Human handover and booking-rule design

The quality of an AI assistant in a practice is decided less by what it can do than by what it hands over, and how fast a person picks it up. Write these rules before launch, in plain language, and give a copy to everyone on reception.

The design principle

I design so that the tool assists and a person decides. UK GDPR treats a decision as based solely on automated processing where there is no meaningful human involvement ^[1], and where a significant decision of that kind is made, the person must be able to obtain human intervention and contest it ^[2]. Rather than argue about which steps might be caught, I keep the tool to steps a patient would not mind a machine doing, and make a person reachable at every point.

When it must hand over

• The patient asks for a person, in any words, at any point.
• Any mention of pain, swelling, bleeding, injury, an accident or distress.
• Any clinical question: what they need, whether something is safe or suitable, what a symptom means.
• Any question about the cost of their own treatment, rather than a published fee.
• A complaint, or obvious frustration.
• A safeguarding concern, a vulnerable patient, or someone speaking for another person.
• Anything it does not understand after one attempt to clarify.

A good handover message is short and honest: "I am going to pass this to a member of the [Practice] team. They will contact you by [time]." Out of hours, urgent messages must also give your arrangements for emergency care, which the GDC expects patients to receive clearly ^[3].

What it may book

Booking rules are the practice's rules, not the tool's. Start with the narrowest list you can, and add to it only after the weekly review shows the current list is working.

Adjust the list to your own appointment book. The test for each line: if the tool books this wrongly, will a person notice before the patient is affected?

Logging and follow-up

• Every handover is logged with the time, channel, reason and the patient's own words.
• Each one has a named owner and a time by which they will respond, in hours and out of hours.
• The owner records the outcome: booked, passed to a clinician, callback failed, complaint opened.
• Unclaimed handovers are checked at the start of each session, before anything else.

Complaints raised through the tool go into your complaints procedure like any other. The GDC expects a written record of complaints and responses, kept separate from patient records ^[4].

Once a week, count handovers by reason and check how quickly each was picked up. A reason that keeps recurring tells you either that the tool needs a better approved answer, or that the task should never have been automated. Test the rules in practice with the AI receptionist evaluation checklist.