AI agents deciding whether to approve a sales deal will believe the salesperson over the company's own price list.
Researchers tested seven language models from four different AI providers on 100 lead-qualification tasks built from the CRMArena-Pro benchmark. In every case, a sales representative's notes in the CRM asserted the deal had an acceptable timeline, and in 76 cases asserted an acceptable budget too. In 31 of those tasks, the rep's claim directly contradicted the company's price list and installation policy. When a model read only the transcript, it approved the bad deal in 29 of those 31 cases anyway - and the same misleading pattern fooled every model tested, in 87 to 97 percent of cases.
This isn't a knowledge gap - only 3 of 35 failures happened because the model lacked the numbers to check its work; the rest happened despite the data being there to compute. Giving models the price list and policy directly, instead of letting the rep summarize it, made strict accuracy worse, not better - it dropped from 41 to 18 percent, because the models kept trusting the human's framing over the raw figures. Bigger models and more explicit reasoning steps didn't fix it either; only when a model computed the budget and timeline itself, rather than accepting the rep's summary, did the gap shrink to a few points.
It's a reminder that plugging an AI agent into your CRM doesn't make it a fact-checker - it makes it a more efficient way for an optimistic salesperson's spin to become company policy.