BUline runs its own work on an AI-Ofis of 24 specialized agents. An employee writes to a single Telegram bot, and the system routes the task to the right agent. Drawing on this experience, we build multi-agent offices for clients too.
The AI-Ofis is 24 agents built on the Claude model. Each has written instructions for one role, its own tools and a shared knowledge base. There is a single entry point: a Telegram bot. Employees send text, files or voice messages. A voice message is first transcribed and shown to the employee for confirmation; the agent starts only once it is confirmed.
Every task is logged: who assigned it, which agent completed it and what the outcome was.
The employee types a task to the bot or sends a voice message.
The speech recognition service produces the text, and the bot shows it to the employee, who corrects any errors.
The system identifies the role from the task content: a contract goes to the legal agent, a report to the finance agent, a translation to the translator.
The agent relies on internal documents, requests data from another agent when needed and prepares the result.
Any text or document going outside the company, and any change in a system, happens only after the employee approves it.
Rules we arrived at while building our own AI-Ofis.
One agent, one job. An agent that “knows everything” makes more mistakes and is harder to check.
Employees don’t have to memorize a list of agents. Routing is the system’s job.
Neither voice commands nor external actions run without approval.
Core company information is in one knowledge base; role-specific documents are in each agent’s own section.
Only leadership can access the finance and legal agents.
Which agent is used most and where errors cluster: instructions are updated based on this.
You don’t need 24 agents right away. We start with the 3–5 roles that take up the most time.
Separately for each agent and overall for the whole system.
| Metric | Calculation |
|---|---|
| Active users | Share of employees who message at least once a week |
| Accepted without edits | Share of results the employee accepted without changes |
| Misrouting | Share of tasks reassigned to another agent |
| Voice recognition errors | Share of transcripts corrected by the employee |
| Turnaround time | Median time from task assignment to result |
“We test every type of agent we offer clients in our own work first. That is how we know where it is strong and where it goes wrong.”
The BUline teamIn Uzbekistan, most employees are on Telegram all day. There is no new app to install or learn, and voice messages are already a habit.
Yes, when needed. For example, the legal agent may ask the finance agent for an amount for a contract. But the final result still goes to an employee for approval.
That is what the confirmation step is for: the bot shows the text, and the employee corrects or rewrites it. Key details such as names, amounts and dates are highlighted separately.
It depends on the task: Odoo, amoCRM, MoySklad, Google Drive, email. Read access comes first; write access is granted after testing and once the approval chain is set up.
It is useful even for a company of 10–20 employees, but start with 2–3 agents. More agents mean more support.
We review your processes and recommend a specific task, KPI and plan for your first agent.