Task and reason
Task: hand the company’s daily operational work to AI agents: taking in clients, preparing documents and texts, publishing, bookkeeping. The condition: any result can be checked, and decisions about money and irreversible actions stay with people.
Reason: the two owners of OM-Digital face dozens of client and internal tasks every day. One AI chat for everything does not solve the problem: it has no role, no queue, no memory of what has been done and no limits of authority. Mistakes are costly. In August 2026 an agent’s report of success failed to match reality three times: a reply to a client was marked as sent but was not in the conversation; a listing was marked as published but could not be found on the platform; an assistant returned “done” within a few seconds without changing a single file. It became clear that what was needed was not a lucky prompt but a system with roles, checks and a log.
Since June 2026 the main operational work at OM-Digital has been done by AI agents, while people set the direction and make decisions. Below is the order in which we built this: six steps that any small company can repeat.
Many are trying. According to a study by Yakov and Partners and Yandex (December 2025, a survey of 150 technical directors from 16 industries), 71% of Russian companies use generative AI in at least one function, and 46% have deployed or are testing AI agents. The winners are those who thought through roles and control in advance.
Short answer
An AI agent reliably handles routine work that can be checked: sorting incoming requests, preparing documents from templates, reconciling data, publishing on schedule, watching logs. Decisions about money and commitments, the last step outward and irreversible operations stay with a person. Results come not from the strongest models but from defined limits of authority, an action log and checking the outcome on the recipient’s side.
How an AI agent differs from a chatbot
A chat answers a question. An agent performs a task. It has a role with limits of authority, a queue of incoming tasks, access to tools (files, mail, a database) and a log where every action is recorded. It does not wait to be asked: it takes a task from the queue, does it and reports back.
Step 1. Dividing the work into roles
We separate roles by type of task. We have separate areas responsible for production (texts, websites, code, analytics), coordination, client intake, sending messages to external platforms, advertising and promotion, incident analysis, finance and legal documents.
The scheme took shape gradually. On 10 July 2026 we made client intake a separate role: before that, incoming requests sat in one queue with production and waited for a long task to finish. After the split, a separate area answers the client, and the queue moves evenly.
Step 2. Setting up a shared task queue
Tasks move between roles through a shared queue. Each record has a type, a recipient, a priority and a status, and each role reads only its own types. There is no overlap by design of the system, not by agreement. Over the period of operation 29,772 records have passed through the queue, 9,058 of them work assignments.
Step 3. Limits of authority
The first zone is decisions about money and commitments. Price, payment terms, deadlines and the signature on a contract are approved by a person.
The second is the last step outward. Sending to a client is done by a separate role, and never by the one that prepared the material. This way every document is seen by a second participant in the process before it goes out.
The third is anything where the cost of a mistake exceeds the cost of the work: operations on live infrastructure, deleting data, any irreversible actions.
Research supports this division too. Economists from Stanford and MIT tracked 5,172 customer support employees with an AI assistant that suggested answers but left the decision to the person: the number of resolved requests per hour rose by 15% on average, most of all among newcomers.
Step 4. An action log
Every significant action is recorded in a shared archive: who did what, when and for which order. This is the company’s memory. On 17 September 2026 the log received 306 records from seven roles, and a week later any of them can tell you where a figure in a report came from or why a decision was made.
Step 5. Checking the result on the recipient’s side
After the August cases we introduced a rule: a task counts as closed not when the action ran without an error, but when the outcome is confirmed. A message to a client is re-read in the conversation itself, a publication is found on the platform, a file is opened and checked against the task. If an assistant finished in a few seconds on a task to edit a file, someone checks whether the file changed.
In addition, before delivery every text goes through an automated check against the company’s list of banned wording: the script checks each text the same way regardless of workload. Any negative conclusion from an agent (“this does not exist”, “this does not work”) is checked by hand before it goes into a document: an agent reports quickly and confidently, and the check takes minutes.
Step 6. Watching the system itself
A separate check makes sure the roles are working and the queue is moving. Silence in the log while the queue is not empty means a failure, even if nobody has reported it.
Where to start for a company looking at agents
With one role and one process that repeats most often and is done the same way every time. Write down what counts as “done” and assign a check: a person, a second agent or automatic reconciliation. Within a month you will see how much time the role frees up and where it makes mistakes. It only makes sense to scale what has passed such a check.
Key takeaways
- Problem: an agent reports success that is not there, and one chat for everything provides no roles, no memory and no limits.
- A six-step solution: roles, a shared queue, limits of authority, a log, checking the outcome with the recipient, watching the system.
- Money, outward sends and irreversible operations stay with a person.
- Measurable result: 306 log records in a day and 29,772 queue records in four months.
- It is better to start with one role and scale after a month of measurements.
Result
- Since June 2026, for four months, daily tasks have been done by seven AI agents with written procedures, while people set the direction and make decisions.
- In a single day, 17 September 2026, the shared log received 306 records from seven working roles: for every action you can see who performed it, when and for which order.
- From 17 June to 24 October 2026, 29,772 records passed through the shared task queue, 9,058 of them work assignments.
- Client intake became a separate role on 10 July 2026: incoming requests no longer wait for a long production task to finish.