Podiatry is a useful test case for AI writing because the documents split cleanly. Roughly half are administrative — appointment preparation, billing explanations, directions, what happens at the first visit. Those can be drafted. The other half are clinical — post-procedure care, wound instructions, footwear and activity guidance — and for those, a language model can help with structure but never with content.
Three document types carried the value: appointment preparation, front-desk replies, and referral letter formatting. Three categories must stay clinical: wound and dressing care, weight-bearing and activity guidance, and anything involving reduced sensation or circulation.
Where it helped
Appointment preparation
Bring your current footwear, bring any orthoses you already use, bring imaging if you have it, arrive ten minutes early for paperwork, wear or bring socks that allow access to the foot. Generated versions were better than what we had been sending because they were complete and consistently ordered. One edit always required: remove anything that does not apply to your practice.
Front-desk replies
The recurring questions — whether a referral is needed, what a first appointment costs, how long it runs, whether you treat children, what happens if a dressing needs changing before the next visit. Written once as a set of standard replies, these removed most of the back-and-forth. Keep the clinical answers out of this set: a question about a symptom gets routed to a clinician, not answered from a template.
Referral letters
Generated text is useful for the frame — presenting complaint, duration, what has been tried, relevant history, what you are asking of the recipient — but every clinical field must be filled from the record. Treat it as a structure with mandatory blanks, not a draft to be skimmed. In our testing, the errors that got through were always in the history section, where a generated summary filled a gap with something plausible.
Where it failed — and where it must not be used
Wound and dressing care
Drafts gave instructions about dressing changes, soaking, and what a healing wound should look like. In a population where a meaningful share of patients have diabetes or circulatory disease, a wrong instruction is not a correction — it is a complication. This content comes from your protocols and the clinician's judgement for that patient, with no generated text in the final document.
Weight-bearing and activity guidance
Drafts stated how long to stay off a foot, when to return to work, or when running is reasonable. These depend on the procedure, the fixation, the healing seen at review, and the patient. We now write these as questions the patient should ask at review, rather than as guidance in a handout.
Statements about sensation or circulation
Generated text sometimes reassured patients about numbness or colour changes. Those are findings that require examination, and reassuring language in a handout can delay someone seeking review. Any mention of altered sensation in a patient document should instruct the reader to contact the clinic.
Medication and dosage
Drafts produced confident text about pain relief and anti-inflammatory use. Nothing about medication belongs in generated patient material — not the choice, not the dose, not the duration. That is a conversation between the clinician and the patient.
What patients actually read
A patient leaving with post-procedure instructions reads three things: what to do today, what means call us, and when the next appointment is. Handouts organised around those three questions outperformed chronological day-by-day sheets in our testing, measured by how few clarification calls the practice received.
The second most-read document is the billing explanation — what is covered, what the gap will be, when payment falls due. Written plainly, it prevents most billing complaints.
One more document earns its keep: a plain-language description of what the first appointment involves, sent at booking. Patients arrive unsure whether they will be examined, treated, or referred on the day, and that uncertainty turns into unreturned paperwork and late arrivals. A short paragraph naming the sequence — history, examination, discussion, plan — reduces no-shows more reliably than a reminder.
Editing time, from our 30-day benchmark
Appointment preparation went from about 16 minutes to 4, near $6 saved at $30 per hour. Front-desk reply sets took roughly 90 minutes to build once and have been reused since. Referral letters dropped from around 12 minutes to 5, with the clinical fields still filled by hand. Post-procedure clinical content was not generated.
A workflow that holds up
Build the standard reply set once and keep it in one place. Generate referral letters from a fixed frame with mandatory clinical blanks. Organise aftercare around today, warning signs, and the next appointment. Wound care, weight-bearing guidance, statements about sensation or circulation, and anything involving medication stay with the clinician — drawn from your protocols, written for the individual patient, and reviewed by whoever is accountable.
Administrative text scales, clinical text does not
Preparation and front-desk replies are worth automating. Patient care instructions are not.
Try Rytr →Frequently Asked Questions
Can AI write post-procedure instructions?
It can organise them. The content must come from your clinical protocols and be specific to the patient and procedure.
Can it answer patient questions about symptoms?
No. Route those to a clinician. A template reply to a symptom question is a clinical risk.
What about medication guidance?
Never generated. Medication is a clinician-patient conversation.
Can it draft referral letters?
The structure, yes. Every clinical field is filled from the record and reviewed before sending.
Are preparation instructions safe to generate?
Yes, with one review for items that do not apply to your practice.
Is this worth it for a solo practitioner?
The front-desk reply set is the largest one-time saving, and it keeps working afterwards.