Photo By:  Immo Wegmann

For several years, using AI at work was fairly easy to understand. An employee opened a chatbot, typed a question and received an answer. IT teams could decide which tools employees were allowed to use and who had access to them.

But today, agentic AI is changing that.

Tools such as Claude Code and Codex can do much more than answer questions or suggest code. They can work with files, use software tools, run code and complete multi-step tasks with varying levels of human supervision among other things. The technology is moving AI from something that simply responds to people toward something that can carry out work on their behalf.

As these tools spread through companies, a new problem is emerging: companies may have a harder time understanding exactly how AI is being used.

The issue is no longer just how many employees use AI. It is what those employees are asking AI to do. The change is especially clear in and among software development. Imagine a developer asking an AI agent to fix a bug. The agent could inspect the relevant code, make changes, run tests and respond to problems it encounters along the way. One request from an employee can lead to a series of actions by the AI.

The employee made one request. The AI performed many steps. That difference creates a new challenge for company security teams.

Traditional security systems are largely built around questions such as: Who is this person? What systems can they access? What permissions do they have? With AI agents, companies also need to ask: What can the AI actually do with those permissions?

That question becomes more important as companies connect agents to internal data, software and business systems. AI agents are increasingly being used for tasks beyond software development, including research, reporting and other business processes.

Harsha Kumar, CEO of NewRocket, has focused on similar issues as companies bring AI into enterprise workflows. His work has included discussions around AI governance, data and the systems needed to make AI useful inside large organizations.

The shift also makes AI adoption harder to measure. A company might say that a certain percentage of its employees use AI. But that number does not necessarily show how much work those employees are delegating to agents.

One employee might use an AI assistant for a few questions each day. Another might use several agents to work through complex tasks over several hours.

The amount of work being done by AI can therefore vary significantly from one employee to another.

The cost can vary too.

Agentic systems can use more computing resources when they perform longer tasks or use multiple tools. That makes AI spending different from traditional software spending, where companies often buy a fixed number of licenses for a predictable price.

Companies are already paying closer attention to this problem. JPMorgan, for example, has introduced a sandboxed environment for some engineers using AI and has placed spending limits on certain users. The goal is to give employees access to powerful AI while creating additional controls around how the technology is used.

There is another challenge which is, oversight.

One approach is to require employees to approve important actions. But as agents perform more steps, asking a human to approve every action may become difficult to scale.

The alternative is to put stronger controls around the agent itself.

That can mean limiting which systems it can access, giving it only the permissions required for a particular task, recording its actions and placing it inside a controlled environment.

This points to a broader change in how companies may need to think about AI governance.

Instead of simply asking how many employees have access to an AI tool, companies may increasingly need to understand how many agents are operating, what systems they can reach, what permissions they have and how much work they are completing without direct human involvement.

That is a much harder problem to measure.

It also changes the meaning of AI adoption.

A company can know exactly how many employees have an AI license. It is much harder to know what those employees are building with the technology, what workflows they are creating and how much authority they are giving to AI agents.

That unpredictability could become one of the defining challenges of enterprise AI.

Agentic AI promises to make companies more efficient by taking on work that once required human effort. But as these systems become part of everyday business processes, they could also make the organization itself harder to understand.

The biggest question for businesses may no longer be:

“How much are our employees actually using AI?”

The question instead may be:

“Do we know what our AI is doing?”