The enquiry system that knows when not to answer.

Most of the AI demos we see in healthcare show the same thing. A patient asks a question, and a chatbot answers it instantly. It looks impressive. It is also the wrong way round.

When we started working with a functional medicine clinic in Ireland, the clinic manager set one condition before anything was built. No email leaves the clinic without her approval. We agreed straight away, and then we went further. We decided the system should know which questions it must never try to answer at all.

Where the hours really went

The clinic had a problem every small practice will recognise. Enquiries came from four places: the clinic’s email, Instagram, Facebook, and phone calls written up by hand. One person answered all of them, personally, from scratch.

Behind that sat three years of client history spread across three spreadsheets and an export from the lab portal. Every week brought more than 20 manual status updates, tracking each sample from kit sent to results received.

None of this was clinical work. It was the work around the clinical work, and it was eating the day of the one person patients most wanted to hear from.

Sorting comes before answering

The first thing the system does is not write anything. It sorts. Spam and supplier mail are filtered out. Every real enquiry is classified into one of four types: a critical case, a request for test information, a message that mentions symptoms, or a direct question about a test.

That classification decides everything after it. A question about what a test costs, or how long results take, can be drafted from the clinic’s own library of 24 templates, with live pricing, in the clinic’s own voice. A message that describes symptoms is a different thing entirely.

The gate

Clinical or regulated enquiries are never drafted by AI. They route to a “speak to your GP first” response, and they stop there for a person to handle.

In the first 30 days, 81 of 214 enquiries were held back. That is nearly two in five. In most software, a number like that would be reported as a failure rate. Here it is the point. It is the guardrail doing exactly what it was designed to do.

We think this is the right way to measure AI in healthcare. The question is not how much it answers. It is whether it answers only what it should, and whether a person can see, every time, which side of the line each message fell on.

What the manager does now

She still approves every reply. The difference is that most replies arrive already written, in her voice, with the right price, and she edits rather than starts from nothing. Over 30 days, the clinic counted 21 hours given back to the team, at 6 minutes for each enquiry handled.

She also owns the template library. When a price changes, she changes it herself, with no development ticket. That matters more than it sounds. A system the clinic cannot change without us is a system that slowly goes out of date.

The lab updates now happen on their own. Test orders move across a five-stage board, advancing automatically when the lab’s emails arrive. Every update lands in one record, which now holds 3,216 clients.

What we would tell any practice manager

Start with the enquiries, not the clinic. The front desk is where the volume is, and it is where a careful system can give hours back without going near a diagnosis.

Write the line between routine and clinical down before anything is built, and make it conservative. You can loosen a gate later, once you trust it. You cannot take back an answer that should never have been sent.

And keep one person’s name on every outgoing message. Patients notice when a clinic still sounds like itself.