AI Agent for a Physiotherapy Clinic: Sessions and Follow-Up

An AI agent for a physiotherapy clinic handles the administrative thread around treatment. It tracks session packages, offers reschedules within the clinic's rules, and checks in on home exercises between visits. It does not diagnose, change a plan or interpret pain, and it hands those moments to a therapist with the context attached. The practical gain is that fewer patients drift away mid-treatment because nobody followed up.

Key takeaways

• Start with three jobs: package balance, reschedules and exercise check-ins. Each is repeatable, visible and easy to review.
• Keep clinical decisions human. Pain changes, exercise modifications, red-flag symptoms and discharge always go to a therapist.
• Give the agent narrow access. Let it read the schedule and package ledger, and let it write only drafts and proposed changes until its behavior has earned more.
• Write the clinic's rules down first (cancellation terms, package expiry, which therapist sees which patient). An agent cannot follow rules that live only in people's heads.
• Measure outcomes you can observe, such as follow-ups that were caught and slots that were refilled, and review a sample of conversations every week.

What does an AI agent do in a physiotherapy clinic?

Physiotherapy has a particular rhythm. A patient is assessed, receives a plan, is assigned a package of sessions, and then returns over several weeks while doing exercises at home. Most of the administrative friction sits between visits: a missed appointment that is never rebooked, a package that runs out without a conversation about next steps, an exercise program the patient quietly stops doing.

An agent is useful in those gaps. It keeps a working memory of each patient's administrative state: sessions used, sessions left, next appointment, assigned therapist, preferred channel and language. It acts through the clinic's existing tools, which are the calendar, the patient record system, messaging channels and the billing ledger. It follows written rules and logs what it did.

What it is not is a clinician. It should never decide whether a symptom is serious, whether an exercise should change, or whether a patient is ready for discharge. Building that boundary into the design, rather than relying on good intentions, is the core of the work. The same logic appears in AI agents for clinics, where scheduling and intake are automated while judgment stays with the care team.

How does an agent manage session packages?

A package is a small contract: a number of sessions, a validity period, cancellation rules and sometimes a shared allowance across a treatment plan. When these rules live in a spreadsheet and a receptionist's memory, errors creep in. Sessions are deducted twice, an expired package is honored by accident, or a patient learns on the day that nothing is left.

The agent can maintain the ledger as a single source of truth and act on it:

• Record a session as used only when the therapist or front desk confirms attendance.
• Apply late-cancellation and no-show rules exactly as written, and flag exceptions instead of improvising.
• Tell the patient their remaining balance and expiry date when they ask, and before the final session.
• Prepare a reassessment or renewal request for the therapist when a package nears its end, so the conversation about next steps is clinical first and commercial second.
• Escalate any dispute about charges to a named person.

Notice what the agent does not do. It does not grant discounts, extend validity on its own or waive a fee. Those are policy decisions, and the clinic should decide in advance who may make them. If the rule is that the clinic manager approves extensions, the agent's job is to prepare the case, with dates and history, and send it to that person.

How can it handle reschedules without breaking the treatment plan?

Rescheduling looks simple until you count the constraints: therapist availability, room or equipment, continuity of care, the spacing a protocol needs between sessions, package expiry and the patient's own limits. An agent that only sees open slots will offer times that are technically free and practically wrong.

Give it rules, not just a calendar. AI Agent Calendar Integration: Scheduling Rules That Hold Up covers the general pattern. In a physiotherapy clinic the rules usually include:

• Prefer the same therapist, and say so plainly when that is not possible.
• Respect the minimum and maximum spacing the therapist set for this patient's plan.
• Never move a patient past package expiry without routing the change for approval.
• Offer a short list of options, not an open calendar.
• Hold a proposed slot for a limited time, then release it.
• Confirm the change in the patient's preferred language and channel, and update the ledger and the calendar together.

When a patient cancels because of a flare-up, a fall or a new symptom, the agent should not simply rebook. It acknowledges the message, offers no automatic slot that assumes the patient is fit to attend, and asks the therapist to review. That one branch matters more than the rest of the scheduling logic.

What can the agent say about home exercises, and what must it route?

Exercise follow-up is where the temptation to overreach is strongest, because the conversation feels clinical. A workable design lets the agent do only four things: send the program the therapist already assigned, using the therapist's own instructions; remind the patient at the agreed rhythm; ask a fixed set of check-in questions written by the therapist; and answer logistical questions, such as where to find a video or how to book. Answers are stored in the record as patient-reported, never as assessed.

Use a decision guide for every incoming reply:

• The patient completed the exercises and has no concerns: log it and thank them.
• The patient missed exercises for ordinary reasons such as time or travel: log it, offer to resend the program, and note it for the next visit.
• The patient reports more pain, new numbness, swelling, a fall, or anything on the therapist's concern list: stop, give no advice, tell the patient a therapist will review, and alert the therapist right away.
• The patient asks whether to change, skip or add an exercise: route to the therapist and do not answer.
• The message sounds urgent: give the emergency instruction the clinic wrote in advance and alert staff.
• The message is ambiguous: ask one clarifying question, and if it is still unclear, escalate.

Patient messages are also untrusted input. A reply, a forwarded note or an attached document might contain text that tries to steer the agent. The OWASP guidance on prompt injection is a sound reference here. The practical rule is that the agent treats message content as information, never as instructions, and that its permissions are limited enough that a manipulated reply cannot do harm.

Illustrative scenario: one patient, one week

This is an illustration, not a case study. A patient recovering from a shoulder procedure has a package with two sessions left. On Monday the agent sends the exercise reminder the therapist set up. The patient replies that they missed several days because of travel and asks to move Thursday's session. The agent checks the therapist's spacing rule, finds two slots with the same therapist that stay inside the package validity, and offers both. The patient chooses one. The agent updates the calendar and the ledger and logs that the exercises were missed for logistical reasons.

On Wednesday the patient writes that the shoulder feels sharper at night. The agent does not suggest rest or a lighter exercise. It replies that a therapist will look at the message, then sends the therapist the message text, the recent exercise history and the upcoming appointment. The therapist calls the patient and decides whether to bring the visit forward. Later, as the package nears its end, the agent prepares a reassessment request rather than a renewal offer, because the therapist, not the ledger, should decide what comes next.

A step-by-step plan to start

• Step one: choose one role, such as package and reschedule coordinator, and write its success criterion in terms you can observe, like follow-ups logged and escalations answered.
• Step two: write the rules. Include cancellation terms, expiry, spacing, therapist assignment, languages and who may approve exceptions. How to prepare company documents for an AI agent helps with turning scattered policies into usable material.
• Step three: have each therapist write their check-in questions and their concern list, in their own words.
• Step four: connect tools with the smallest access that works. Least privilege for AI agents explains why read access and draft-only writes are a good first position.
• Step five: tell patients they are talking to the clinic's automated assistant, and make a route to a person available in every conversation.
• Step six: run in draft mode, where the agent proposes and staff approve. Collect anonymized real situations into a test set, as described in Build a golden test set for your AI workflow, and include the awkward cases: ambiguous messages, upset patients, urgent symptoms.
• Step seven: widen autonomy one action at a time, starting with the ones that carry the least clinical risk, such as sending reminders.

How Be Human approaches this

Be Human designs, implements and operates this kind of agent as part of a clinic's own infrastructure. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review and incident response. Authorization is enforced in the tools themselves, not only in the instructions, so the agent cannot do what its role does not allow. Health data is sensitive, so location, retention and access are agreed before launch, and the clinic stays responsible for its duties under the privacy rules that apply to it, such as LGPD in Brazil, overseen by the ANPD, or equivalent law elsewhere.

Frequently asked questions

Can an AI agent give physiotherapy advice to patients? It should not. The agent can pass on the exercise program a therapist assigned and ask the therapist's check-in questions, but anything that involves interpreting symptoms or changing a plan goes to a therapist. Clinics should write this boundary into the agent's rules and test it with realistic messages.

Will patients accept messages from an agent? Most accept them when the messages are useful, brief and honest about who is writing. Say clearly that it is the clinic's automated assistant, use the patient's preferred language and channel, and always offer a way to reach a person. Trust drops quickly when an agent pretends to be human or cannot hand over.

What does the agent need access to? At minimum, the calendar, the package ledger and the approved exercise programs, with read access first and drafted changes second. Clinical notes should stay out of reach unless a specific task truly needs them. Access is reviewed regularly, and anything unused is removed.

How do we know it is working? Choose measures tied to the role: follow-ups that were caught and answered, slots that were refilled after a cancellation, escalations that reached a therapist promptly, and disputes about balances. Review a sample of real conversations each week with a therapist and the front desk. If reviewers keep correcting the same behavior, fix the rule or the instruction before extending the agent's autonomy.

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