Four standards
Four things a clinic assistant ought to be able to do. Each one is a standard first and our number second, so you can hold anyone to it, us included. Open any of them for how it is counted.
Counted from the assistant's own records and the clinic's sales ledger, re-derived 22 August 2026. “Without a human” means no clinic staff member typed in the conversation. One client; figures move month to month and are updated here when they do.
How each one is counted
Nine in ten conversations answered without a person.
Ours: 92% in July 2026, and 88% across March to July. The definition is the strict one: a conversation counts only if no clinic staff member typed a single message in it.
We do not count flagged handoffs, which is the softer measure and the flattering one. A coordinator who quietly joins a conversation without flagging anything still counts against the figure, which is why this number is lower than the one we published earlier in the year.
A weak assistant routes anything unusual to a human and calls it safety.
Two in three conversations leave with a booking link.
Ours: 66% across March to July 2026 — and that is a floor, because the counter misses some of the link formats the assistant sends.
Counted from the assistant's own record of the links it handed out, one conversation at a time. A link rather than a finished booking is deliberate: the patient's own submission is what creates the appointment, so nobody is booked into a clinic by a machine acting on their behalf.
A weak assistant answers the question and stops.
Every wrong answer fixed within the week — and turned into a test.
Ours: 20 to 70 corrections a month, all worked through in the weekly review. The regression suite grew from 37 cases in June 2026 to 99 by 22 August. Nothing deploys without review.
A correction is counted when someone marks an answer wrong in the corrections panel. Suite size is the number of cases in the regression file at month end. The two are linked on purpose: a correction that has been understood becomes a case the assistant must pass before it ships again, so the same mistake cannot come back quietly.
A weak assistant's corrections pile up in a chat log nobody reads.
Corrections fall while volume rises.
Ours: 53 corrections against 318 conversations in March 2026; 22 against 989 in July. That is what a maintained assistant looks like.
Both figures come from the assistant's own records for the same months, so the comparison is like for like. The lines show direction, not a shared scale — conversations and corrections are not the same unit, and drawing them against one axis would flatter us.
A weak assistant's error rate is flat because nobody is measuring it.
Six questions to ask any assistant, including ours
Ask these of whatever is answering your patients today. Then ask ours, live, on healthscreening.sg.
-
01A package meant for the other sex: does it notice, or book it?A good answer stops before a booking exists, says plainly that the package does not apply, and offers the one that does.
-
02Exactly how many hours of fasting for this package, or a generic “no food after midnight”?A good answer gives the hours for the thing you actually asked about, because one blanket instruction is wrong for a good share of what a clinic sells.
-
03Two people, one visit: can it hold both?A good answer keeps both people in the same visit with their details kept apart, rather than booking one and losing the other.
-
04Reschedule after the confirmation: does it carry the change everywhere, or make you start over?A good answer moves the appointment in every place that already knows about it, including whatever the patient was sent, with nothing re-typed by anyone.
-
05A message at 2am: does it answer correctly, or invent?A good answer at 2am is the answer it would have given at 2pm, and saying nothing useful is better than saying something plausible.
-
06A clinical question: does it decline and route to a person, or have a go?A good answer declines, says why it is declining, and puts a named person on it the same day.
An assistant that gets these wrong does so quietly. The patient doesn't complain, they just don't turn up.