What AI is really doing in clinics, and what it is not

The strongest published evidence that AI helps clinics is not about chatbots, diagnosis or robots. It is about paperwork. The largest real deployment so far is AI note-taking for doctors, and the clearest lesson from it applies just as much to the front desk: AI earns its place where the work is frequent, repetitive and mostly text.

Where the evidence is: doctors' notes

The Permanente Medical Group, Kaiser Permanente's physician group in Northern California, rolled out AI scribes that listen to a visit and draft the clinical note. Over 63 weeks, 7,260 physicians used it across about 2.5 million patient visits. The group's follow-up study in NEJM Catalyst estimated it saved 15,791 hours of documentation, the equivalent of 1,794 eight-hour working days.

Patients noticed. In the group's survey, 47% said their doctor spent less time looking at the computer, and none reported a negative effect on the visit.

The caveats matter as much as the headline. The hours are an estimate against doctors who did not use the tool, not the result of a controlled trial. Most of the benefit went to heavy users, and doctors who used it rarely said fixing the drafts took longer than typing. It worked because it was put on the most repetitive job in the building, and only for the people who used it every day.

Why this matters if you never buy a scribe

The scribe is proof of a pattern, not a product recommendation. In Canada, the Canadian Federation of Independent Business estimated in 2023 that doctors spend about 18.5 million hours a year on unnecessary administrative work, which it put at the equivalent of 55.6 million patient visits. That is an estimate scaled from a Nova Scotia study, not a count, but the direction is not in dispute.

The same kind of work fills the front desk: confirming appointments, chasing cancellations, answering the same five questions, sending the follow-up someone meant to send. It is frequent, it follows rules, and getting it slightly late costs patients rather than lives. That is exactly the shape of work where automation pays back.

Most of what helps the front desk is not AI at all

This is the part the hype skips. The best-evidenced front-desk tool is the humble text reminder. A Cochrane review of 8 randomised trials covering 6,615 people found text reminders raised attendance from 67.8% to 78.6%, though the authors rate the evidence as low to moderate quality. In a randomised study of more than 158,000 Kaiser Permanente visits, adding a second text cut no-shows among high-risk patients by 7% in primary care and 11% in mental health, relative to one text.

None of that needs a language model. A reminder is a rule: a fixed trigger and a fixed message. AI earns its place at the few points where something has to be read and understood, such as a patient's free-text reply, a voicemail, or a question that arrives at 9 PM. Putting a model where a rule would do makes a system slower, dearer and harder to fix.

What AI is not doing in clinics

• Replacing the front desk. The published deployments take work off people; none of them remove the person who handles the exception.
• Making clinical decisions at reception. Anything that sounds like a symptom should go to a human, every time, and a well-built system routes it there.
• Working without someone checking. The scribe study's doctors still reviewed every note. Front-desk automation needs the same: a person who can see what it sent.
• Delivering a guaranteed number. Every figure in this article comes from a specific setting with its own caveats. Your clinic will land somewhere different.

A sensible order for a clinic

1. 1.Reminders and confirmations first. Rule-based, cheap, and the best evidence base of anything here.
2. 2.Then the first reply: missed calls, after-hours messages and web enquiries answered in minutes, with anything clinical handed to a person.
3. 3.Then follow-up: the cancelled patient who never rebooked, the treatment plan nobody chased.
4. 4.Clinical notes last, and from a vendor your province or college accepts. We do not build scribes; we build the front-desk side.