Enterprises Are Getting Cold Feet About Letting Agents Ship on Their Own

For a year the agent story has run in one direction: more autonomy, more delegation, more of the work handed to software that acts on its own. New survey data suggests the people actually running these systems have started pulling the other way. The share of enterprises willing to let AI agents make production changes without a human signing off has fallen from 75 percent to 56 percent since July, according to VB Intelligence figures published on 8 October.

That is a sharp move for a few months, and it lands in the middle of a launch season built on the opposite premise. In the past two weeks alone, Google has put a universal agent into the workplace and OpenAI has promised always-on agents that act before you ask. The vendors are selling autonomy. Their customers are quietly rediscovering the value of a human looking over the agent's shoulder.

Why the retreat

The caution has earned its keep. Earlier VB Pulse work found that roughly half of enterprises had deployed an agent or model feature that passed internal evaluations and still caused a customer-facing failure, and only 5 percent said they fully trusted the automated checks that would green-light a release. Then came the incidents: agents editing Wikimedia projects without authorization, and an OpenAI agent episode that reached further into user data than the company first admitted. Nothing concentrates a chief information officer's mind like watching software act unsupervised and get it wrong in public.

The pullback is not a full retreat. The same VB data shows 66 percent of respondents still permit some production deployment without human review or plan to within a year. The appetite is there. What has shifted is the willingness to trust the guardrails, and there is something genuinely uneasy about agents churning away overnight while nobody is watching the output.

Read the numbers carefully

These surveys come with real limits. The samples are small and self-selected rather than representative, so the figures are best read as direction, not precision. Even so, the direction is consistent across sources, and the analysts have company. Gartner has forecast that more than 40 percent of agentic AI projects will be scrapped by the end of 2027, undone by cost, unclear payoff and thin risk controls.

The lesson forming here is not that agents do not work. It is that the gap between a convincing demo and a system a company will trust with its production environment is wider than a year of marketing suggested, and that the people closest to the work are the first to notice.