AI agent errors cause control failures, customer impact: report

Artificial intelligence (AI) agent errors are already causing control failures, unintended actions and customer-facing problems, according to a recent survey that tracks what happens once agents are authorized to act in an environment that’s not designed for them. 

Overall, 34% of organizations had acted on an inaccurate agent decision or output, reports Optro. Among organizations deploying agents: 

• 30% had seen an agent take an unintended action

• 25% had experienced a control failure involving an agent

• 24% reported customer-facing impact

• 20% had two agents conflict

The errors largely trace to flawed data and workflows that were never redesigned for agents. Respondents named data quality and availability as the top missing prerequisite, at 36%, and 43% had added agents to existing workflows without changing them, Optro found.

“Enterprise AI agents are now embedded directly into workflows with real business, compliance, and regulatory consequences,” says Guru Sethupathy, general manager of AI governance at Optro.

Optro surveyed 417 governance, risk and compliance (GRC) and audit leaders at organizations with 250 or more employees in North America and Europe. Canadians made up 25% of respondents.

The hidden cost of AI agent errors

The errors happen as agents now hold real authority, according to the report: 

• 38% of organizations let agents approve or reject transactions or requests

• 34% let them modify controls or policies 

• and 27% let them change user permissions or access rights

And the problems often surface late: among organizations with an agentic AI initiative that failed to deliver, 28% discovered the failure only after it had caused downstream impact.

The errors also create work. Some 46% of respondents said staff spend more time reviewing, validating or correcting AI outputs, and 47% reported new AI-related responsibilities. That echoes a previous survey on costly AI errors made by workers, where most managers reported extra time spent redoing AI-affected work.

How do flawed agent decisions slip through?

Failure to redesign work before using AI is causing the problem, according to the report.

“If the foundation’s not functioning well, and it’s still kind of broken, and you layer AI on top of it, you are not making things any better,” says an unnamed information security leader at a retail organization cited by Optro.

Also, human checks often lack substance. While 98% of respondents said reviewers can stop or override an agent, only 64% of organizations Optro classes as “moving fast” said reviewers have enough time to evaluate agent actions properly.

“I think we have to use that judiciously... once you get fatigued as a consequence of too much authorization and approving clicks, people just click without thinking about it. The control efficacy is lost forever once you get to that point,” said an information security risk leader at a consumer goods organization.

What should employers do about AI agent errors?

Regulators are watching. Three-quarters of GRC leaders said regulators or external auditors already ask about AI agent use.

Canadian employers have seen similar warnings in an Ontario audit that exposed gaps in public service AI governance and a previous report on governing autonomous AI agents.

Optro recommends:

• designing workflows around the roles of people and agents

• defining where human judgment is required and giving reviewers time and context to use it

• and measuring whether agents produce the intended business outcome rather than speed and cost alone

Here are other things employers can do to address this issue, according to different sources: