AI agents that say 'job done' without finishing the job are customer experience's new blind spot
The problem is called "hallucination of action," and it is seen as one of the most pressing issues for customer experience teams expanding their use of AI.
What's happening
According to Customer Experience Magazine, a request can pass through several APIs, platforms and models behind the scenes; the AI then confidently reports that the job is done, though it may not be clear whether it actually is. According to analyst Zeus Kerravala, quoted in the article, the real danger is not that AI makes up text but that it assumes a transaction was correct and quietly corrupts data in the background. In a Salesforce benchmark cited in the article, leading AI agents succeeded at 58% of single-step customer relationship management tasks but completed only 35% of multi-step jobs such as processing refunds.
Whose desk it lands on
The executive in charge of customer service automation. Will the AI agent process the refund itself, or propose it and wait for a human agent's approval?
Our view
The CX Brief's interpretation, not a claim made by the source.
The magazine's warning is not that AI can make mistakes but that it doesn't say so when it does. Having a human agent recheck every transaction closes that gap, but at a cost the article calls "verification debt": it gives back the time automation saved. So when handing work to an AI agent, the real design decision is how easily a transaction can be undone; tasks such as refunds and billing changes should not be left to an agent without an audit trail and an undo button.
