AI Agent for a Solo Physician Practice: Refills and Referrals
Which refill, referral and inbox tasks a solo physician can delegate to an AI agent, and which clinical decisions stay with the doctor.

A solo physician can hand three kinds of routine work to an AI agent: preparing refill requests for the doctor's approval, tracking referrals until the specialist's report returns to the chart, and sorting the inbox into clear categories with a firm escalation path for anything clinical. The agent prepares, tracks and routes. The physician keeps every diagnosis, every prescription decision, every judgment about urgency and every signature.
Key takeaways
- Refills are a queue to prepare, not a decision to make: the agent gathers the request and the physician approves, changes or declines it.
- Referrals are lost at the handoffs, so the agent should track each one from order to returned report and flag what is missing.
- Inbox triage works when the agent sorts and drafts only administrative replies, and any message describing symptoms goes to a human path.
- Set permissions before tasks: read access to the chart, write access only to a draft queue, and no authority to send clinical content or sign anything.
- Start with one role and one checkable success criterion, then expand only after recurring reviews confirm the routing is reliable.
Where a solo practice loses time in refills and referrals
A solo practice absorbs administrative work personally. Refill requests arrive through the patient portal, phone messages, faxes and pharmacy calls, often with incomplete details. A referral is a chain of steps: the order, any authorization the payer requires, scheduling with the patient, the specialist's report and the result filed back into the record. Most of the delay happens between those steps, not inside the clinical decision itself.
The usual failure points look like this:
- A refill request arrives in three places and gets answered from memory, without checking the current medication list.
- A referral is sent, but nobody confirms that the specialist office received it or that the patient has a date.
- A specialist report arrives by fax or attachment and sits unfiled until the next visit.
- Administrative messages and symptom messages share one inbox, so the urgent ones wait behind scheduling questions.
An agent is useful here because these failures are about tracking and routing. They do not require the agent to know medicine. They require it to notice what is open, what is missing and who needs to act.
Refill requests: what the agent prepares and what the physician approves
A refill workflow is a preparation pipeline with a single approval point. The physician should define in writing which requests may be approved on a routine basis and which always require review. The agent then works through each request in a fixed order:
- Collect every refill request from the portal, inbox, fax and pharmacy messages into one queue, recording the patient, medication, strength, pharmacy and date of the last visit.
- Compare the request against the active medication list in the chart and flag any mismatch, such as a changed strength, a discontinued drug or a duplicate prescription.
- Attach the relevant excerpt from the last note and any monitoring the physician has tied to that medication, such as a lab that is due.
- Prepare a proposed response: approve as written, request a visit, or hold for review. The proposal stays in the queue and is not sent.
- Route the item to the physician with the reason for the proposed action, so the decision takes a few seconds rather than a search through the chart.
The physician's written protocol decides which items can be pre-approved. A routine renewal of a stable medication, where the physician has already set the rule, is a reasonable candidate. A new medication, a dose change, a controlled substance, an abnormal recent result, or a patient the physician has not seen within the interval they defined is not. The agent does not decide that a refill is safe. It decides only which bucket the request belongs in.
Referral tracking: from order to returned report
Referral tracking is the strongest case for delegation because it is mostly follow-up. The agent records each order with the specialty, the receiving office, the urgency the physician selected, the patient's preferred scheduling window and any authorization status. It then watches for the events that should happen and raises a flag when they do not.
Illustrative: A physician orders a physical therapy evaluation on a Monday. The agent records the order, the receiving clinic's contact route and the patient's preferred mornings. A week later, with no appointment confirmed, it drafts a reminder for the patient and a status request for the clinic, and places both in the physician's queue for approval. When the evaluation note arrives, the agent files it against the referral and flags any recommendation that asks for a clinical decision.
A referral tracker should show, at minimum:
- The order date, the urgency the physician chose and the receiving office.
- Whether the patient has been contacted and whether an appointment is confirmed.
- Whether any authorization or supporting document is required and whether it has been sent.
- Whether the specialist's report has returned, and where it was filed.
- Which open referrals have passed the interval the physician set for follow-up.
For a practice that sends many referrals to rehabilitation or other therapy services, the physical therapy workflow described in Superintelligent Agents for Physical Therapy: Referrals and Visit Notes shows a similar pattern of tracking handoffs without changing the clinical plan.
Inbox triage: the sorting rule and the escalation rule
Inbox triage should begin with categories the physician can read at a glance. A practical set is administrative (billing questions, forms, address changes), refill request, appointment, result question, clinical symptom, and unclear. The agent assigns one category to each message and records the reason.
The escalation rule matters more than the sorting. The agent may draft replies only for administrative messages, using templates the physician has approved. Any message that describes symptoms, a medication side effect, worsening, distress or a possible emergency goes to the clinical path immediately, and the patient receives the emergency instructions the physician has written for that situation. A symptom message must never wait in a low-priority queue, and the agent must never be the one that judges a symptom as minor.
Inbound text is also untrusted input. A patient message, a faxed document or a pasted referral letter can contain text that is written to steer the model, whether by accident or on purpose. The OWASP guidance on prompt injection prevention describes why such content should be treated as data to classify, not as instructions to follow. In practice, the agent's routing rules should live in the physician's protocol, not in whatever a message asks it to do.
For practices that have a behavioral health component or a separate intake process, the limits described in AI for Behavioral Health Clinics: Intake and Scheduling Limits are a useful reference for deciding which messages must never be answered by automation.
Permissions, untrusted input and the approval boundary
The design principle is least privilege. The agent should read only what its role requires, write only to a draft or queue location, and have no ability to send clinical content, place an order, sign a prescription or alter a diagnosis. Authorization decisions should be explicit, reviewed on a schedule and logged. The OWASP authorization guidance is a practical starting point for checking that each action is permitted for each role, not just for the system as a whole.
The NIST AI Risk Management Framework offers a useful vocabulary for this work: define the context, identify the risks, measure how the system behaves and manage the residual risk with named owners. For a solo physician, the owner is the physician. Every boundary below should be written down before the agent is connected to any live system:
- Which decisions the physician approves in every case, without exception.
- Which routine decisions the physician has pre-approved, and the exact conditions that must hold.
- Which messages trigger immediate escalation and what the patient is told in the meantime.
- What the agent may read, what it may write to, and what it may never touch.
- Who reviews the audit log, how often, and what prompts a change in permissions.
- How privacy obligations are handled, including the rules that apply under local law such as Brazil's LGPD or the rules that govern the practice's jurisdiction.
The last point should be checked with the practice's own advisers. An agent design can support compliance, but it does not replace it.
Starting with one role and a checkable success criterion
Be Human's managed approach starts with one useful role rather than a full practice system. For a solo physician, that is usually refill queue preparation or referral tracking. The first engagement defines the memory the agent may consult, the tools it may call, the approval gates a physician must pass, and one observable success criterion, such as the proportion of routed items that the physician agrees were sorted correctly in a recurring sample. Operations after launch can include monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled changes. The service is delivered remotely, in English and Portuguese, and it is built so the physician stays in control of every clinical decision.
A practical first step is to write down the last twenty refill requests and the last ten referrals, then mark which steps the physician performed personally and which were only tracking or copying. The steps in the second group are the candidates. Choose one, define the approval point and the escalation rule, and test it against real history before any live use.
If you are not sure which of your workflows should go first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Can an AI agent send refill approvals to a pharmacy without the physician? In the design described here, no. The agent prepares a proposed response and places it in the physician's queue. Sending, signing and any change to the prescription remain with the physician, and routine approvals happen only under a protocol the physician has written and reviewed. Pre-approval rules should be narrow, and anything outside them should go to review.
Is it safe to let an agent sort patient messages? Sorting and drafting administrative replies can be safe when the categories are clear and the escalation rule is strict. Any message that describes symptoms, side effects or possible emergencies must go to a human path immediately. The agent should not judge whether a symptom is minor, and its routing should be checked on a recurring sample so that errors are caught early.
What does a solo physician need before starting? You need a written protocol that states which refills can be pre-approved, which messages must escalate, what the agent may read and write, and who reviews the audit log. You also need a clear success criterion and a way to check it. Privacy and regulatory obligations should be reviewed with your own advisers before any system is connected to patient data.
Does an agent replace the electronic record or the practice's compliance work? No. The agent works alongside the record and the existing compliance processes. It can track, prepare and route work that the record already holds, but it does not become the source of clinical truth or a substitute for the practice's obligations. Its value depends on clear boundaries, regular review and the physician keeping authority over every clinical decision.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


