How an AI agent confirms appointments, and when it steps aside

Remi is the AI agent we built to confirm appointments for clinics. At a four-physician ENT clinic in Ontario, over the 30 days ending September 9, 2026, it confirmed 833 of the clinic’s 901 confirmed appointments: 92%. The front desk confirmed the other 68 by hand. In the same window, 102 patients told Remi they could not make it, early enough for the clinic to rebook the slot. The full case study has the method and every figure.

Confirming by hand is a phone job: call the patient, leave a voicemail, try again later, note the answer. This post explains how Remi does that job instead, how it reads a patient’s reply, and exactly when it hands the patient to a person.

How does an AI agent confirm a patient’s appointment?

Text first, then call, and stop at the first answer. Each patient gets a text and an email about their visit, sent from a number the clinic owns and written in the clinic’s own words. Patients who do not reply get a phone call in a natural voice, placed during the calling hours the clinic has set. If the call reaches voicemail, Remi leaves the clinic’s message. The answer is written straight to the schedule, and only the exceptions reach staff.

At the ENT clinic, that meant 2,957 outreach attempts in 30 days, 2,397 texts and 560 calls, and not one of them made by a person. 964 patients replied by text, 40% of texts sent, and 185 answered a call, 33% of calls placed. Per 100 sent, texts produced 30 confirmations and calls 21. The calls matter because they reach the patients who do not text back. The average time to an answer was 30 hours, after 1.2 rounds of outreach.

Visits further out follow a longer plan built around the date of the appointment. Visits booked at short notice follow a quicker one. In both cases the clinic decides the timing and the wording.

How does Remi read a patient’s reply?

It sorts every reply into one of four outcomes: confirmed, unable to attend, wants to reschedule, or has a question. “Yes.” “See you then!” “I can’t make Tuesday, is Thursday possible?” “Which entrance should I use?” Each one updates the appointment straight away, so the schedule always shows the latest answer.

A plain “Yes” is recognised by a few lines of ordinary code, with certainty. The language model is kept for the replies that need it. At the ENT clinic, four in ten texts got a reply, and three of every four replies were a confirmation.

When does an AI appointment agent hand a patient to the front desk?

Whenever the next step needs a person, and at no other point. Remi’s job is narrow on purpose: it establishes whether a patient is coming, records the answer, and gets out of the way. It does not reschedule, because choosing a new slot means weighing the clinic’s calendar against the patient’s needs, and that belongs to the front desk. It does not answer questions beyond the schedule, because where a scheduling question ends and a clinical one begins is not for an agent to judge, so nothing outside the schedule goes through it. It does not decide what to do about a patient who has gone silent close to their visit, because the right move depends on who the patient is and what the visit is for. In each of those cases Remi stops, hands the patient over with the conversation so far, and the person at the desk picks it up with nothing to reconstruct. These are the four triggers:

• A patient cannot attend. The front desk is told immediately, while there is still time to offer the slot to someone else. At the ENT clinic this caught 102 cancellations in 30 days.
• A patient wants a different time. Staff handle the reschedule and receive the request with the conversation attached.
• A patient asks a question. Remi speaks only about the schedule. Anything else, and anything clinical, goes to the team.
• A patient has not responded close to the visit. The front desk decides what happens next.

Once a patient has answered, Remi is done. One clear answer is enough, and the patient is not contacted about that visit again. If a phone number cannot be reached at all, Remi tells the front desk so the record can be corrected.

What does the front desk see?

A calendar of every visit and where it stands: confirmed, awaiting a reply, being contacted now, or in need of a person.

Each appointment has its own timeline with every text, email, call and reply in order. Anyone at the desk can open a visit and see the whole story in a few seconds, pick up where Remi left off, or cover for a colleague.

How does Remi avoid contacting the wrong patient?

It never chooses who to contact. Remi identifies the appointment, and the patient’s number or email is looked up from the clinic’s own schedule. A patient is only ever contacted about their own visit. The rest of the design follows the same rule, that facts come from the clinic’s records and not from the agent:

Everything runs on the clinic’s clock. “Three days before the visit” and “during calling hours” mean what the front desk means by them.

Staff instructions take priority. If a coordinator adds a note for a particular patient, that note shapes the message the patient receives.

The schedule is checked before anything goes out. Clinics share the schedule they already produce, and Remi verifies it against the schedule’s own totals before any outreach begins.

Nothing is re-keyed. At the ENT clinic, Remi is wired into the booking system: it reads the schedule and writes each confirmation back. Valaren wrote the scripts for the clinic’s ENT bookings and onboarded and trained its coordinators.

What does this change for a clinic?

The day starts with a short list of patients who need something, not a long list of calls. At the ENT clinic, Remi confirmed about 28 appointments a day, and every coordinator on the account uses it. We estimate that returns about 70 hours of coordinator time a month, at five minutes per confirmation.

In September 2026 the clinic added a fifth physician and its schedule roughly doubled, to 2,064 visits from 1,012 in August. Remi’s share of confirmations held at 90%, and the median reply came within 23 minutes.

Patients get a clear message they can answer in a few seconds, and when they need a person, they reach one. The front desk keeps every conversation that needs it.

Every figure is the clinic’s own Remi dashboard: for the 30 days ending September 9, 2026, and for September the visits dated September 1 to 30. The hours figure is an estimate at five minutes per confirmation; everything else is counted. More on Remi and our work in healthcare.