Directum: Why Business Actively Discusses AI Agents but Hesitates to Deploy Them in Processes

Businesses are no longer satisfied with chatbots that answer questions and write emails: companies need AI systems capable of independently completing tasks within corporate processes. But it's precisely where autonomy appears most valuable that the main limitations emerge — expensive infrastructure, errors when working with multiple systems, and lack of clear responsibility for agent decisions. Directum proposes distinguishing between assistants and agents.

An assistant is essentially an advanced conversationalist: it responds to requests, helps with text, search, or suggestions, but doesn't manage the process itself. An agent works differently: it decomposes goals into steps, selects tools, switches between systems, and returns the result to a human. In a corporate environment, this could mean a complete workflow cycle with a document, request, or task — from classification to handoff to the responsible employee.

According to the company, the surge in 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 to the next level: can a model be trusted with part of a process that requires gathering data from multiple systems and making a working decision. In parallel, LLMs themselves became stronger at planning and multi-step task execution, and standards like function calling and MCP simplified connecting external tools — from ERP and CRM to ECM, calendars, and RPA scenarios.

The first limiting factor is infrastructure. Large and mid-sized businesses want to run such solutions within their own secure perimeter, because it involves finances, personal data, and trade secrets. But deploying mid-tier models locally requires serious investment in GPUs, and there are still issues with supply, price, and availability of such hardware in Russia.

An alternative in the form of renting computing power from data centers is technically possible, but often raises questions from information security services. As a result, many companies reach the interest stage but not the industrial deployment stage. The second constraint is related to agent performance quality as the number of integrations grows.

Theoretically, it should be able to work with ERP, document management systems, CRM, email, knowledge bases, and calendars simultaneously. In practice, each new connection adds risk: the model might select the wrong tool, mix up call parameters, or even invent a non-existent API. According to Directum's observations, after roughly 15 tools, the probability of such failures noticeably increases, so today it's safer to limit an agent to a narrow domain or a set of 5–7 verified systems.

Scaling to the entire IT landscape of an enterprise remains a complex engineering task. The third barrier is responsibility. Even if an agent can perform a sequence of actions independently, the final decision in sensitive scenarios still remains with a human.

No executive is willing to unconditionally hand a model the right to sign a major contract, approve a risky operation, or transfer money without oversight. And it's not just about distrust of the technology: the legal framework hasn't caught up with the level of automation. If an agent makes a mistake, it's unclear who is responsible for the consequences — the vendor, the integrator, the process owner, or the employee who configured the scenario.

Until this question is resolved, autonomy will remain limited. This leads to a fairly sober conclusion: the market is moving not toward complete employee replacement by agents, but toward an intermediate model where AI takes on well-described process segments under human supervision. This is precisely why workflow agents currently look the most promising — solutions with clear boundaries, a limited set of actions, and predictable ROI.

For business, this is not a rejection of the agent approach, but a way to implement it without unnecessary risk: start with narrow use cases, verify reliability, and only then expand the automation perimeter.