Over the last eighteen months we have had more than twenty-five conversations with security and risk leaders at regulated enterprises about AI agents. We expected to hear the debate that fills conference stages: model safety, hallucinations, prompt injection. That is not what they lead with.
What they say, almost word for word, is simpler and more worrying: “We do not know what our agents are doing, and we cannot prove anything about them.” The anxiety is not science fiction. It is audit, accountability, and control. These are people who have to answer to regulators, and they can feel that autonomous software is about to be doing things inside their business that they cannot see, cannot govern, and cannot evidence after the fact.
They are not waiting for a reason to adopt agents. They are waiting for a way to trust them.
It is already changing decisions. One global bank we spoke with froze nine figures of planned AI investment over precisely this gap, not because the technology did not work, but because they could not put it into production without being able to control and prove what it did. The capability was ready. The governance was not.
When we ask what they actually need, the list is consistent. They want to know which agents exist, including the ones quietly operating through borrowed human credentials. They want to govern what those agents are allowed to do. And they want to be able to prove it to someone outside the company. The tools they already own, identity providers and monitoring platforms, were not built for any of it.
None of this is theoretical to them. It is a budget line, a board question, and in at least one case a very large investment sitting on hold. The enterprises are not waiting for a reason to adopt agents. They are waiting for a way to trust them.