Agency · 10 work groups
Nothing gets lostin work chats.
The agency films and runs content for its clients. All the work happens in Telegram: ten client groups, direct messages, voice notes, photos of edits. The project manager can’t read it all and hears about a problem when the client is already unhappy. We built an AI assistant that reads the chats instead and sends short lists: what’s on fire, who promised what and which client is waiting for a reply.
Live · since 13 October 2026 · an agency and the head of a client company
🔥 On fire: overdue, today, remarks
📎 Show message✅ Done
🔴 Edit two videos for “North Studio” — Boris · due yesterday
🟠 Send the Reels script to “Corner Café” — Kate · today
⚠️ “Corner Café”: the old logo is back in the video
🆕 Looks done: VK covers for “North Studio” — please confirm
⏳ Client waiting for a reply: North Studio · 3 working h
📎 Show
Anna: when will the final cut be ready?
- 10client work groups processed by the assistant
- 3×a day it sends what’s new11:00, 16:00 and evening
- 2 hwithout a reply and the client is flaggedworking hours; red after 4
- 1 tapto the source message for every item
Everything important is in chats, and there’s too much of it
A client posts a change to a video in the group. An editor promises by voice note: “done by Wednesday”. A week later nobody remembers the promise, and the client is asking for the third time.
A task tracker doesn’t help if nobody copies things into it. A manager can’t copy dozens of messages a day by hand. So the assistant doesn’t ask the team to change habits: people write where they always did, and it reads and collects.
Work chats
Read-only: the assistant never posts in work chats
Everything to text
AI finds
To a person
- 3 times a daywhat’s new: remarks, promises, “looks done”
- Every hour“client waiting for a reply” over 2 working hours
- Saturdaysthe week’s meeting agenda as a Word file
Every item has a “📎 Show message” button
Three short lists instead of a hundred messages
Three times a day the assistant sends what’s new: client remarks, promises it found and things that look done. It doesn’t close those itself; it asks for confirmation. “On fire” holds overdue items, today’s tasks and client remarks. “Stuck” holds anything in progress for more than three days without movement.
Every item has a “📎 Show message” button. The model can word something wrongly, so the source is always one tap away.
🔥 On fire: overdue, today, remarks
📎 Show message✅ Done
🔴 Edit two videos for “North Studio” — Boris · due yesterday
🟠 Send the Reels script to “Corner Café” — Kate · today
⚠️ “Corner Café”: the old logo is back in the video
🆕 Looks done: VK covers for “North Studio” — please confirm
🎤 0:09 · “What’s open for North Studio?”
“North Studio” has 3 open:
📎 Show message
🔴 two videos — Boris, was due yesterday
🟡 VK covers — Kate, by Friday
⚪ agree the November shoot — no date
A client is waiting — the assistant will say so
During the day, once an hour, the assistant checks the latest messages in client chats. If a client asked something and has had no reply for more than two working hours, a flag goes out. After four hours it turns red. The team’s internal chats aren’t counted.
On Saturday mornings the assistant compiles the week’s meeting agenda and sends it as a Word file: what was done, what’s overdue, what’s stuck and which remarks clients made.
⏳ Client waiting for a reply: North Studio · 3 working h
📎 Show
Anna: when will the final cut be ready?
Ask by voice
You can ask the assistant by text or voice: “what’s open for this client?”, “what did Kate promise?”. You can set a task in one sentence: “ask Boris to edit two videos by Wednesday”. You can reply to one of its messages and ask “who said this?”.
Voice notes and videos are transcribed on the company’s own computer, without sending audio to cloud services. Text in photos and documents shared in chats is read automatically too.
A strict mode for the head of the company
The same assistant works for the head of a company with escape rooms, a cinema and a theatre, but more strictly. It shows only what needs that person’s own action or oversight, without small edits or other people’s tasks.
It also spots staff payouts in chats. A message like “sent 3000 from Bratskaya” becomes a card with buttons, and once confirmed the payout goes into payroll. Each morning the AI checks the head’s open items against chats and payments and clears what’s already done.
Looks like a payout: “sent 3000 from Bratskaya”
Yes, record itNoAnother employee
Employee: Boris · 3,000 RUB · Bratskaya till
Record it in payroll?
How it was built
How we built it
The assistant reads chats through an employee’s own account and never posts in the groups. Each run takes only what’s new since the last one. The full history, documents included, was processed once on connection.
The rules for what counts as a task were tuned per person. A project manager needs every promise the team makes; the head of a company needs only their own decisions.
- The agencyWhich chats to read, who is on the team, which tasks matter, when to send summaries.
- SemCueChat reading, voice transcription, task extraction by the model, de-duplication, the bot with buttons and questions.
- The managerConfirms “looks done”, sets tasks by voice, works through the meeting agenda.
For technical readers
How it’s built
- Reading
- Telegram via a personal account, read-only; the bot sends messages privately.
- Transcription
- Whisper large-v3 on the company’s own computer, text recognition in photos, text from PDF, Word, Excel and PowerPoint.
- Model
- A language model with no tool access: text in, a structured list out. A second pass re-checks and merges.
- Schedule
- Reviews 3 times a day, a “waiting for a reply” check hourly in working hours, an agenda on Saturdays.
- Links
- Tasks sync with Asana; payouts go to payroll after confirmation.
Where the numbers come from
Assistant description and run log
- Assistant description and run schedule, October 2026.
- Decisions on the strict mode and chat payouts, 14 October 2026.
- Messages on this page are typeset in the assistant’s format with fictional data.
Next system
AI business diagnostics for a consultant →