Most numbers about AI chatting come from the vendors selling it, often without saying where they come from. Here are ours, taken straight from our production database. They cover three AI personas we run ourselves, almost entirely on Telegram. The numbers are aggregated and no persona is named.
The numbers, 21 September to 20 October 2026
| Measure | Value |
|---|---|
| Fans who wrote | 441 |
| Messages received from fans | 40,056 |
| Messages sent by the AI | 29,304 |
| Messages sent by a human | 2,268, or 7% of messages sent |
| AI cost for the chat | $129, about $4.30 a day for the three personas |
| Cost per message sent by the AI | about 0.44 cents |
| Cost per fan who got an AI reply | about $0.35 for the month (371 fans) |
| Paid offers made | 653 |
| Sales | 62, by 25 buyers: 61 in Telegram Stars (30,500 ⭐) and 1 by card (€6) |
| Sales with no human message in the previous 24 h | 46 of 62, or 74% |
The cost covers everything the chat uses: writing replies, checks before sending, fan memory, reading the photos fans send. It does not include social posting or hosting.
What it earns, next to what it costs
The 30,500 Stars are worth about $400 to the creator, at the withdrawal rate of about $0.013 per Star reported by third-party guides (Postmypost); the rate varies and is fixed when you withdraw. So the AI cost about a third of what it brought in over the period.
That is an order of magnitude, not a promise: three personas, one month, and audiences that are still young. But it is the question to ask any tool: what does it cost, next to what it sells?
Where the money goes
- 74% to write the replies. That is the core of the work.
- 19% to check every reply before it goes out, for example that it does not give away for free what is paid. A fifth of the budget, and we keep it: a reply that goes wrong costs more than a check.
- 6% for memory: keeping what each fan says.
- Under 1% to read the photos fans send.
What we learned
81% of the cost goes to fans who never buy
That is the real expense: fans who talk a lot and never buy. Not ignoring them, but replying at a slower pace and without an offer on every turn, is where profitability is decided.
People buy after a real conversation
Buyers sent a median of 44 messages before their first purchase, from 2 for the quickest to 380 for the most patient. Offering media in the first exchanges rarely sells: first a relationship, then a desire expressed by the fan.
About one offer in ten gets paid
653 offers, 62 sales: 9.5%. A declined offer is not a failure as long as it does not spoil the conversation. That is why a no should close the topic, with no sales follow-up.
Humans are still useful
The AI sent 93% of the messages, and three sales in four happened with no human writing in the 24 hours before. The last quarter had a human step in: an AI chatter does not replace the whole team, it leaves them the fans who matter and the delicate cases.
The limits of these numbers
- Three personas, one month: a sample, not a market average.
- These are our own personas, not customers.
- Cost depends on the AI models used and their prices, which change.
- We know of no independent study on what a team of human chatters costs: compare with what you pay yourself, nights included.
To see how Klovae sells without pushing, read the Telegram AI chatbot page.
Updated 21 October 2026