enterprise tiers; consumer API tiers generally do not carry one. Sending PHI to an endpoint with no BAA is a reportable disclosure.
Beyond the BAA: minimum necessary applies to prompts, so send the smallest slice of the record that answers the question rather than the whole chart. Log every access at the field level. Zero data retention should be contractual, not assumed. This architecture is not optional decoration around the agent — it is the first thing to build.
Where agents earn their keep in a practice
| Workflow | Typical burden | Agent scope |
|---|---|---|
| Prior authorisation | Hours per week of staff time, high denial-on-technicality rate | Assembles the packet from the chart, checks it against the payer's published criteria, flags what is missing before submission |
| Eligibility and benefits | Repeated per appointment, often on hold | Queries the clearinghouse, reconciles the response against what the patient reported, surfaces mismatches at booking rather than at check-in |
| Referral coordination | Chased manually, frequently dropped | Tracks the loop from order to consult note received, escalates the ones that go quiet |
| Claim denial triage | Denials sorted by hand, many never reworked | Classifies the denial reason, routes the reworkable ones with the correction already drafted |
| Appointment no-shows | Revenue lost, slots wasted | Predicts likely no-shows from history, runs the confirmation sequence, offers the slot onward when a cancellation lands |
| Clinical documentation | — | Different product, different regulatory question. Out of scope here. |
The integration reality: EHRs are hard
This is where healthcare agent projects actually stall. Epic, athenahealth, eClinicalWorks, NextGen and Cerner all support FHIR to some degree, and US regulation has pushed real progress on read access. Write access is another matter, and vendor-specific app review processes can add months to a timeline before a single call is made in production.
Plan for it. Scope the first agent around what you can read plus what a human can commit, and treat deeper write integration as a later phase with its own schedule. A project that assumes day-one write access to the EHR is a project that will be late.
A prior authorisation agent, traced
Detects the trigger
An order is placed for a service the patient's plan requires authorisation for. The agent picks this up from the order rather than waiting for someone to notice.
Retrieves the payer's criteria
Looks up the current medical policy for that plan and procedure — these change, and stale criteria are a common denial cause.
Assembles evidence from the chart
Pulls only the elements the criteria call for: the relevant diagnoses, the conservative treatment already tried, the specific imaging. Minimum necessary, by design.
Identifies the gap
If the criteria require six weeks of documented physical therapy and the chart shows four, it says so now — before submission, while it can still be fixed.
Queues for clinical sign-off
A clinician reviews and submits. The agent never asserts medical necessity on its own behalf.