Directum: Why Businesses Talk Up AI Agents but Hold Off on Deploying Them

Chatbots that answer questions and draft emails are no longer enough for businesses. Companies want AI systems that can take a task all the way to a result inside corporate processes. But that is exactly where autonomy looks most valuable, and where the main limits show up: expensive infrastructure, errors when working with many systems, and no clear accountability for an agent's decisions.

Directum suggests separating assistants from agents. An assistant is essentially an advanced conversation partner: it responds to a request and helps with text, search or a suggestion, but it does not run a process itself. An agent works differently: it breaks a goal down into steps, picks tools, switches between systems and hands the result back to a person. In a corporate setting, that can mean a full cycle of work on a document, request or task, from classification to handoff to the responsible employee.

The company says the surge of interest in agents is no accident. After the wave of pilots in 2024, businesses stopped asking whether AI can write an email or summarize a meeting and moved on to the next level: can a model be trusted with part of a process that requires collecting data from several systems and making a working decision? At the same time, LLMs themselves have become stronger at planning and multi-step task execution, and standards such as function calling and MCP have made it easier to connect external tools, from ERP and CRM to ECM, calendars and RPA scenarios.

The first obstacle is infrastructure. Large and mid-sized businesses want to run such solutions inside their own protected perimeter, because they involve finances, personal data and trade secrets. But deploying mid-range models locally requires serious GPU investment, and supply, price and availability of such hardware in Russia remain problematic. Renting capacity from data centers is technically possible, but it often raises concerns among information security teams. As a result, many companies reach the stage of interest but not of production launch.

The second limit is the quality of an agent's work as the number of integrations grows. In theory, it should be able to work with ERP, electronic document management, CRM, email, knowledge bases and calendars at the same time. In practice, every new connection adds risk: the model may pick the wrong tool, mix up call parameters or even invent a nonexistent API. According to Directum's observations, the likelihood of such failures rises noticeably after roughly 15 tools, so today it is safer to restrict an agent to a narrow domain or a set of 5–7 proven systems. Scaling across an entire enterprise IT landscape remains a difficult engineering problem.

The third barrier is accountability. Even if an agent can carry out a chain of actions on its own, the final decision in sensitive scenarios is still left to a human. No executive is ready to hand a model, without reservations, the right to sign a major contract, approve a risky transaction or transfer money without oversight. And it is not only distrust of the technology: the legal framework has not caught up with the level of automation. If an agent makes a mistake, it is unclear who answers for the consequences: the vendor, the integrator, the process owner or the employee who configured the scenario. Until that question is settled, autonomy will remain limited.

The conclusion is fairly sober: the market is moving not toward agents fully replacing employees, but toward an intermediate model in which AI takes over well-defined parts of a process under human supervision. That is why workflow agents look best right now: solutions with clear boundaries, a limited set of actions and predictable ROI. For businesses, this is not a rejection of the agent approach but a way to adopt it without excess risk: start with narrow cases, test reliability and only then widen the automation perimeter.