An agent can own the mistake. It cannot own the cost.

In September an AI agent set up translation for this site. It chose the model, wrote the pipeline, and ran it. Two runs spent about twenty dollars and produced nothing I could keep, because of rules the agent had written itself. When I asked for the money back, the agent could explain exactly what went wrong and fix it. It could not pay. Neither could the support agent I asked next.

That is not a failure of politeness. An agent can be responsible for an outcome in the sense of having caused it. It cannot be accountable, because accountability means someone carries the cost, and the agent owns nothing.

So the accountability has to sit somewhere else, and it has to be decided before the action. A spending limit the agent cannot raise. An estimate shown before the run. Paid work saved even when a later step fails. The person who authorizes the spend should see the ceiling, not the receipt.