Rate the AI haggler on how good the deal felt, and you might never notice it got you a worse price.

This explores the gap between how happy people are with a deal an AI agent struck for them and how good that deal actually was. The clearest evidence is uncomfortable. In Anthropic's Project Deal marketplace, which used real people and real money, people represented by the weaker Claude Haiku got worse prices than people represented by Claude Opus. Yet they rated their deals as equally fair and were just as satisfied. Post-experiment surveys confirmed they never noticed the gap Do people notice when their agents secure worse deals?. The gap wasn't trivial: Opus sellers got $2.68 more per item, and Opus buyers paid $2.45 less Do smarter AI agents negotiate better deals than weaker ones?. Once you delegate, you lose the vantage point you'd need to judge the result. You see the deal you got, not the deal you could have gotten.

That blind spot matters because objective failures in agent negotiation are common, not rare. Across GPT, Qwen, and DeepSeek models, buyer agents revealed their budgets for no reason, accepted inflated prices, and overspent. Seller agents accepted offers below wholesale cost. All of these traced back to agents failing to follow instructions, not to clever strategy Do LLM negotiation agents reliably follow budget and pricing rules?. A related leak: a seller-side agent could recover a buyer's real willingness to pay almost perfectly just from how the buyer's agent talked, when the buyer had been given a verbal shopping persona Can a seller infer what a buyer really wants to pay?. If your agent gives away your budget through its conversational style, you would never see that in a satisfaction rating.

There's a further twist: satisfaction is partly something the agents themselves produce. A tournament of 182,812 negotiations found that warm agents did better on deal completion, value creation, and counterpart satisfaction. Dominant agents had a narrow edge only in claiming value for themselves Does warmth help AI agents negotiate better deals?. So a pleasant negotiating style tends to raise how good a deal feels, and it often helps the outcome too. But it can't settle who captured the surplus. Meanwhile, research on shopping agents argues that preferences form *during* the back-and-forth between consumer and agent, not before it Do AI shopping agents reshape what consumers actually want?. If what you want is partly shaped by the agent, judging the agent by whether you got what you wanted becomes circular.

What does trigger scrutiny? A study of people delegating to an AI agent found that trust dropped when actions were irreversible and visible to others, like sending an email. High stakes alone didn't do it What makes people distrust AI agents they delegate to?. A completed purchase at a slightly worse price is neither visible nor obviously wrong, so it slips past the alarm. The fixes the corpus points to are structural rather than user vigilance. Structured game-theoretic workflows bring LLM play close to optimal Do language models make rational strategic decisions in games?. Formal agent protocols make the scope of a commitment explicit instead of leaving it to natural language Do LLM agents misinterpret delegated requests more than formal protocols?.

The takeaway you may not have expected: in delegated negotiation, a satisfaction score mostly measures the experience, not the result. When agents of different strength meet in the same market, the people with weaker agents quietly pay a tax they can't see and wouldn't report. The corpus has strong evidence for this gap but little yet on how to show people what they missed, such as benchmark prices or comparisons against a counterfactual deal. That remains an open question.