Most patient messages are routine: opening hours, whether you see new patients, roughly what a check-up costs, can I book for Tuesday. An AI can answer those well, in seconds, at 11pm.
A small number are not routine, and those are the ones that decide whether an AI front desk is safe. This is the list we work from, and it applies whichever system you use.
Messages that should always go to a person
- Anything that sounds urgent or clinical: severe or unbearable pain, heavy bleeding, facial swelling, trouble breathing, mentions of A&E or 999.
- Complaints and anything legal: "I want to complain", solicitors, the ombudsman, Trading Standards.
- A request for a human: "can I speak to a real person", "are you a bot".
- Questions about medicines and doses: "what painkiller should I take", "can I have ibuprofen with this".
- Abuse or distress.
The trigger should be the message, not the AI's confidence. A model that is sure of itself is exactly the one that will happily answer a dosing question it shouldn't.
What the patient should be told
Going quiet is the wrong hand-off. The patient should know straight away that a person has their message, and, for anything urgent, what to do right now.
For an urgent message, something like: "I'm sorry you're dealing with this. I've flagged your message as urgent for the practice team. If you need help right now, please call the practice directly. Out of hours, call NHS 111, or 999 if it's life-threatening."
For a medicines question: "I can't give medical advice here, but I've passed your question to the team. For questions about medicines, please speak to your GP or pharmacist."
Never promise a call-back time you can't guarantee. "Someone will call you in five minutes" at 9pm on a Sunday is worse than saying nothing.
What your team needs to see
A hand-off only works if someone picks it up. Your team needs:
- one place where handed-off conversations sit, separate from the general inbox;
- the reason it was handed off (urgent, complaint, asked for a person, medicines);
- the full conversation, so nobody asks the patient to repeat themselves;
- a named owner, and a check at the start of each session.
Once a person has taken a conversation, the AI should stay out of it until they hand it back.
Stalled is not the same as escalated
A stalled conversation is a different problem. Nothing has gone wrong; the patient has simply gone quiet, or nobody has replied yet. The useful checks are time-based: an enquiry with no first reply after a few minutes, a conversation where the patient hasn't answered for a couple of hours after your last message, and anything open with no activity for a day.
Those need a nudge, usually drafted for a person to approve, with a cap on how many follow-ups are sent and a gap between them. They don't need the urgent treatment above, and mixing the two queues buries the messages that matter.
How ReplyOS handles this today
In ReplyOS, a message with urgent clinical wording, a complaint or legal language, a request for a person, or abuse stops the automatic reply and puts the enquiry in the Needs human queue. On web chat the patient is told it has gone to the team, with NHS 111 and 999 for anything urgent. When a member of staff takes the enquiry, the AI stops replying on it.
Stalled conversations are detected separately, at 10 minutes without a first reply, 2 hours without a patient answer and 24 hours without activity by default, and the follow-up is suggested for staff to approve unless the practice turns on automatic sending. There is a cap of three attempts with a six-hour gap.
We're still tightening parts of this, particularly the messages patients receive on SMS, WhatsApp and email after a hand-off, and we'll update this post when that changes.
Questions
Which patient messages should an AI receptionist never answer on its own?
Urgent or clinical messages (severe pain, bleeding, swelling, breathing problems), complaints and legal threats, requests to speak to a person, questions about medicines or doses, and abuse. These should go straight to a member of the team.
What should the patient be told when a message is handed off?
That a person has their message, and for anything urgent, to call the practice, NHS 111 out of hours, or 999 if it is life-threatening. Don't promise a call-back time you can't guarantee.
Is a quiet conversation an escalation?
No. A conversation that has gone quiet needs a time-based follow-up, ideally approved by a person and capped. Keep it in a separate queue from urgent hand-offs.
Sources
- NHS, "When to use 111" The guidance the urgent holding message points patients to.