Anthropic turns to lawyers on how AI agents should negotiate
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Anthropic wants law firms to help it define how AI agents working for someone should haggle without giving up on a fair result. In Project Deal, an internal test, Claude agents closed 186 deals for 69 employees. With Claude Opus 4.5 rather than Claude Haiku 4.5, participants sold for more and bought for less, and those on the losing side did not realize it.
Anthropic is asking law firms for advice on how AI agents that act for people should negotiate. The Financial Times reported it on Sunday, in a story picked up by The Decoder.
The company wants to learn how a lawyer defends a client's interests in a negotiation while keeping an eye on a fair result. Those guidelines would serve agents that, on their users' instructions, would bargain over prices, settle complaints or try to get better terms in a contract.
The issue goes back to an experiment of its own, Project Deal. For one week in December 2025, Anthropic set up a buy-and-sell market among staff at its San Francisco office in which every negotiation was handled by agents running on Claude, its AI assistant. In all, 69 employees took part, each with a budget of $100, and in the real version of the market their agents closed 186 deals for a little over $4,000. Afterwards, the items changed hands for real.
In other, parallel versions of the same market, some participants were represented by Claude Haiku 4.5, the company's smallest model, instead of Claude Opus 4.5, the most advanced one it had at the time. With Opus, the same item sold on average for $2.68 more and was bought for $2.45 less. In the survey that followed, those who had the small model did not seem to realize they were losing out. Anthropic warned at the time that, if such gaps between agents reach real markets, whoever ends up worse off might not know it.
It is not its first step in this area. According to The Decoder, Anthropic is already looking for lawyers, among other profiles, for a four-month program that will test Claude on questions requiring judgment in economics, law, philosophy, political science and cognitive science, with pay of between $100,000 and $167,000.
For now, Anthropic has not made public any document on these norms.
Why it matters · analysis and opinion
An AI company turning to lawyers to teach its agents how to negotiate shows that the challenge is no longer only technical but one of rules: how hard an agent working for someone can push without overstepping with the other side. Its own experiment left an uncomfortable signal, because the chosen model changed the outcome and the person at a disadvantage did not notice. If these agents reach shopping, complaints or contracts, clear norms will be needed, along with a way to check that they are followed. Almost everything practical is still unknown: which norms will come out of these consultations, when, and whether they will be public. Anyone thinking of handing purchases or complaints to an agent would do well to review each deal before accepting it.
Source: Financial Times · Written with the help of AI: how we make the news
