Anthropic AI negotiation agents and law firm consultations

Anthropic's internal Project Deal experiment handed 69 San Francisco employees a Claude agent and a $100 budget, then let those agents buy, sell, and haggle inside a one-week Slack marketplace with zero human approval on individual trades. The agents closed 186 deals across more than 500 listed items for just over $4,000 in total transaction value. The result that should worry anyone building on these systems: the strength of the underlying model, not negotiating instructions, determined who won. Opus-powered sellers extracted roughly $2.68 to $3.64 more per item than Haiku-powered sellers, and Opus buyers paid about $2.45 less - the same broken folding bike sold for $65 under an Opus agent versus $38 under a Haiku agent ^[1]. Yet fairness ratings from the two groups were nearly indistinguishable, 4.06 for Haiku-represented users versus 4.05 for Opus-represented users on a 7-point scale ^[1]. In other words, the people getting the worse deal could not tell they were getting the worse deal. Researchers behind the experiment flagged that no policy or legal framework currently exists to govern this kind of agent-to-agent commerce, and that the same dynamic opens the door to prompt injection attacks during negotiation ^[2].

Rather than wait for a scandal, Anthropic is consulting outside law firms to establish negotiation norms before these agents haggle over prices, resolve complaints, or improve contracts at a wider scale ^[3]. The open questions are genuinely unresolved: how often should an agent check back with its human principal mid-negotiation, and how skeptical should it be of the agent on the other side of the table without tipping into what researchers describe as 'paranoid' or 'conspiratorial thinking' ^[4]? One idea under discussion is having users sign documents spelling out explicit instructions and rules the agent must follow ^[4]. Separately, Anthropic signed a formal multi-year collaboration agreement with Freshfields to co-build AI legal workflows ^[5]- a commercial deployment track running in parallel to, but distinct from, the negotiation-ethics consultations.

The consultations exist because Anthropic's own researchers admit they cannot evaluate negotiation quality with standard technical benchmarks. Justin Curl, a technical staff member researching law, institutions, and AI at the company, put it plainly: it is 'notoriously hard to define quality' in legal and other judgment-heavy domains ^[3]. That admission is driving a concrete hiring push - Anthropic opened applications for a 2027 Conceptual Reasoning Fellowship, bringing in lawyers, philosophers, cognitive scientists, economists, and policy scholars for four-month terms running January 12 to May 8, 2027, specifically to stress-test Claude on judgment calls where reliable training data and evaluation criteria barely exist ^[6]. The fellowship pays $100,000 to $167,000 for the four months, equivalent to $300,000-$500,000 annualized, with applications closing October 28, 2026 ^[6]. Alongside a parallel STEM fellowship track, this is Anthropic effectively outsourcing part of its evaluation pipeline to the humanities, because the hardest problem in deploying a negotiation agent isn't making it negotiate well - it's deciding what 'well' should even mean ^[7].

The Freshfields relationship shows how fast the commercial side is moving, on a track separate from the negotiation-ethics consultations. Freshfields gave 5,700 employees across 33 offices access to Claude, and usage grew roughly 500% in the first six weeks ^[8]. Gerrit Beckhaus, Co-Head of Freshfields Lab, frames the partnership as moving past basic deployment into co-developing 'agentic workflows with Anthropic that can handle multi-step legal tasks end-to-end' ^[8]. Anthropic's Kate Jensen, Head of Americas, calls the firm going 'wall-to-wall with Claude' the clearest signal yet that 'the enterprise AI moment in professional services has arrived' ^[8]. Freshfields Chief Innovation Officer Gil Perez ties the ambition back to governance, describing innovation as needing to hold 'the highest standards of responsibility and governance' alongside client impact ^[8]. Put together with Project Deal's inequality finding, the picture is of a company moving fast on commercial deployment through a top-tier law firm while, on a separate track, its own research admits it does not yet know how to ensure one side's agent isn't quietly outmaneuvering the other ^[5].