The window of intent is very short
Someone who decides on Sunday night to finally book a check-up has usually un-decided by Tuesday. A follow-up that lands after the resolve fades is a follow-up to nobody.
Someone enquires about a health check at eleven at night, which is when people worry about these things. The call back comes two days later, by which point they have either booked elsewhere or talked themselves out of it entirely.
This agent follows up the same evening. It books the slot, explains the fasting instructions slowly, confirms on WhatsApp, and is explicitly forbidden from saying anything clinical.
Healthcare enquiries are unusually fragile, because the person making them would often prefer an excuse not to go.
Someone who decides on Sunday night to finally book a check-up has usually un-decided by Tuesday. A follow-up that lands after the resolve fades is a follow-up to nobody.
A front desk handling walk-ins, calls and billing does not have a spare hour to chase web enquiries, and outbound is always the first thing dropped.
Fasting, timing, what to bring, what to stop taking. Delivered fast over a bad line, these produce a patient who arrives unprepared and has to be rebooked.
An unconfirmed slot is roughly a coin flip. A reminder call the day before is known to work and is almost never made consistently.
Enough to book correctly. Nothing that resembles taking a history.
The moment that matters in healthcare is not the booking. It is this one.
Good evening, this is Kanika calling from the hospital. You had enquired about a full-body health checkup on our website. Is this a good time?
Yes. Actually I've been getting chest pain for a few days, is that something serious?
I am really sorry to hear that, and I want to be straight with you: I am a coordinator, so I am not the right person to say what it might be. The doctor is. Let me get you the earliest slot I can.
Okay. When can I come?
Let me confirm the exact consulting times and send them to you in writing rather than telling you something that turns out to be wrong. Is tomorrow morning workable, or would evening be easier?
Morning is fine.
I will send the slot, the address and the preparation instructions on WhatsApp. One thing to note now: for the blood work you will need to come fasting, so nothing after dinner except water.
It declines to assess, routes to a consultation, and does not guess the timing either. Both refusals are deliberate.
This is the whole reason a healthcare deployment needs a written brief rather than a general-purpose assistant.
It never diagnoses, interprets a symptom or a report, or suggests a treatment. It will not say whether something sounds serious. If a caller describes symptoms, it acknowledges kindly and routes them to a consultation.
It never invents a doctor's name, availability, package price or wait time. It works from the department list, consulting days and published package contents you supply, and confirms anything else in writing.
It answers cashless questions only for insurers you have confirmed as empanelled. A wrongly promised cashless admission is discovered at the counter, by a patient, on a bad day.
A language model with no constraints will answer a medical question if asked one, because answering questions is what it does. On a recorded call from a hospital's number, that answer becomes the hospital's statement. The risk is not that the model is usually wrong; it is that it is occasionally wrong, confidently, to an anxious person who will act on it.
So the healthcare persona is built the other way round: a short list of things it does (book, confirm, explain logistics) and an explicit list of things it must refuse, with the refusal wording written and approved rather than improvised. It is a coordinator, and it says so when pushed.
The vocabulary is tuned too, which matters more here than in most sectors. Department and procedure names are exactly the words generic speech recognition mangles, and a transcript that records gynaecologist as something else is a booking error waiting to happen.
Every persona has its own script, its own vocabulary and its own list of things it will not say.
The agent is a coordinator: it books, confirms and explains logistics, and it is explicitly instructed never to diagnose, interpret or advise. Where your own protocols require specific disclosures or consent language on outbound calls, that gets written into the brief before it goes live. We would rather agree that in advance with your team than assume it.
No. It can tell a patient that reports are ready and arrange a consultation to go through them. It will not read out or interpret a value, and that boundary is not adjustable.
Yes, and that is often where it pays for itself fastest. Confirmation the day before, with the preparation instructions repeated in writing, addresses the two most common causes of a wasted slot.
It is instructed to slow down, acknowledge, and route to a human rather than continue a script. Where you want a hard handover rule, for example any mention of a specific set of symptoms, we build that in as an escalation.
Yes. The structure is the same, and what changes is the service list, the preparation instructions and the department vocabulary.
That refusal is the product. Everything else in the category can book an appointment.