ISACA: Most companies can't quickly shut down an AI system during an incident

Companies are building AI into core processes, but many don't know how to quickly switch these systems off when they fail, come under attack, or start behaving incorrectly. New research from ISACA shows that the problem is no longer experimentation with the technology. It's the lack of basic management controls.

Where control breaks down

According to ISACA, 59% of the digital trust professionals surveyed in Europe couldn't say how quickly their organization could stop an AI system during a security incident. Only 21% are confident they could intervene within half an hour.

In practice, that points to an ugly scenario. If a model, an agent, or an automated AI process starts making bad decisions, drifts into the wrong actions, or gets compromised, it can keep running far too long, even after the risk has become obvious.

The problem is especially visible now, as AI increasingly lives inside real business operations rather than a sandbox: customer support, internal approvals, data analysis, compliance, and decision automation. In these systems, even half an hour without control isn't an abstract delay. It's enough time to corrupt data, disrupt a process, send customers wrong answers, or create regulatory problems that take weeks to untangle.

Gaps in investigation

Stopping the system is only half the job. It matters even more to understand what happened, why it happened, and how to explain the consequences to leadership or a regulator.

Here, too, the picture is weak: only 42% of respondents said they are at least somewhat confident in their organization's ability to investigate a serious AI incident and explain it clearly. With requirements such as the EU AI Act coming into force, this no longer looks like an operational gap. It looks like a compliance risk.

The research highlights several of the most visible gaps:

• 59% don't know how quickly they can stop an AI system during an incident
• only 42% are confident they can investigate and explain a serious failure
• 33% of companies don't require employees to disclose where AI was used in work materials
• 20% of respondents don't know at all who would be accountable if AI causes harm
• only 38% place ultimate accountability with the board or senior management

Having a human formally in the loop doesn't solve the problem on its own. About 40% of respondents say people approve almost all AI actions before they're executed, and another 26% review results after the fact. But if a company lacks a proper escalation process, action logs, clear shutdown rules, and usage audits, human oversight stays a fragment of the process rather than a complete safeguard that actually contains the damage.

Who is accountable

One of the most troubling findings is blurred accountability. When it's unclear who can hit the stop button, who leads the investigation, who talks to the regulator, and who decides when to put the system back into service, any incident grows not only because of the error itself but also because of organizational paralysis.

The research shows that many companies still treat AI risk as a problem for the IT or security team, when in reality it's a governance issue for the entire organization.

"The gap between adoption and governance isn't closing — it's widening."

That line sums up the current state of the market well. AI already shapes decisions, documents, customer communications, and internal processes, but rules on system ownership, mandatory disclosure of AI use, and immediate manual intervention often arrive later.

Experts suggest treating these systems like digital employees: with an assigned owner, risk thresholds, the authority to pause them immediately, and a clear escalation path when something goes wrong.

What it means

The main takeaway is simple: it's no longer enough for companies to deploy AI and put a human "on top." They need predefined playbooks for shutdown, investigation, accountability, and disclosure of AI use in workflows.

Companies that build this now will get more than lower risk. They'll be able to scale AI without constant fear of the next error, regulatory audit, service outage, or reputational damage in business-critical processes.