AI for Medical and Allied Health Practices: What's Appropriate and What Isn't

Healthcare is the sector where AI is most oversold and where the consequences of getting it wrong are most serious. This piece is deliberately more cautious than our others.

Nothing here is clinical or legal advice. Practices are responsible for their own compliance obligations.

Where it genuinely helps

Reception overflow and after-hours

The clearest case. Most practices have periods where reception cannot answer everything — Monday mornings, lunch cover, the hour after close.

An agent can answer, book routine appointments against real availability, handle "are you open" and "where do you park", and take a callback request for anything else.

The critical design point: the agent must not triage clinically. Its job is to route, not assess. Anything suggesting urgency goes to a human immediately, and the threshold for "suggesting urgency" should be set deliberately low.

Appointment reminders and rescheduling

Unglamorous and reliably valuable. Non-attendance is a direct revenue loss and a wasted slot another patient wanted.

Automated reminders with an easy reschedule path reduce non-attendance meaningfully. This is not sophisticated technology and it does not need to be.

Recall and follow-up administration

Identifying patients due for follow-up and generating the contact list. The clinical decision about who needs recall stays with the practitioner; the administrative work of finding and contacting them does not.

Documentation and correspondence drafting

Drafting referral letters and summaries from structured notes, for practitioner review and sign-off.

The word "draft" is doing the work in that sentence. This must be a starting point a clinician edits, never something that goes out unreviewed.

Where it should not go

Anything resembling clinical triage

An agent must not decide how urgent a patient's symptoms are. The failure mode — telling someone their chest pain sounds routine — is catastrophic and entirely foreseeable.

Route by keyword to a human, with a low threshold and generous false-positive rate. Being over-cautious costs a phone call. Being under-cautious costs much more.

Anything that reads as advice

Patients will ask an agent medical questions. It will happen on the first day.

The agent must decline plainly and offer to connect them with the practice. "I can't help with that, but I can get you an appointment or have someone call you back."

Autonomous contact about results

Test results are conversations, not notifications. An agent can arrange the conversation. It should not deliver the content.

The privacy dimension

Health information is sensitive information under the Privacy Act and attracts stronger protection than ordinary personal information.

Practical requirements:

• Know where the data goes. Most voice platforms process offshore. For health data that may be unacceptable depending on your obligations — ask directly and get the answer in writing.
• Minimise collection. An agent booking appointments needs a name, contact and appointment type. It does not need symptoms, and should be built not to ask.
• Decide retention deliberately. Consider whether recording is needed at all. Zero-retention configurations exist where audio is processed live and nothing is stored.
• Tell patients. APP 5 requires notification at or before collection.
• Check your indemnity. Some policies have conditions around automated patient contact. Worth a call before deploying.

Questions to ask any vendor

1. Where is voice data processed, and can it be kept onshore?
2. Is audio recorded? Can that be switched off entirely?
3. What exactly triggers escalation to a human, and can we set it ourselves?
4. What does the agent say when asked a clinical question?
5. Can we see the full conversation brief, not a summary of it?
6. What happens if the escalation contact does not answer?
7. Who at your end can access our patient data, and under what circumstances?

A vendor who has worked in health will answer these immediately. One who has not will improvise, and you will hear it.

A realistic expectation

For a practice with two to six practitioners, sensible automation typically returns:

• Reception time back, mostly at the peaks where it hurts most
• A measurable reduction in non-attendance
• Fewer abandoned calls outside hours
• Faster correspondence turnaround

What it does not do is replace reception staff. Practice reception involves reading distress, knowing regulars, and managing a waiting room. Those are not automatable and should not be.

The practices that do well with this treat it as covering the gaps rather than replacing the people.