We're adopting agents before we know how to shut them down: only 40% are confident they can reverse their failures, says Veeam

By El Ecosistema Startup · November 6, 2026.

Eighty-eight percent of organizations already use or are testing AI agents. Yet only 40% of those responsible for deploying them are confident they can precisely isolate and roll back a failure in that technology. The figures come from the Data and AI Trust Gap report by Veeam, a data protection company, presented at VeeamON London.

Veeam surveyed 600 senior executives (CEOs, CIOs, CISOs and CDOs) in financial services, healthcare, manufacturing, retail and technology across North America, Europe and Asia-Pacific. Fieldwork was carried out between March 16 and April 6, 2026, according to the coverage cited by the outlet. It is worth remembering that Veeam sells solutions in this space, so the study also serves its commercial interests.

Other findings from the report, attributed to Veeam: only 7% consider themselves truly AI-ready, and 95% say problems with their data have already held back their progress. Only 28% believe they can detect AI systems operating outside approved parameters.

Visibility is worse when respondents are asked about specifics. Among those already using AI, 29% know which systems an agent accessed, 25% what actions it executed, 24% what decisions it influenced and 22% what data it used.

The most striking finding is organizational. According to Veeam, where the CISO (chief information security officer) owns agent risk, organizations are 24% more likely to detect anomalous behavior. Where responsibility is split across several teams, they are 47% less likely. These are relative differences between groups of respondents, not a controlled experiment, and they do not prove causation.

There is also a rift at the top: 65% of CEOs believe their company has a complete AI inventory, compared with 48% of technical leaders.

In response, Veeam is introducing Agent Commander, built around three capabilities: detecting which agents exist and when they drift; protecting with controls and access limits; and undoing, that is, recovering clean data after an incident. This is the company's own proposal, and the article provides no independent evidence of its effectiveness.

For context, the source notes that Gartner projects more than 40% of agentic AI projects will be canceled before the end of 2027 due to costs, unclear value or inadequate risk controls. In addition, 61% of respondents say the EU AI Act (the European AI regulation) has influenced their investments, and 47% say maintaining audit trails of AI decisions is their biggest compliance concern.

**Our take.** The thesis is simple: the bottleneck for agentic AI is no longer the model's capability but the organization's ability to answer for it. An agent is software with permissions that acts fast. If nobody knows what it touched, the problem is not one of intelligence but of operations, and it looks more like backup and access management than AI philosophy.

This finding fits a thesis we have been advancing: the real perimeter of agentic AI is non-human identities and credentials. An agent running on shared accounts is impossible to trace or contain cleanly. Giving each agent its own identity, scoped permissions and a named owner is probably the most cost-effective step a company can take today.

The organizational findings warrant caution, but they point in a reasonable direction. Detecting drift requires a definition of normal, a single place where signals land and someone with the authority to pause an agent without waiting for a committee. A diffuse division of responsibility fails on exactly those three points.

Winners and losers: the winners are companies that treat reversibility as a design requirement before scaling, along with data governance and recovery vendors, which have a natural market here. The losers are those deploying a second and third agent without having sorted out the first, and probably a good share of the 40% of projects Gartner expects to be canceled.

In the short term, the transition will be messy: agent incidents, uncomfortable audits and regulation driven by failures. But that friction is healthy. In the long term, the promise of agents that free people from routine work will only be credible if we know how to stop them and undo their mistakes. Trust is not declared; it is built with infrastructure, and the abundance we anticipate depends on it.

A practical test for any team: can you answer, within five minutes, which systems an agent touched, what actions it executed, what data it used and who can shut it down? If not, that is your next priority.