A completed order is an attended appointment inside a fixed window, counted over every eligible order issued in the period

Every vendor in this market publishes a percentage. Almost none publishes the denominator. This page publishes ours, with the exclusion rules, the control arm and the window, so an operations director can reproduce the count on their own extract.

We wrote it because three separate buyers, at three separate institutions, asked the same thing before believing any figure: what is the denominator, what group are you comparing against, how do we define the methodology together. None argued about the result. All three argued about the count.

Why does the denominator decide everything?

A completion rate is a division. The numerator is hard to inflate: the appointment either happened or it did not, and the schedule records it. The denominator is where the argument is won or lost, because every exclusion a vendor allows itself lifts the figure without a single additional patient being seen.

Count the orders that already left the consultation with an appointment, and they enter pre-resolved and drag the rate upward. Drop the patients with no valid phone number in the record, and you have removed the hard half of the job. Two counting decisions, no extra patient seen, several points of rate.

Which orders enter the denominator?

The base population is every medical order issued inside the period that had no appointment by the close of the consultation day. Each one carries its issue timestamp, the patient identifier, the service requested and the issuing site. The rules below run on top of that base.

We publish the count of every exclusion next to the result: raw total, eligible total and the difference by rule. An exclusion that is not published cannot be audited.

What counts as a completed order?

There are four possible lines and each gives a different figure: the patient contacted, the appointment booked, the appointment attended, the result reported. We count the third.

Booking is the easiest number to lift and the one most often published. You can book a slot the patient never asked for, at a time that does not work, and the figure still rises. What rises with it is the missed appointment count, and why booking more the naive way makes a service's no-show worse deserves a page of its own. Anyone reporting bookings is reporting a promise.

We stop short of the reported result for an attribution reason. Between attendance and the report sits the lab's or the radiologist's turnaround, which belongs to the provider's internal operation. Including it would measure their processing capacity under our name.

Attendance is read from the provider's own attendance mark on the schedule, the same one it bills the service against. If their schedule says the patient arrived, the patient arrived.

How do you know Surface produced the completion?

Because a share of the orders from the same period never passes through Surface. Every eligible patient is randomised into one of two arms before anyone touches them, on the last digit of a stable identifier. One arm enters the follow-up; the other stays on the process the institution already had. Both arms are drawn from the same period, the same sites and the same service mix.

Randomisation happens at the patient level: a patient with three referrals falls entirely into one arm, because they cannot receive two different treatments inside the same WhatsApp conversation.

Once assigned, they stay in their arm whatever happens. Bad phone number, never replied, asked not to be contacted, conversation closed with no answer: the order keeps counting where it landed. This is the intention-to-treat principle from clinical trials, where all randomised participants should be analysed in their randomised group, irrespective of compliance with the trial protocol. Pulling the unreachable patients out of the treated arm and leaving them in the control is the cleanest and most common way to manufacture a difference.

Before-and-after is the industry's favourite comparison because it costs nothing: take last quarter, install the system, compare. It overstates, because it inherits the season, any marketing campaign running in parallel, a headcount change in the contact centre, and the regression to the mean of a bad quarter. The Cochrane group that reviews organisation-of-care interventions explicitly discourages including uncontrolled before-after studies, because it is difficult, if not impossible, to attribute causation from them.

Contamination is the risk that shows up on the ground and keeping the arms clean is on us. If the scheduling team works the control cases from the Surface dashboard, or a mass reminder campaign goes out to the whole base, the gap flattens. When it happens we say so and discard the period.

We report two counting units. Per order, each exam counts one: a patient with one MRI and a patient with fifteen blood tests contribute one and fifteen. Per patient, each person counts one. The per-order figure is always lower and lets you budget appointment slots; the per-patient figure describes what the person experienced.

How long does an order get to be completed?

Without a fixed window, "the order was completed" is a claim with no closing date: the numerator keeps growing while the denominator stays frozen, and the rate climbs on its own as time passes. Any vendor reporting "cumulative" completion without saying how long they counted is riding that effect.

We measure from the issue date at 7, 14 and 21 days, and publish all three. The decision window is 21 days because the services that dominate the outpatient mix - laboratory, imaging, check-ups - are executable inside that horizon. Part of what we do is accelerate something that was going to happen anyway, and a long window hides that.

Orders with a clinical horizon beyond the window leave the cohort and are measured in their own. This cut-off is not our invention: the public referral-loop measure in Medicare's quality programme counts only referrals issued on or before 31 October of the measurement year, so that every referral counted had a comparable amount of time to close.

How is the count executed, step by step?

1. Freeze the extract.

   Every order issued in the period, with its issue timestamp, patient identifier, service requested, issuing site and scheduling status at the close of the consultation day. It is the same file you can start with before any integration exists.

2. Apply the rules and keep the discards.

   Every exclusion is counted separately and reported. Raw total, eligible total and the itemised difference sit in the same table as the result.

3. Randomise before anyone is touched.

   Chance picks the arm on the last digit of a stable identifier, and the assignment is recorded before the first message goes out. A whole patient lands in one arm.

4. Freeze the rules for the length of the window.

   No change of protocol, channel or contact hours while the cohort runs. A mid-flight change forces you to close the cohort and open another.

5. Match against the attendance mark.

   It is read from the provider's schedule. Bookings are reported separately, as an intermediate funnel stage, alongside those contacted and those who replied.

6. Report both units and all three windows.

   Twelve cells: two units, three windows, two arms. Plus the n of each arm and the difference. They publish together or not at all.

What do we publish and what do we withhold?

We publish aggregate ranges across deployments, and the method that produced them. Never an identifiable client's figure, never revenue attributed to an institution anyone could deduce, never screenshots of real patient conversations.

We have no client we are yet permitted to name, and we say it that way instead of implying scale. We would rather publish the method than a testimonial. A testimonial takes ten minutes to write; a denominator can be audited.

The rule we apply to ourselves: if a percentage appears on this site without its denominator beside it, it comes from the literature and not from a measurement of ours. Use it as a test against us and against every other vendor.

Under what conditions would Surface's effect be indistinguishable from zero?

A vendor who cannot name its own failure conditions has measured nothing. These are ours, and we walk through them before signing any pilot.

• No supply, no completion. If the schedule has no availability inside the window for the scarce services, we convince the patient and still fail to book them. Surface does not create capacity. At a saturated provider the ceiling is set by appointment slot availability.
• Small denominator. If the provider already books most orders at the desk, there is little left to move and the gap between arms compresses until it disappears.
• Contaminated arms. The same team working both arms with the same tool, or an institutional campaign landing on the whole base. The period is discarded and has to be run again.
• Insufficient sample. With a few hundred orders per arm, a difference of a few points sits inside the noise. The correct response is more volume or more weeks of accumulation.
• Long window. Given enough time both arms converge, because the control patient eventually books on their own. Part of the effect is acceleration, and acceleration runs out.
• Dirty numerator. If the schedule's attendance mark is unreliable, the numerator is noise in both arms and no figure on this page means anything. This is caught before measuring, by matching the mark against billing.