Why we stopped saying “AI agent governance”

We built Kimetai to govern AI agents. Then we spent a few months talking to the people who would have to buy it, and the same thing happened in almost every conversation: the automation that worried them had no AI in it at all.

Two species, one blind spot

A webhook that replicates invoicing data into accounting, and an agent that reads a customer email and decides its priority, look nothing alike on a whiteboard. Operationally they are the same object: software that acts on your systems without a human clicking, holding a credential, moving data that matters. Both need an identity, an owner, a policy and a record. Neither has any of the four in most companies.

The legal perimeter never split them

This is the part that settled the decision. Article 30 of the GDPR does not ask whether a model was involved. Neither does asset management under NIS2, nor the ICT asset inventory under DORA, nor the invoicing-record regimes rolling out across Europe. Exactly one of the five frameworks in play is AI-specific — and it is the one whose heaviest obligations arrive last.

Naming the category “automation governance” was not marketing. It was the legal perimeter catching up with what we were already seeing.

What changed in the product, and what did not

Discovery now inventories both species from the first connection, the console shows flow:// and agent:// in the same list, and classification returns exposure framework by framework instead of an AI Act label alone. What did not change: the decision point still sits in front of the action, and the audit trail is still sealed as it happens.

We have not dropped the term “AI agent” either — it is precise when the thing genuinely is one, and half our roadmap is about model-driven steps. The rule we follow now is simply the honest one: talk about automation when the obligation applies to both, and about agents when the distinction actually matters.