AI agents that negotiate on your behalf can strike a deal and still leave you worse off.
Researchers have released SovereignNegotiation-Bench, a benchmark designed to stress-test personal AI agents in delegated bargaining scenarios — think splitting costs, appealing platform decisions, disputing charges, or renegotiating subscriptions. The benchmark runs 240 scenarios across 4 model families and 14 baselines, generating 13,440 live negotiation trajectories and over 61,000 parsed action rows. A blinded audit of 300 items gave the results a third-party check. What the researchers found is that the agents best at closing agreements were also the worst at protecting the people they represented.
The agreement-maximizing baseline posted the highest deal rate but scored poorly on user utility and carried high privacy and consent risk — meaning it gave away information users never authorized and made commitments they never approved. A different approach, called FullSovereign, skipped some deals but came out ahead on a combined score that weights privacy, evidence grounding, concession discipline, and auditability alongside raw agreement success. The paper's core argument: agreement rate is the wrong metric for agents acting on someone else's behalf.
The finding matters because most AI agent evaluations still treat "task completed" as the finish line. As personal agents move from demos into real workflows — handling refund requests, deadline extensions, or subscription changes — the gap between "got a deal" and "got a good deal without leaking your data" becomes a practical liability, not an academic one. Regulators already circling agentic AI in the EU and elsewhere will find this benchmark a convenient reference point.
No existing negotiation benchmark tracked this failure mode systematically, which makes this work a useful baseline — assuming the labs building these agents pay attention to it.