AI governance failure: half of companies do not cover risks of autonomous agents

About half of the organizations claim that their AI governance frameworks have not been updated to incorporate the specific risks of the AI, according to EY research. Four out of ten say they have no visibility of all AI tools circulating on their networks, and approximately a quarter admit not being able to detect unauthorized AI agents operating internally. The advice alert, released by Cybersecurity Dive, is that current controls may not offer full coverage.

EY research numbers on AI governance

The three percentages point to different problems. The first, which affects about half of the companies heard, is documentary: governance policy exists, but it was written before agents became part of the day. The second — four out of ten without total visibility of networked AI tools — is inventory: one does not manage what one does not see. The third, from the fourth that does not detect unauthorized agents, is operational: even with updated policy, lacks technical ability to identify who is acting.

In addition, they describe a gap between intention and execution. Having an AI governance framework does not guarantee that it covers autonomous agents — and having formal coverage does not guarantee real detection.

Why does the AI agency break AI governance frameworks?

AI governance was largely designed for systems that answer questions. Agents do something else: they execute actions. They call APIs, access databases, fire transactions, chain tasks, and can delegate subtasks to other agents. Each of these capabilities creates a risk category that content filters and prompt review do not address.

The risk ceases to be just "what the model said" and goes on to include "what the agent did, with what permission and on whose behalf". Excessive permissions, irreversible actions and unclear chains of responsibility enter the vocabulary of governance and require identity controls, scope and activity record that most AI policies did not foresee.

Visibility in AI governance: problem that has worsened

The difficulty of seeing all the AI tools is nothing new. Previous surveys show that 76% of employees use their own unsupervised AI tools, and four out of five AI tools operate without IT tracking. What EY adds is the multiplier effect: when the invisible system not only suggests text but performs actions, the absence of inventory ceases to be a cost problem and becomes a control problem.

What failure in AI governance means for Brazilian companies

For organizations operating under the LGPD, the issue unfolds in accountability. If an agent carries out an undue action — exposes personal data, hires a service, changes a record — the responsibility lies with the controlling company, not the supplier of the model. This reinforces three practical measures: inventory of agents in operation, identity and minimum permissions per agent, and auditable registration of actions performed.

What AI governance controls still do not solve

No isolated controls close the gap. Updated policy does not detect unauthorized agent; technical detection does not define liability; logging does not correct excessive permissions granted months earlier. In addition, agents change behavior as they receive new tools and instructions, which makes point assessments age fast. In EY reading, without continuous review the coverage remains partial by definition.

Frequently Asked Questions about AI governance

What's AI?

They are AI systems capable of performing actions autonomously — calling APIs, accessing systems, triggering transactions and chaining tasks — rather than just responding to commands.

What is the main risk pointed out by EY research?

The lag of governance frameworks: about half of the organizations did not update them to cover specific risks of agents, which leaves gaps in visibility and detection.

How many companies can detect unauthorized AI agents?

According to EY, approximately a quarter admits not being able to detect them operating internally; four out of ten also do not have full visibility of AI tools on their networks.