An account that looked fast until somebody measured the tail
The case where the median looked fine and only the ninetieth percentile told the truth.
Calls booked per month
421.7
1,265 in the last 90 daysAverage over the last 90 days, counted from bookings that carry an actual scheduled call time. No such figure exists for the period before installation, because nobody was recording bookings then.
- Ninetieth percentile time to reply5.2 h1.7 min
- Replies within 5 minutes52.2%98.3%
- People who got a reply within 24 hours83.9%98.2%
A training and nutrition program sold online to people who want to change their physique without a rigid diet or daily gym sessions. Traffic comes from organic content, so the inbox receives thousands of story replies and comment responses, only a fraction of which turn into an actual conversation.
A 3.6 minute median looked fine. Only the ninetieth percentile revealed that every tenth lead was waiting more than five hours.
The starting point: good median, terrible tail
This account was not neglected. The median reply time before installation was 3.6 minutes, better than most companies that measure this at all. Looking only at the middle of the distribution, there would be nothing here to fix.
The ninetieth percentile, however, was 5.2 hours. In other words: nine conversations in ten were handled quickly and every tenth waited more than half a working day. Replies within five minutes covered 52.2 percent of messages, and 83.9 percent of people writing in got an answer within 24 hours.
What the measurement showed afterwards
The median fell from 3.6 minutes to 59 seconds, but the second number matters more: the ninetieth percentile dropped from 5.2 hours to 1.7 minutes. The response time distribution stopped having a tail. Replies within five minutes rose from 52.2 to 98.3 percent, and the share of people answered from 83.9 to 98.2.
Follow-up moved too: from 44 to 68.2 percent of conversations, at 1.58 reminders per conversation instead of 0.81. The booked call rate for the last 90 days is 10.1 percent, with a confidence interval from 8.7 to 11.7 percent across 1,499 classified conversations.
What cannot be computed from this account
We do not report a booked call rate for the period before installation. In the entire year preceding the cutoff, only 27 conversations contained at least three messages with real content, which is what the classifier needs in order to judge whether a call was booked. The rest were one-off prompts that went nowhere.
At a sample of 27 the confidence interval runs from 3.9 to 28.1 percent. We draw it on the chart as a dashed bar next to today's, so you can see why we use one of those numbers and not the other.
Why this is not one hundred percent
Two groups are left without a reply, both of them on purpose. The first is friends and accounts the client themselves follows: the tool has a setting to leave them alone, because nobody wants an agent pitching their mates. The second is contacts rejected immediately against the disqualification criteria the client sets. On top of that there are conversations a human took over and never finished. One hundred percent on this metric would mean the tool answers everybody indiscriminately, and that is a defect rather than a feature.
Every metric we measured
The full table, including the metrics that did not improve. Sample sizes sit below it.
| Change | Year before Setor | Last 90 days | Change |
|---|---|---|---|
| Ninetieth percentile time to reply | 5.2 h | 1.7 min | 180× faster |
| Replies within 5 minutes | 52.2% | 98.3% | 46.1 pp |
| People who got a reply within 24 hours | 83.9% | 98.2% | 14.3 pp |
| Median time to reply | 3.6 min | 59 s | 3.7× faster |
| Median reply outside working hours | 2.6 min | 58 s | 2.7× faster |
| Conversations with at least one follow-up | 44% | 68.2% | 24.2 pp |
| Follow-ups per conversation | 0.81 | 1.58 | 2× more |
| Conversations held | 614 | 13,014 | 21× more |
What each of these numbers actually measures
- Ninetieth percentile time to reply
- The boundary of the slowest tenth of replies: nine in ten arrive faster than this.
- Replies within 5 minutes
- Share of pairs where the reply arrived within five minutes.
- People who got a reply within 24 hours
- Share of conversations where a person wrote something in text and received a reply within 24 hours of their first message. It counts people, not messages.
- Median time to reply
- The middle of the distribution of time between a lead's message and the business reply that follows it. Pairs longer than 24 hours are dropped, because that is a follow-up.
- Median reply outside working hours
- The same figure, computed only for messages outside Monday to Friday 9:00 to 17:00, in the seller's timezone.
- Conversations with at least one follow-up
- Share of threads in the window where the business wrote again after at least thirty minutes without a reply. The denominator is every thread, single-message ones included, so the figure is understated on both sides of the comparison.
- Follow-ups per conversation
- Average number of such reminders per thread.
- Conversations held
- Conversations in the window with at least three messages where both sides spoke.
What these numbers rest on
Before window
March 2025 → March 2026
614 conversations in window
3,309 reply pairs
After window
Last 90 days
13,014 conversations in window
69,445 reply pairs
Cutoff point
March 19, 2026
The first trace of Setor on this account.
A conversation counts when the lead wrote something in text. That is the denominator for every percentage on this page apart from the time based metrics. A pair is a lead's message and the business reply that follows it. That is the denominator for the time based metrics.
How we counted thisSee the report on 1.58 million messagesInstagram DM automation for coaches
Want to see these numbers on your own account?
On the call we go through your inbox and show you how many messages currently go unanswered.