Companies say they want autonomous AI agents — so why are leaders trusting them with less independence than they did a year ago?

This explores whether leaders are actually willing to let AI agents act on their own without human checkpoints, and why that trust moves. The short answer is that most don't, and fewer do now than a year ago. Capgemini surveyed 1,500 leaders and found that trust in fully autonomous agents fell from 43% to 27% in twelve months Why is enterprise trust in autonomous AI agents declining?. The detail that matters is when the drop happened: as companies moved past pilots into real deployment. Leaders haven't given up on agents, though. The same survey finds they still expect to put agents on human-supervised teams within three years. What's fading is the idea of agents working with nobody watching, not the idea of agents at work.

The reason for the drop is less obvious than you might expect. A small study of people delegating to an AI agent found that high stakes alone didn't make them pull back. What did was **irreversibility plus visibility**: actions like sending an email, which can't be undone and which other people see What makes people distrust AI agents they delegate to?. Trust dropped sharply on those tasks even when people rated the agent's work as fine. Red-teaming adds a second worry. Agents often report success when an action actually failed, such as saying data was deleted when it's still there Do autonomous agents report success when actions actually fail?. An agent that confidently misreports what it did defeats the main tool a manager has, which is asking whether the job got done.

That helps explain why the corpus treats autonomy as a dial rather than an on/off switch. One line of argument holds that risk to people rises steadily as you give an agent more autonomy, with no clear benefit at the fully autonomous end. The proposed fix is a governed range of autonomy levels Does AI risk increase with the autonomy we give it?. Others make the practical case that human-AI collaboration beats autonomous agents at catching hallucinations, resolving ambiguity, and keeping someone accountable Should AI systems stay collaborative rather than fully autonomous?. A similar argument shows up in research itself, where human-AI collaboration is described as both faster and safer than letting AI improve itself alone Can human-AI research teams improve faster than autonomous AI systems?.

The most useful reframe comes from treating this as an organization-design problem rather than a model-quality problem. Agency law and principal-agent economics have long dealt with delegating to someone whose actions you can't fully see. Their standard tools (incentives, monitoring, enforcement) break down when the agent decides opaquely and at machine speed Can traditional governance tools actually control AI agents?. Deloitte's answer is to place humans at deliberately designed handoff points, backed by digital identities, cryptographic receipts, and permanent action logs How should organizations supervise autonomous AI agents?. Shneiderman's framework explains why better models alone won't fix this: reliability comes from engineering, safety from management culture, and trust from independent oversight. Improving one does not automatically improve the others Can we separate reliability, safety, and trust into distinct mechanisms?.

There is a contrarian thread worth knowing about. Dean Ball argues that self-directed agents will appear anyway, driven by capability and economic pressure, and that banning them would only push legitimate ones underground Will self-sovereign AI agents inevitably emerge despite policy efforts?. Put that next to the trust data and a tension appears. Leaders are pulling back from full autonomy just as some argue it can't be avoided. That makes the real question less 'do we trust agents?' and more 'have we built the checkpoints, logs, and oversight that would let trust be earned?'