Short answer. A working AI team is not one chatbot that was “given access to everything”. It is three things: roles with clear boundaries, a shared task queue through which the roles talk to each other, and a log in which every action is visible. Remove any one of the three and you get a demo, not a working business.
Why one universal agent does not work
The first idea that comes to mind is to take a strong model and hand it everything at once. In practice such an agent loses the thread. It searches for orders, replies to the client, writes the text and counts the money all at the same time — and at some point answers a client in the tone of an internal memo, or abandons a job halfway through because a new message arrived.
People solve this with a division of labour, and here it works exactly the same way. As soon as a role has one clear area of responsibility and one entry point for tasks, quality jumps.
How our team is arranged
| Role | What it is responsible for |
|---|---|
| Coordinator | The single entry point: takes tasks from the founders, sets priorities, distributes them across roles and checks the result |
| Production | The work itself: texts, websites, apps, scripts, documents |
| Secretary | Taking on clients, clarifying the task, registering it in the CRM, replying in under an hour, delivering finished work |
| Order sourcing | Working through marketplaces and email, submitting bids, handling incoming messages |
| Marketing | Content plan, publications, site promotion |
| Finance and law | Income and costs, contracts and offers, tax questions |
| Internal investigations | Working out failures: why a task stalled, why a message did not arrive, what to fix in the process |
The last role deserves a separate word. It looks redundant right up until the day it turns out that a message to a client was “sent” according to the report but does not exist in the conversation. That happened to us, and it is exactly why we now have a hard rule: any sending or publishing action is confirmed not by the absence of an error but by checking the end result — open the conversation and see the message sitting there.
A task queue instead of conversations
The roles do not “chat” with each other in free-form text. Every handover of work is an entry in a shared queue: from whom, to whom, what to do, priority, status. A task goes through “waiting → in progress → done”, and at any moment you can see where it is stuck and who is holding it.
This is the boring part of the design, and it is also the main one. Free-form correspondence between agents looks impressive in a demo and falls apart in the second week: tasks go missing, two roles do the same thing, and working out afterwards who decided what is impossible.
The log: not one action goes unrecorded
Every significant action — a reply to a client, delivered work, a fault found, a decision taken — goes into the company's shared log. Through it you can pull up the history at any point: what happened on a particular order, when the client replied, why the price is what it is.
An AI employee has no memory of yesterday unless it is written down. The log is the company's memory: without it every day starts from zero.
What changed in practice
- A reply to a client in under an hour, at any time of day. For enquiries where the person is already ready to pay, a tighter mode applies: every channel is checked every fifteen minutes for two hours.
- Work does not stop in the evening. The company runs around the clock, as long as there is a connection.
- Nothing gets lost. A client who wrote three weeks ago and went quiet does not fall through: the system reminds us about them by itself.
- The cost of an hour of work stopped being a constraint on routine — checks, rounds of the marketplaces, reports, proofreading.
Clear limits: what our AI does not do
This section matters more than the previous one, and most accounts of AI-run companies leave it out. Here is what stays with people:
- Decisions about money. Real payments, price changes, discounts beyond the usual — founders only.
- Confirming payment. That the money has arrived is confirmed by a human in their own banking app. A screenshot from the client does not count as proof.
- Irreversible actions. Anything that cannot be undone goes through a person.
- Disputes with a client. A conflict, non-standard terms, legally significant questions — escalation, not an independent decision.
- Work that cannot be done automatically. We openly decline orders that mean permanently running someone else's accounts by hand: refusing is better than promising.
Limits are not a sign of a weak design but the condition on which it can be trusted. A system with no forbidden zones will sooner or later do something irreversible.
How much of the client's time this takes
We work through the company's processes, design the roles, write the rules, set up the shared task queue with a coordinator, the log and the reporting — and hand it over with training. Rollout takes 30 days and starts at 235 000 ₽. The detailed scope is on the page for building an autonomous AI organisation.
It makes sense to start not with the hardest area but with the most repetitive one: handling incoming enquiries, preparing standard documents, first-pass checking of data. There the result shows in the very first week, and it is easier to agree where to go next from there.
Do AI employees replace people?
They take over repetitive work: rounds, checks, first replies, preparing standard documents. Decisions about money, disputes and everything irreversible stay with people — for us that is written down as a rule, not as an intention.
What does a company need in order to adopt this?
An understanding of its own processes and a person who can answer questions about them. Perfect documentation is not required — working through the processes is part of the job.
How many roles are needed at the start?
Usually three or four: intake, production, a coordinator and the log. The rest are added once it becomes clear where things actually get stuck.
What happens if the AI makes a mistake?
There will be mistakes; the question is whether they are visible. That is why the design includes a log, a check of the end result of every sending action, and a separate role that investigates failures. A mistake spotted the same day is cheap.
Does this only work for IT companies?
No. The design is not about programming but about each role having one area of responsibility and a shared task queue. Work in retail, services or manufacturing is put together the same way — what changes is the content of the roles, not the structure.