ISACA: most companies are not ready to quickly shut down an AI system during an incident

Companies are actively embedding AI in key processes, but in many cases they don't understand how to quickly shut down such a system in case of failure, attack, or erroneous behavior. A new ISACA study shows that the problem is no longer about experimenting with technology, but about the absence of basic management control.

Where Control Breaks

According to ISACA, 59% of surveyed digital trust specialists in Europe could not say how quickly their organization is able to stop an AI system during a security incident. Only 21% are confident they can intervene within half an hour. In practice, this means an unpleasant scenario: if a model, agent, or automated AI process starts making erroneous decisions, drifting into wrong actions, or is compromised, it can continue working for too long — even after the risk becomes obvious.

The problem is especially noticeable now, when AI is increasingly not in a sandbox, but inside real business operations: customer support, internal approvals, data analysis, compliance, and decision automation. In such systems, even half an hour without control is not an abstract delay, but time during which you can corrupt data, disrupt a process, send incorrect answers to customers, or create regulatory problems that will take weeks to sort out later.

Gaps in Investigation

Stopping the system is only half the task. More importantly, it's necessary to understand what exactly happened, why it occurred, and how to explain the consequences to management or a regulator. But the picture here is also 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. Against the backdrop of requirements like the EU AI Act coming into force, this already looks not like an operational shortfall, but as a risk for compliance. The study specifically highlights several most notable gaps:

• 59% don't know how to quickly 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 even know who will be responsible if AI causes harm
• only 38% see ultimate responsibility lying with the board of directors or top management

Meanwhile, formal human involvement doesn't solve the problem by itself. About 40% of those surveyed say that people approve almost all AI actions before execution, and another 26% check results after the fact. But if a company doesn't have a proper escalation scheme, action logs, clear shutdown rules, and usage audit, human control remains a fragment of the process, not a full-fledged defense system that actually mitigates damage.

Who is Responsible

One of the most unpleasant conclusions is blurred responsibility. When it's unclear in a company who exactly can press the stop button, who conducts the investigation, who communicates with the regulator, and who decides whether to return the system to operation, any incident begins to grow not only because of the error itself, but also because of organizational delays. The study shows that many companies still perceive AI risk as a problem for the IT or security team, although in reality it's a matter of management at the level of the entire organization.

"The gap between implementation and management is not shrinking — it's growing."

This thesis well describes the current state of the market. AI already influences decisions, documents, customer communications, and internal processes, but the rules for owning the system, mandatory disclosure of its use, and immediate manual intervention often emerge later. Experts recommend treating such systems as digital employees: with a designated owner, risk thresholds, the right to immediate pause, and a clear escalation route if something goes wrong.

What This Means

The main conclusion is simple: companies can no longer just deploy AI and put a person "on top." You need pre-planned scenarios for stopping, investigating, assigning responsibility, and disclosing AI use in work processes. Those who build this now will not only have fewer risks, but also the ability to scale AI without constant fear of the next error, regulatory check, service outage, and reputational losses in business-critical processes.