Hallucination
A confident, well-formed answer that is simply false. Worst on multi-step reasoning and specialised domains, which describes key account work exactly.
A language model does not know when it does not know. Asked a question at the edge of its knowledge it generates a confident, well-formed answer that is simply false, complete with a plausible citation, and the International AI Safety Report finds the problem worst exactly where account work lives: multi-step reasoning and specialised domains. A model asked about your customer’s supply chain will produce a paragraph. Whether the paragraph is true is a separate question it cannot answer for you.
The book’s line is worth memorising. A fluent wrong answer in front of your most important customer is more dangerous than no answer at all. Which is why the method keeps a human in the loop at every point where the output goes to a customer, and why the living account plan keeps its raw sources, so that any claim in the plan can be traced back to the transcript or the report it came from.
Set out in full in: Bengio et al., International AI Safety Report (2026); Creating Customer Value Propositions with AI (Kogan Page, 2027), chapters 2 and 7.
Related terms
Automation bias
The human tendency to trust the machine even when it is wrong. The danger is not that AI replaces your expertise but that it lulls it into disuse.
Human in the loop
Adopt, but keep your hands on the wheel. The machine amplifies judgement. It does not supply it.
The three registers
Productive, creative, collaborative: three different jobs AI can do for an account team. Not a maturity ladder, and the collaborative one needs an expert in the loop.
See the method run on one real account in three days: the One Page Proposition.