AI for Physical Therapy Clinics: Plans of Care and Reminders
How a physical therapy clinic can use an AI agent to track plan-of-care dates, authorizations and reminders while clinicians keep every treatment decision.

An AI agent can help a physical therapy clinic keep plan-of-care dates, authorization status and patient reminders current by reading the records it is allowed to see, flagging what is due and sending messages from approved templates. It should not decide whether a patient needs more visits, a new goal, a different intervention or discharge, because those decisions belong to the treating therapist. The goal is fewer lapsed plans and missed follow-ups, with clinicians reviewing every exception.
Key takeaways
- A plan of care is a set of dates and decisions. The agent should track the dates and route the decisions to people.
- Authorization and certification deadlines belong in a visible queue with a named staff owner.
- Reminders should come from approved templates, and only routine messages should send without review.
- Clinical content, goal changes and discharge stay with licensed clinicians. The agent holds no permission to change them.
- Start with one queue that has a clear deadline, prove it works in flag-only mode, then widen the scope.
Why plan-of-care dates slip through a busy clinic
A plan of care is a working document with dates attached to it. Certification periods, referral visit limits, payer authorizations and re-evaluation deadlines usually live in different places: the electronic record, a payer portal, a fax tray, a shared inbox and a therapist's notes. Each of those sources is reasonable on its own. Together they make it easy for a date to pass without anyone deciding that it should.
The problem is rarely a lack of care. The administrative clock runs independently of the clinical conversation. A therapist may plan to extend treatment at the next visit while the front desk does not know the current authorization ends that week. A patient may feel better and stop attending, and nobody notices the plan is still open until the paperwork comes due.
An agent is useful here because it watches the calendar continuously and does not get pulled away by a busy morning. It is not useful if it is given a vague job such as helping with patient follow-up. The first step is to name the specific dates it tracks and the people who respond to them.
What the agent should track
Start with a short list of records and dates, narrow enough that one person can review it each day. A workable first scope includes:
- Start and end dates of each active plan of care or certification period
- Referral details, including any visit count or duration the referral allows
- Authorization requests with their status, submission date and expiry date
- Re-evaluation and progress report deadlines set by the treating therapist
- Scheduled visits that fall outside an active authorization window
- Patients who have not returned within an interval the clinical lead defines as a follow-up trigger
Each item needs an owner. The agent can flag a date, but a named person on the staff side has to answer it. A queue with no owner is only a list of worries. Clinics that already run recall queues for another kind of care can compare the logic in AI Agent for a Chiropractic Practice: Intake and Visit Recalls, since the same questions about ownership and timing apply.
Where reminders end and clinical judgment begins
Reminders are the easiest part to automate and the easiest part to get wrong. A message confirming tomorrow's appointment is routine. A message implying that the patient should keep coming because the plan is still active is a clinical statement. The boundary should be written down before anything is switched on, and it should be the same for the agent and for staff.
This decision guide keeps both applying the same rule:
- If the item is administrative, such as an expiring authorization or a missing signature, the agent flags it to the named billing or front desk owner.
- If a patient asks about pain, symptoms or whether to continue treatment, the agent does not answer. It routes the message to a clinician with the question and the relevant visit history.
- If the plan itself needs to change, such as new goals, a different frequency or discharge, the agent prepares a summary for the therapist and takes no action on the record.
- If the message is routine and uses an approved template, such as an appointment confirmation or a request for intake paperwork, the agent may send it.
- If a payer or outside party asks for clinical information, the agent drafts a response for staff review and sends nothing.
Messages from patients, payers and referring offices should be treated as untrusted input. A message can contain text that tries to change how the agent behaves, which is why the OWASP guidance on prompt injection prevention matters before any agent reads an inbox. Permissions should never depend on what a message asks the agent to do. Clinics that have already drawn limits around intake and scheduling in mental health care will recognize this pattern, which is covered in AI for Behavioral Health Clinics: Intake and Scheduling Limits.
Staff should be able to answer one question about any message the agent sends: which rule allowed it.
Handling authorizations without guessing
Coverage rules differ by payer, plan and region, and they change over time. An agent should never infer whether a visit is covered from patterns it has seen before. A safer design gives it a reference table that the clinic maintains: the payers it bills, the authorization requirements for each, and the contact route for questions. The agent compares the record against that table and flags mismatches. Billing owns the table, and changes go through the same review as any other clinic policy.
Access should follow the principle of least privilege. If the agent needs to read authorization status from a payer portal or a scheduling system, its connection should carry only the scopes it needs, and staff should be able to revoke it without touching anything else. The OWASP authorization guidance describes the same idea in general terms: a system can be allowed to see a field without being allowed to write it. For a clinic, that usually means the agent reads expiry dates and writes only to the staff queue.
An illustrative rollout for a four-therapist clinic
Illustrative scenario: a four-therapist outpatient clinic has a recurring problem with plans that expire while patients are still attending. The front desk tracks expiring authorizations on a spreadsheet that is updated when someone remembers. The clinic decides to try an agent on one queue only: plan-of-care and authorization expirations.
The rollout follows a sequence the clinic can check at each step:
- Name one staff owner for plan-of-care compliance and one clinical reviewer who approves reminder wording.
- Export a read-only list of active episodes with start dates, end dates and authorization status.
- Agree with the clinical lead on the flagging window, meaning how far before an expiry an item should appear in the queue.
- Write the reminder templates in plain language and have the clinical reviewer approve each one.
- Run the agent in flag-only mode, and compare its queue each day with what staff actually catch.
- Review every disagreement, fix the rule or the data behind it, and only then allow routine reminders to send.
The point of flag-only mode is to find out where the rules are wrong before a patient receives a wrong message. In a case like this, the first comparison often reveals records where the end date was never updated after a re-evaluation. That is a data problem, and fixing it does more for the clinic than any change to the agent.
How Be Human runs this as a managed service
Be Human starts with one defined role, the memory and tools that role needs, the approval boundaries for each action, and one observable success criterion, such as the share of expiring plans flagged early enough for staff to act. For the example above, the agent's memory holds the reminder templates, flagging windows and payer reference table, and the clinical boundary is written into the role itself. After implementation, managed operations cover monitoring of the queue, recurring evaluations of whether flags were correct, maintenance when the scheduling or records system changes, periodic permission reviews, incident response if a message goes out in error, and controlled improvements to the rules. The approach follows the risk management logic in the NIST AI Risk Management Framework, with humans responsible for the decisions the agent is not allowed to make. The referral and visit note side of the same clinic is covered in Superintelligent Agents for Physical Therapy: Referrals and Visit Notes. Service is remote and available in English and Portuguese.
Choosing the first task
Keep the first task small enough to measure in a normal week of work. Good candidates are queues that already have an owner and a clear deadline, such as expiring authorizations or re-evaluations due. Poor candidates are open-ended goals like improving patient engagement, because nobody can tell whether the agent did its job. Once the first queue is reliable and its exceptions are understood, the same pattern can extend to other administrative dates in the clinic. Take the free Be Human AI diagnostic to find the first task worth handing to an agent in your clinic.
Frequently asked questions
Can an AI agent decide when a patient's plan of care should end? No. The agent can track the end date, flag that it is approaching and prepare a summary for the treating therapist. Discharge, changes in goals and additional visits remain clinical decisions, and the agent should have no permission to alter the plan.
Does the agent need access to the full clinical record? Usually it does not. A well-scoped setup reads only the fields needed for dates, authorization status and appointments. Access should be granted through connections that staff can review and revoke at any time.
How do you know the agent is working? Define the success criterion before rollout, then compare the agent's daily queue with what staff catch during flag-only mode. Track missed items and false flags, and review them in a recurring evaluation so that rules improve based on evidence rather than impressions.
Can the agent work in Portuguese as well as English? Yes, the role can be configured for both languages. The approved reminder templates should still be written and reviewed by the clinical reviewer in each language, rather than translated automatically without checking.
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.


