What 1.58 Million Messages Say About Selling in Instagram DMs
We analyzed 1,580,874 inbound messages across 141 accounts. When leads actually message, how fast they get an answer, how many messages separate a conversation from a booked call, and how many bookings needed a follow-up.
Rather than repeat what everyone says about selling in the inbox, we counted it. Below are four numbers from 1,580,874 messages that leads sent to 141 businesses running Setor AI between 1 February and 31 July 2026.
- 65.45%of inbound messages arrive outside working hours
- 88 smedian time to a reply from the business
- 16lead messages in the median booked conversation, against 3 with no outcome
- 2 in 3booked calls involved at least one follow-up
Method, exclusions and what this data does not tell you are at the bottom. Numbers first.
1. Two thirds of messages arrive after hours
Inbound only, meaning sent by the lead rather than by the business. 65.45 percent landed outside Monday to Friday, 9:00 to 17:00.
| Hour | Share of inbound lead messages | |
|---|---|---|
| 00:00 | 1.73% | |
| 01:00 | 0.71% | |
| 02:00 | 0.37% | |
| 03:00 | 0.31% | |
| 04:00 | 0.37% | |
| 05:00 | 0.57% | |
| 06:00 | 1.3% | |
| 07:00 | 2.77% | |
| 08:00 | 4.1% | |
| 09:00 | 5.17% | |
| 10:00 | 5.73% | |
| 11:00 | 5.99% | |
| 12:00 | 6.1% | |
| 13:00 | 5.96% | |
| 14:00 | 5.79% | |
| 15:00 | 5.69% | |
| 16:00 | 5.74% | |
| 17:00 | 5.92% | |
| 18:00 | 6.09% | |
| 19:00 | 6.38% | |
| 20:00 | 6.75% | |
| 21:00 | 7.05% | |
| 22:00 | 5.91% | |
| 23:00 | 3.51% |
working hours, Mon-Fri 9:00-17:00outside working hours
The curve climbs all day and only breaks after 21:00. The busiest hour of the day is 21:00, and it beats every working hour: 7.05 percent against 6.1 percent for midday, the best working hour on the board. From 17:00 to midnight you get 41.61 percent of everything, more than all eight working hours combined.
The rest splits out as 25.17 percent on weekends and 13.48 percent between 22:00 and 6:00.
We checked whether one large client was driving this. The biggest single organization accounts for 9.4 percent of observations, and the median across 124 accounts with at least 200 messages is 64.05 percent, with quartiles at 58.32 and 68.26. The pattern is consistent across accounts rather than averaged out of two extremes.
A setter working 9 to 5 covers slightly under a third of the traffic in time. The rest waits until morning, by which point the lead has messaged somebody else.
2. Median reply: 88 seconds
We measured 1,044,860 pairs of “lead message, then business reply” across 137 accounts, discarding any gap longer than a day, because past that it is not a reply, it is a follow-up.
- 87.8 smedian time to a reply
- 32.3%answered within a minute
- 87.4%answered within five minutes
- 93.2%answered within fifteen minutes
The ninetieth percentile is 512 seconds, eight and a half minutes. Even the tail of the distribution lands while the other person still has the phone in their hand.
One honest caveat. In the data, the “from business” flag does not distinguish an AI reply from a human who took the conversation over. This is the account’s response time, not the model’s. On accounts running an AI setter the overwhelming majority will be automated, but this data cannot separate the two and we are not going to pretend it can.
3. A booking is a conversation, not a message
This one surprised us most. We compared how many messages a lead sends in conversations that ended with a booked call against ones that stalled.
| Conversation type | Median lead messages | |
|---|---|---|
| Ended in a booking | 16 | |
| Disqualified | 4 | |
| No outcome | 3 |
The median booked conversation carries 16 messages from the lead. No outcome: 3. Disqualified: 4. On the business side, 25 and 7 respectively.
That is a fivefold gap, and it contradicts the popular picture of DM automation. Booking a call is not the result of one well-written message or a clever flow with a link in it. It is the result of a conversation that runs for a dozen or more turns.
A tool that sends one excellent message does not solve this. What solves it is something that can hold a sensible conversation across sixteen exchanges.
Be careful with the direction of that relationship. We are not claiming longer conversations cause bookings. It is more likely that both follow from genuine interest on the lead’s side. The practical conclusion holds either way: your process has to survive a long conversation, because short ones do not book.
4. Two thirds of bookings needed a follow-up
We define a follow-up strictly, from messages alone: a message from the business preceded by another message from the business that the lead never answered, separated by a gap of at least thirty minutes. That gap excludes multi-part sequences within a single turn.
| Booked | Not booked | |
|---|---|---|
| conversations | 10,709 | 207,010 |
| at least 1 follow-up | 66.19% | 59.59% |
| at least 2 follow-ups | 35.27% | 29.78% |
| at least 3 follow-ups | 17.82% | 13.27% |
| mean follow-ups | 1.37 | 1.15 |
Two thirds of booked conversations passed through a moment where the lead went quiet and somebody came back to them. A third needed that at least twice.
And again, honestly: the gap between the groups is small, six and a half percentage points. That is not evidence that follow-up causes bookings, because booked conversations run longer and therefore have more opportunities to contain one. What is strong is the absolute fact: two thirds of the conversations that worked contained a silence somebody had to break. A process without follow-up concedes those two thirds.
Following up without sounding like a pest is its own craft, and the stage-by-stage breakdown sits in what an AI setter does in your inbox.
5. The same inbox before and after installation
The four numbers above describe the market. Separately we measured what happens on a single account when the same inbox moves from manual handling to automated.
The cutoff is the first trace of Setor on the account, not the date the official API was connected. That distinction turned out to matter: four of the five accounts studied were using our extension months before the API connection, so a window measured from the API date would compare a tool against the same tool. The before window is the year preceding that point, the after window is the last 90 days.
| year before Setor | last 90 days | |
|---|---|---|
| people answered within 24h | 53.3% | 96.7% |
| median time to reply | 2.3 h | 52 s |
| 90th percentile time to reply | 18.2 h | 1.7 min |
| replies within 5 minutes | 22.2% | 97% |
| median reply outside working hours | 2.4 h | 51 s |
| conversations with a follow-up | 17.6% | 71.1% |
| booked call rate | 4.3% | 6.7% |
This is one account, the only one where every metric is measurable against a clean window: 1,479 conversations in the baseline period and 5,462 today. The booking rate rests on a narrower sample, 1,301 conversations before and 1,500 after, because the classifier needs a conversation with real content.
Almost every second person writing in got no answer within a day, follow-up barely existed, and a reply outside working hours took over two hours. With those three things fixed, the booked call rate rose by half.
Booked calls from before installation do not exist in any system, because bookings are only recorded once an account is connected. We established them by reading conversation content with a language model, using the same measure on both sides. Before computing anything we checked the classifier against 500 conversations from the period where the outcome is recorded: precision 98.3 percent, recall 92.8 percent, erring toward undercounting.
The full breakdown, including the accounts where no such comparison is possible, is on the client results page.
What this adds up to
The four numbers describe one picture. Leads message in the evening and at weekends, they need an answer in minutes rather than hours, their path to a booked call runs a dozen or more exchanges, and two thirds of the time it requires returning to a thread that went dead.
None of those four conditions is met by a person working office hours, and that is arithmetic rather than a criticism of people. The same picture explains why rule-based flows fail on higher-priced offers: a script built on keyword matching will not hold sixteen turns of conversation.
More on the role itself and where it fits is in what an AI setter actually is, and the cost side is broken down in what an appointment setter costs.
Method
Every number comes from one set of aggregate queries against our production database, run on 19 August 2026 in read-only mode.
- Window: 1 February to 31 July 2026.
- Sample: 1,580,874 inbound messages, 141 organizations. Response time: 1,044,860 pairs across 137 organizations. Conversation length and follow-ups: 217,719 conversations matched to a contact, of which 10,709 booked.
- Inbound message: sent by the other party rather than the business account, not deleted.
- Hours: Europe/Warsaw, the seller’s timezone rather than the lead’s.
- Working hours: Monday to Friday, 9:00 to 16:59.
- Exclusions: the demo account used for Meta app review.
- No message content, no account names, no personal data, no per-company breakdown.
What this data does not say
It does not say what time leads buy. We measure when a message was sent, not when a decision was made.
It does not say how many of these conversations turned into revenue. A booking is a scheduled call, not a sale, and attendance and close rate are separate data we have not been collecting long enough to publish anything from.
It does not say what time it was in the lead’s own timezone. Hours are counted in the business’s timezone, because it is their calendar that determines whether anyone is at the inbox.
And it does not say how many replies came from AI versus a human. That flag in our database does not separate the two.
Check it on your own inbox
You do not need our data for this. Open your Instagram inbox and go through the last fifty inbound messages. Count two things: how many arrived outside your working hours, and how many of your conversations that ended in a booked call contained a gap longer than half a day.
If the first number lands near two thirds, you are in the market norm. The only question left is who handles that second portion when you are not at the inbox. That is the gap Setor AI closes: it answers every message in seconds, at 21:00 and at 3:00 alike.