When AI Agents Act on the Business’ Behalf, Who Is Accountable?

Vladimir Tikhomirov argues that enterprise infrastructure built around a human decision-maker isn’t equipped for AI agents that can interpret, decide and act independently, and that businesses need to rebuild around machine identity, permissions and traceability before they can safely remove human approval steps.

Enterprise software has, for decades, been assembled around the idea that a person makes the decisions, while the system carries them out. With AI agents now able to interpret tasks, call other systems and execute actions on behalf of people or even entire companies, I think the infrastructure we have today is simply not built to keep that level of autonomy stable and under control.

Enterprise systems still expect a human at the controls

All of this would be less of a problem if AI systems and agents were arriving into enterprise infrastructure that had already been rebuilt for their full-scale use. But this is not really where we are today, and I doubt we could have been there. Most enterprise systems are still very much honed with humans in mind.

Google, for example, recently ran a survey of 1,402 global IT leaders and found that 83% of organisations needed upgrades to infrastructure to support production-grade agentic AI. The gap is substantial.

Most enterprise mechanisms and controls still assume there is a human somewhere in the process who can be identified and authenticated. They are also the ones bearing liability, which is why permissions are attached to employees, while approval flows are designed around somebody clicking an “approve” (or “decline”) button.

When AI agents and autonomous systems step onto the scene, everything changes. An agent can, of course, receive one high-level instruction or retrieve information from several sources at once in the blink of an eye. It can then initiate another action almost immediately. But giving it ordinary user credentials and hoping existing controls will do the rest is, in my eyes, a fairly clunky solution.

And the accountability problem, being more profound than the infrastructure one, is what takes the stage. If the system can interpret, decide and act by itself, then we need to understand where the decision actually came from — and who, eventually, is responsible for it.

Who is accountable when agents act together?

Once several autonomous systems start working together, responsibility for their actions gets much less straightforward.

Say, one agent reads internal data, another recommends an action, a third executes it, while a fourth is left choosing further steps based on the result. And if a system failure is detected, pointing to the last agent in the chain that made the final decision tells us very little.

That is why what matters here is preserving decision context. Major international organisations are already working in this direction, as the OECD proposed to maintain documentation throughout the AI lifecycle, including model inputs, decision-making criteria, rationale, risks and mitigation measures.

With agents, though, I think this has to go one step further.ompanies need to be able to reconstruct the actual chain of action: what information an agent had when decisions were made, what degree of authority it was given, which system it communicated with.

Without this in place, accountability risks rapidly turning into archaeology.

And companies have yet to keep up with this. For example, Deloitte’s 2026 State of AI research found that only 21% of surveyed organisations had a mature governance model for agentic AI. So, there is a large divergence between how much autonomy businesses want to give these systems and how well they can trace and govern what will happen afterwards.

From here, I don’t think the solution is to put a human approval step back in front of every action. That would defeat much of the purpose. But in order to remove those approvals safely, companies had better impose clearer limits on what an agent is allowed to do, finer traceability and a way to rewire its decisions.

The faster agents work, the faster failures spread

There is one more thing that changes once humans are taken out of individual steps: the speed at which mistakes disseminate.

Usually, human involvement is treated as a bottleneck. And, frankly, sometimes it is. But the delay which humans bring also works as a safeguard. A suspicious payment may wait for approval, an unusual request gets checked twice, or somebody simply notices that the result does not seem to look right.

Yet, in the case of autonomous agents, this phase disappears, which makes a wrong instruction or excessive permissions move between systems in seconds, further spreading mistakes across the chain.

And what is the way out? To set up controls around systems in a way that they can work at roughly the same speed.

There are some approaches that big tech is already implementing. Cloudflare recently introduced identity and wallet tools for AI agents, with spending caps, approved merchant lists and maximum transaction sizes. What I find interesting here isn’t the wallet itself, but its underlying idea, where if a machine is allowed to act on behalf of an organisation, it should probably have its own identity and permissions.

So, this, I think, is what enterprise AI should develop towards. Bringing a human back in front of every action would throw us almost back to where we started. The answer, then, is to rebuild the infrastructure around the fact that agents do not behave like ordinary enterprise software.

Identity, permissions, traceability and machine-speed controls will have to become part of that infrastructure from the beginning. Otherwise, companies may give agents more autonomy without actually having the systems required to keep this autonomy under control.