300 K2Tech employees work with AI agents: connectors alone are not enough—you need ground rules

Until recently, having AI write an email was enough. Now the frustration is different: it does not know whom to send it to, does not put meeting outcomes into documents, and does not update the task in the tracker. A person still serves as the link between email, chats, and the tracker. Vladimir Sergeev of K2Tech, where around 300 people work with corporate agents every day, described what this gap looks like from the inside. The piece was written by a systems integrator and occasionally promotes its own stack, but that does not diminish the engineering lessons.

Summaries do not solve overload

The idea of having “everything in one window” failed to solve the central problem: task context still lives in employees’ heads. The window cannot tell you which decision has already been approved or where the current deadlines are recorded. In a survey by K2Tech and VK Tech of more than 800 representatives of large Russian companies, 56% expect a fully fledged all-in-one ecosystem to emerge within the next three to five years. But a single window is not enough.

AI also speeds up text production, leaving colleagues to make sense of the output. Researchers at BetterUp and the Stanford Social Media Lab call this workslop: text that looks convincing but leaves the recipient to work out what the author meant. Even a good summary does not eliminate the workload: the text is shorter, but there are more decisions per paragraph, and each needs checking.

The numbers support that skepticism. In an experiment involving Microsoft 365 Copilot and 7137 employees at 66 companies, people spent roughly two fewer hours a week on email, but researchers found no significant changes in the number or mix of tasks completed. An observational study by Berkeley Haas at one technology company revealed another pattern: new processes fill the time freed up, leaving people switching between assistants’ outputs all day.

An agent needs context, not just access

Imagine a meeting where people spend half an hour discussing options, then conclude: “All right, let’s keep doing it the old way for now.” Someone familiar with the project’s history will understand that the decision has not changed. The agent has to distinguish an agreed decision from a proposal left unresolved. That is why Sergeev now says aloud during meetings: “This is a note for the agent.” People are adapting to the bot, rather than the other way around.

The intended workflow looks like this: a videoconferencing agent identifies agreements, a tracker agent creates tasks with assignees and deadlines, and a CRM agent finds the customer record. A person confirms the changes, and only then do they reach the systems. But the workflow has to be assembled and tested separately for each process. The agent must find the right project, verify the assignee, and avoid creating a duplicate if the recording is processed again. Details like these determine whether automation works.

MCP has eased some of the pain: a single connection protocol has replaced a patchwork of custom connectors. But it does not explain what the data means. A won deal in the CRM does not mean a completed project, and a “closed” status may require a document check. Relationships between records and the sources to consult for different questions must be explained explicitly to the agent.

A second technique from K2Tech’s experience is to write key skills as plugins with a fixed sequence. First read the email, then query the CRM, cross-check against the project record, and only then reply. Simply asking the model to prepare a response can lead it to skip an important source.

Limits on autonomy

How much to delegate depends on the cost of an error. An email draft is easy to fix; recovering money after an incorrect transaction is harder. Agents therefore have three modes: read-only, draft for approval, and database writes after operator confirmation. What matters is not just model reliability, but also how visible and reversible an error is.

Access is handled through impersonation: the bot acts on an employee’s behalf, using their tokens and roles through SSO and the directory. But permission to change data does not mean every change should happen automatically, so the available actions are restricted to the task. AIDA tech’s SamurAI platform, on which the setup is built, provides more than 100 built-in tools, 580 ready-made skills, an activity log, and support for more than 10 LLM providers, including local ones. Employees interact with it through the corporate messenger eXpress.

The market’s limitations are also acknowledged plainly: some Russian vendors have no MCP connectors of their own, and sometimes not even a documented API. A custom connector can break after a product update. Sometimes AI is a closed feature within a product: you can get a summary, but cannot connect your own model or knowledge base.

What to do in practice

When choosing a communications platform, examine its API, support for connecting your own model, RAG over corporate data, on-premises deployment, preservation of access permissions, and auditing. The author names VK WorkSpace, CommuniGate Pro, and eXpress as candidates, while stressing that the required features must be checked for the specific version.

The main rule is to first establish who is responsible for the decision, which actions require approval, and what criteria define a completed task. If a process already contains unnecessary approvals, the agent will simply reproduce them. During the pilot, measure how much manual searching and data transfer has disappeared, whether repeat approvals have decreased, and how much time is spent checking the result. If a person receives five notifications from agents instead of one email, the deployment has yet to pay off.