Superintelligent Agents for Physical Therapy: Referrals and Visit Notes

An AI agent for a physical therapy clinic works best as a narrow role that tracks incoming referrals, flags missing paperwork, keeps the scheduling state of each patient aligned with the referral, and drafts visit notes for the treating therapist to review. It does not diagnose, change a treatment plan, or sign anything. The value comes from a clear memory of what each referral requires and from approval points that stop unverified information from reaching a chart, a payer or a patient.

Key takeaways

• Start with one queue, usually incoming referrals that are missing an element your clinic requires before a first visit, before touching visit notes.
• Let the agent draft notes from material the clinician provides, but keep clinical judgment and the signature with the treating therapist.
• Give the agent read access to referral and schedule records and write access only to drafts, task lists and approved message templates.
• Define a success criterion you can observe in your own records, such as how many referrals reach a first visit with nothing missing.
• Review the agent's memory and permissions on a fixed schedule, because staff, referring offices and payer rules change over time.

Where an agent fits in a therapy clinic

Clinics lose time in the gaps between steps. A referral arrives by fax, portal or email. Someone copies the referring provider's details into the system, coverage checks happen later, the first visit gets booked before the order is confirmed, and the last session's notes sit in a different folder. A well-scoped agent holds the state of each referral across those gaps, so a person can see at a glance what is finished and what is waiting. Its role is administrative memory and drafting, not clinical reasoning.

A practical split looks like this:

• Hand to the agent: intake of referral documents, checking required fields against your checklist, reminders to referring offices, drafting visit notes from clinician-provided material, and a daily list of open tasks per patient.
• Keep with clinicians: diagnosis, changes to the plan of care, judgments about medical necessity, discharge decisions, signing notes, and any message that changes care.
• Keep with front desk or billing staff: payer submissions, authorization decisions, account changes and refunds.

Clinics with intake and recall work may also find the chiropractic intake and recall example useful for mapping where reminders end and clinical decisions begin.

Referral tracking, step by step

Referral tracking usually fails at the start, when nobody has agreed on what a complete referral looks like. Settle these points before you connect any system:

• Write down the states your clinic uses, such as received, incomplete, verified, scheduled, first visit completed and closed. Keep the list short enough that staff remember it without checking.
• Write the required elements of a complete referral in plain language, such as the referring provider, the order, the diagnosis information the order provides, the visit count if your payer requires one, and a reliable contact route.
• Choose one intake source, such as the fax inbox or a portal export, and leave the others for a later phase.
• Give the agent a status field and a missing-items list for each referral. It may change the status only when a document is attached and a staff member has confirmed the match to the patient.
• Define the request template for each referring office, and have staff approve the wording during the first weeks before messages go out without review.
• Set a daily review where the front desk sees every referral that has stayed incomplete longer than your normal turnaround.

Scheduling needs its own rules. If your clinic books first visits against a verified referral, the agent should be able to see that status but not override it. Practices that manage waitlists and limited slots can look at the behavioral health intake and scheduling limits for a parallel case of drawing the line between booking and deciding.

Visit notes: drafting without signing

Visit notes carry the highest risk of a confident but wrong sentence, because a draft can read as finished before anyone has checked it. The safe pattern has the agent draft from structured inputs the clinician supplies, such as measured range of motion, exercises performed, the pain rating the patient reported, and a short dictation. Every line should show whether it came from the clinician or was inferred. Fields the clinician left empty stay empty in the draft and are marked as missing.

Illustrative scenario: a therapist finishes a session with a patient recovering from shoulder surgery. She dictates three lines: which exercises were done, that the patient tolerated the load well, and that the next session will focus on overhead reach. The agent places those facts in the matching note fields, leaves the pain rating blank with a visible flag, and adds the next visit date from the schedule. The therapist enters the rating, corrects one sentence and signs. The agent then lists a home exercise reminder as a task, which a staff member confirms before anything reaches the patient.

The note becomes a final record only after the clinician signs it. Your template, your documentation rules and the health privacy rules where you practice decide what must appear in it, so write those requirements down before the first draft is generated.

The approval points a clinician must keep

Approval points work best when they are written as a short decision guide that the agent follows and the systems enforce. Use these questions:

• Does the output change a clinical record? If yes, a clinician reviews and signs it before it is saved as final.
• Does it message a patient about care? If yes, it uses an approved template or a staff member sends it.
• Does it contact a referring office? Staff approve the first messages to each office, and any request for a new or changed order goes to the clinician.
• Does it touch payer or authorization data? The agent may prepare a packet, but a billing staff member submits it.
• Does the information come from an outside source? A scanned referral or pasted text may contain instructions. Treat it as data to read, never as a command, and route anything unusual to a person.
• Is the request outside the agent's assigned role? It stops and names the owner who should handle it.

Prompt instructions alone are not a control. The clinic's systems should refuse any action the agent is not permitted to take, whatever the agent is asked to do. Authorization decisions belong in the system, not in the model's judgment.

Access, records and untrusted documents

Connect the agent with the smallest set of permissions that lets the role work. Start with read access to referral and schedule records, grant write access only to drafts and task lists, and use separate credentials for each integration so one can be revoked without stopping the others. Log every action with the record it touched, the input it used and the person who approved it where approval was required. Those logs are what let you investigate a wrong message later.

Referral documents arrive from outside your clinic, and nobody controls what they contain. A scanned form or pasted email may include text written to change the agent's behavior. Keep extracted content in the data field, never in the instruction field, and make sure no document can trigger a message, a record change or an export on its own.

Health data rules depend on where your clinic operates, so involve your privacy officer before connecting patient records. Map which data the agent holds, how long drafts are kept, and who can see them. The NIST AI Risk Management Framework is a useful reference for turning those decisions into a documented practice rather than a verbal agreement.

How Be Human approaches the operation

Be Human begins with one role, which for a therapy clinic is usually referral tracking or draft visit notes rather than both at once. Before anything goes live, the engagement defines the agent's memory, the tools it may call, the approval boundaries and an observable success criterion. Managed operations then cover monitoring, recurring evaluations of drafts against clinician corrections, integration maintenance, scheduled permission reviews and incident response if a message reaches the wrong office. Changes to the agent's behavior are made as controlled improvements, and the clinical lead approves any change to the approval rules. The service is remote and available in English and Portuguese.

How to know the agent is working

Pick two or three measures that a clinic owner can check without relying on a vendor report. Good candidates are the number of referrals that reach a first visit with nothing missing, the time between receipt and verified status, how often clinicians heavily edit drafted notes, and how many outgoing messages had to be corrected or recalled.

Sample a fixed set of drafts each month and compare them with the signed versions. Build the sample from real cases, include difficult ones such as referrals with conflicting information, and record each failure with its cause: missing input, wrong template, unclear instruction, or a permission that was too broad. Anthropic's guidance on evals for AI agents describes this same practice of checking behavior against a defined set of tasks rather than trusting general impressions.

If the measures do not improve after two review cycles, narrow the role before adding anything new. A smaller agent that reliably flags missing referral items is worth more than a broad one that clinicians must double check. If you want to find the first task worth handing to an agent in your clinic, take the free Be Human AI diagnostic.

Frequently asked questions

Can an AI agent write physical therapy visit notes on its own? It can draft a note from material the clinician provides, but the treating therapist must review, correct and sign it before it becomes part of the record. The draft should separate what the clinician entered from what was inferred, and it should leave blank fields blank. Final clinical documentation remains the clinician's responsibility.

Does an AI agent replace a front desk or referral coordinator? No. It keeps the state of each referral, drafts requests and lists what is missing, so staff spend less time searching for information. Staff still approve outbound messages to referring offices and make authorization decisions.

Which referral details should be checked first? Start with the items your clinic already requires before booking a first visit, such as the referring provider, the order and reliable contact details. Write them into a checklist the agent can read, and confirm the list with your clinical lead before connecting any system.

Is it safe to connect the agent to scheduling and patient records? It can be, when access starts read-only, is limited to what the role needs, can be revoked, and is logged. Involve your privacy officer first, because health data rules depend on where you practice, and schedule permission reviews so access stays matched to the role.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.