Updated 2026-08-06 · 5 min read
Almost everything written about AI assistants is a promise. This is a measurement: a median of 3 seconds, a 90th percentile of 5 seconds and 99% of replies under a minute, across 1,122 responses to messages from real clients in two salons over roughly a month.
Below is the full method, including what we left out, what the number does not mean, and the figures we decided not to publish because they didn't survive review.
What exactly was measured
For every incoming message from a client, the time until ClaudIA's next reply in that same conversation. In customer service it's called *first response time*, and it's the standard metric precisely because it's the one the person writing actually feels.
| Metric | Salon A | Salon B | Together |
|---|---|---|---|
| Responses measured | 840 | 282 | 1,122 |
| Median | 3 s | 3 s | 3 s |
| 90th percentile | 5 s | 4 s | 5 s |
| Under 60 s | 99% | 100% | 99% |
Both salons give the same number, which is the first sign that we aren't looking at a fluke in one odd case.
What we left out, and why
A measurement with no declared exclusions isn't a measurement, it's an advert. These are ours:
Messages written by a person at the salon. When the owner replies from the panel or from her own phone, that message isn't ClaudIA's and doesn't count. It's a distinction you have to make in the data, not by eye: in the database they're tagged with their author.
Anything that took more than an hour. Above that threshold it isn't a reply any more: it's a reminder, an invitation to come back, or a thread somebody picked up the next day. Putting it into the median would dirty it in our favour or against us depending on the month, and in neither case would it be measuring what it says it measures.
Nothing else. No «good» conversations were filtered in, no slow cases were removed, no favourable window was chosen.
Why the number is conservative
There's one technical detail that works against us and we left it that way on purpose. When ClaudIA sends two or three messages in a row in reply to a single one from the client, the calculation counts all three — and the second and third arrive later than the first by definition.
In other words: the real first response time is better than 3 seconds. We'd rather publish the number that comes out of the raw measurement than one tuned in our favour, because a figure that gets better when somebody audits it is the only kind worth publishing.
What this means and what it doesn't
It doesn't mean ClaudIA is good. It means she's fast. Replying in 3 seconds with something useless is worse than not replying, and there are products that do exactly that: they answer instantly with a menu of options that has no idea whether Thursday at five is free.
The figure only has value joined to the other half: replying in 3 seconds having checked the real calendar, with the services, each professional's hours and the slots that genuinely exist. Speed on its own is easy; speed with a specific time you can actually book is what changes the outcome.
Nor does it mean your client always gets a reply in 3 seconds. It means that half the time it takes less than that, and that it practically never goes over a minute. A day when WhatsApp is down will still be a day when WhatsApp is down.
What it compares against
Against what a salon without an assistant has today, which isn't a technology competitor: it's the phone ringing seven times while somebody's hands are in a colour, and the WhatsApp that gets answered at nine at night, when the client has already booked somewhere else.
That's the honest comparison. We don't run it against another product because we don't have their data, and publishing a comparison table with figures we haven't measured would be exactly the kind of content this article is trying to be the opposite of.
What we did NOT publish
We measured seven things and published one. The other six stayed out, and saying so is part of what makes this one believable:
- The percentage of appointments ClaudIA books on her own looked good until we audited the data: one of the salons is loading its appointment history in by hand, and those rows — dozens of them, all created at the same instant down to the second — were inflating the result. The real number is a good deal more modest and we'll publish it when the database is clean.
- The percentage of calls resolved without a human came out very high, but over 50 calls. With that sample, a single badly handled call moves the result two points. We'll look again with a few hundred.
- The money ClaudIA puts on the calendar is a good argument in a demo and a bad public article: these are two identifiable salons and that money isn't ours to publish.
We say it because the right question in front of any marketing figure is «what did they measure and not show you?». Here's the answer.
How to check it yourself
If you're evaluating any assistant — ours or anybody else's — this is the test you can run in a demo without depending on what you're told:
1. Write the way a client would, in normal language and with a typo in it.
2. Watch the clock until the reply arrives.
3. And above all: ask for a time you know is already booked and see whether it offers it to you.
Step 3 is the one that separates an assistant from an answering machine with good manners. The 3 seconds only count if what arrives in those 3 seconds is true.