AI agents in the enterprise: execution can be delegated, accountability can't, and almost no one is ready

By Revista C-Level · November 11, 2026.

Enterprise AI agents are no longer just assistants that answer questions: they now query databases, use applications and carry out tasks without constant human supervision. That is the thesis of the Revista C-Level article, which draws on the technology firm Noventiq and on data from Colombia.

The most worrying figure comes from a Cloud Security Alliance survey cited by Noventiq: 82% of the organizations surveyed had unknown AI agents operating in their technology environments. In other words, most of those companies lacked a complete inventory of which agents were acting inside their systems.

The contrast lies in maturity. According to the Colombia AI Pulse study, published by Endeavor Colombia in August 2026, 61% of the companies surveyed consider AI a strategic priority. Even so, 93.5% remain in the exploration and development stages, and only 0.2% have reached advanced levels.

According to Noventiq, the main risk is excessive permissions: agents that can modify records, access confidential information or execute financial transactions without adequate controls. The firm proposes five preconditions:

- Each agent gets its own digital identity, with an assigned owner. - Data governance: what information it can access and under what conditions. - Limited permissions, following the principle of least privilege (access only to what is strictly necessary for its function). - Monitoring and traceability, to detect anomalies and reconstruct decisions. - Human oversight of sensitive operations, with mechanisms to stop the agent.

The article adds that these measures must cover the agent's entire lifecycle, including permission changes and decommissioning. Its central quote comes from César Tejada, general manager of Noventiq Colombia: “We can delegate execution, but we cannot delegate corporate responsibility.”

**Our take.** One caveat about the source: Noventiq sells security and technology adoption services, so it has a commercial interest in the risk being perceived as large. The 82% figure comes from a third-party survey that the article cites only indirectly, without detailing its sample. I would treat it as an order-of-magnitude signal, not a precise measurement.

Even with that caution, the diagnosis is sound and fits something we have been seeing: accountability for agents is moving from theory onto balance sheets and into contracts. In general, the problems with agents tend to be about governance rather than intelligence: who created them, what they can touch and who answers if they fail. It is the same old problem of service accounts and legacy access, but now with systems that act on their own and at machine speed.

What is interesting about the Colombian data is that slowness can be an advantage. If 93.5% are still exploring, there is room to get identity, permissions and traceability right before scaling, rather than fixing them after an incident. Companies that automate first and govern later, as the 82% of unknown agents suggests, will pay dearly for it.

In the short term, this means new and unglamorous work: auditing, inventorying and restricting. It is an uncomfortable transition, but also an emerging job: overseeing agents. In the long term, the promise that agents will take on routine work and free people up will only be fulfilled if trust is built on verifiable controls. Without them, productivity will be fragile and the first crisis will stall adoption.

Tejada's quote sums up the principle well: technical autonomy can grow without limit, but accountability remains human. That rule, more than any model, will decide which companies scale successfully.