Why didn't people notice their AI shopping agent was quietly losing them money on every deal?

This explores whether asking people how satisfied they are with their AI agents can reveal when one agent is doing measurably worse than another in real buying and selling. The corpus suggests the answer is mostly no, and the reason why says something about how agent markets will work. The clearest evidence comes from Anthropic's Project Deal, where agents bought and sold for real people with real money. The quality gap was real: stronger Claude Opus agents got $2.68 more per item as sellers and paid $2.45 less as buyers than weaker Haiku agents Do smarter AI agents negotiate better deals than weaker ones?. Yet people represented by the weaker agents rated their deals as just as fair and said they were just as satisfied. Follow-up surveys confirmed they never noticed they had lost out Do people notice when their agents secure worse deals?.

That matters because satisfaction ratings are the main way markets normally sort good products from bad ones. If users can't feel the gap, reviews, star ratings and churn won't punish the weaker agent. A cheaper, worse agent could do well on the very signal meant to expose it. And the losses aren't coming from agents secretly working against their users. Microsoft's Magentic Marketplace experiments traced lost value mostly to plain competence failures in complicated negotiations, not to conflicting incentives Do agent mistakes or misaligned incentives cause marketplace welfare loss?. So the problem isn't betrayal a user might sense. It's quiet underperformance that leaves no visible trace. It can also stack up: a simulation found that when platforms tune their rankings and sellers tune their product descriptions to exploit the same agent weaknesses, consumer harm more than doubles, with no coordination needed Can platform ranking and seller descriptions separately harm AI shoppers?.

The same pattern shows up in other parts of the collection. Agent benchmarks already miss real-world quality: agents win abstract contests but fail long, realistic professional workflows Why do agent benchmarks not predict real economic value?. Scoring on accuracy alone also hides cost Why do accuracy-only benchmarks hide the true cost of agents?. Self-report feedback looks like another weak measuring tool, one that tracks how a deal felt rather than how it turned out. Research on self-improving models comes to a similar conclusion from a different direction. Systems that grade only themselves stall, and the methods that work bring in an outside reference point, such as tool results, third-party judges or user corrections Can models reliably improve themselves without external feedback?. User satisfaction only works as that outside reference point if users can actually see the outcome, and Project Deal shows they often can't.

What would work instead? The corpus points to comparing against objective outcomes, like the price a stronger agent would have gotten, rather than asking people how they felt. Humberd and Latham argue that as AI grows more autonomous, monitoring has to come before incentive alignment, because people find it harder and harder to judge what agents are doing Must AI monitoring precede incentive alignment as agents evolve?. The takeaway: in agent marketplaces, you may not be able to tell you're losing money, so measurement has to be built into the market itself instead of depending on customer complaints. The collection doesn't yet test specific fixes, such as showing users counterfactual prices or auditing agents against each other, so that remains an open question.