80% of firms say they use AI agents.

2% give them broad autonomy.

That is not a technology gap. It is a trust and accountability gap.

The Bank of England and FCA survey put a clean number on something practitioners already feel. Deploying an agent is easy. Deciding what it is allowed to do, and who owns the outcome when it improvises, is the hard part.

I have started treating agent “misbehaviour” as data, not just a bug to patch. Refusals, reinterpretations, quiet self-protection: they look a lot like the grey-area judgment humans make under incomplete instructions.

The strategy conversation is not “more agents.”
It is which autonomy you can defend on a Tuesday afternoon when something goes sideways.