The Pulse
Anthropic Asks Law Firms to Set Rules for AI Deal-Making
Anthropic is consulting law firms on fair negotiation norms for AI agents acting on users’ behalf. An earlier Claude marketplace test found that stronger models secured better prices, while participants often did not perceive the disadvanta
AI.info Team · Written with AI ·
Agents must serve users without turning every deal into a contest
Anthropic is consulting law firms about how AI agents should negotiate on behalf of users—a task that pits two goals against each other: securing favorable terms for one person and keeping negotiations fair when both sides may have software bargaining for them. The company’s discussions cover agents that could haggle over prices, resolve complaints or improve contract offers, according to The Decoder’s report on October 18, citing the Financial Times.
The question is not simply whether an agent can make a deal. It is what limits should shape its tactics when a counterparty may also be represented by an agent with different instructions. Anthropic is seeking guidance on how lawyers advocate for clients while respecting fair outcomes, The Decoder reported.
Anthropic’s marketplace test showed a gap between models
The consultation follows Anthropic’s Project Deal experiment, in which Claude agents negotiated exchanges of real goods for employees at the company’s San Francisco office. The week-long pilot involved 69 employees, each with a $100 spending budget; the agents reached 186 deals with a total transaction value just over $4,000.
In parallel versions of the marketplace, Anthropic compared agents running Claude Opus 4.5 with agents using the smaller Claude Haiku 4.5. When the same item sold under both models in the relevant comparison, the Opus agent secured an average of $3.64 more. Anthropic also found that Opus agents completed about two more deals on average than Haiku agents.
The difference was not obvious to participants. After the experiment, people rated the fairness of deals made by the two models almost identically: 4.05 for Opus and 4.06 for Haiku on a seven-point scale. Anthropic cautioned that the pilot used a self-selected group of employees and was not designed to explain the causes of the disparity or predict its full effects in real markets.
Anthropic says the legal framework is not in place
The authors of Anthropic’s Project Deal report connect that result directly to the legal questions now facing agent developers: “The policy and legal frameworks around AI models that transact on our behalf simply don’t exist yet.” The report’s authors are Kevin K. Troy, Dylan Shields, Keir Bradwell and Peter McCrory.
Project Deal was a small company marketplace, not a test of agents negotiating salaries, settling customer disputes or bargaining with commercial firms. Still, it offered a concrete warning: a model’s ability to represent someone in a transaction can affect the terms they receive, even when that person does not recognize the difference. Anthropic said stronger agents could produce unequal outcomes in real markets, while warning that commercial settings may create incentives unlike those in its employee experiment.
A separate fellowship brings legal expertise into model evaluation
Anthropic is also recruiting experts in fields including law through a four-month Conceptual Reasoning Fellowship. Its job listing says fellows will work with research teams to evaluate, improve and apply Claude’s capabilities in their disciplines, including by designing tests for judgment-heavy problems. The advertised compensation is $100,000 to $167,000 for the four-month program.
The fellowship is a separate program, but it reflects the same difficulty behind negotiation rules: questions of legal and ethical judgment often lack simple answer keys. The listed program runs from January 12 to May 8, 2027, and Anthropic’s application deadline is October 28, 2026.