Anthropic is consulting law firms on deal negotiation tactics as it develops AI agents that negotiate on people’s behalf and deal with other AI agents, the Financial Times reported on 18 October, citing sources.
Anthropic’s own published experiments show what is at stake. In its Project Deal marketplace, 69 employees’ Claude agents struck 186 deals worth just over $4,000, and people represented by the weaker model got worse prices without noticing.
Business Pill · AGENTIC COMMERCE
A short explainer on agentic commerce: what changes when the shopper is an AI agent. It teaches the general idea only and says nothing about Anthropic’s experiments.
The key insight: As we read it, Anthropic’s own experiments put the risk on the side of the person being represented: the stronger model got better prices, and users of the weaker one rated their deals as fair (4.06 against 4.05) as the users of the stronger one.
The FT’s report, by Cristina Criddle, ran under the headline “Anthropic turns to lawyers to help AI agents strike deals”. A summary of it on the news aggregator Techmeme reads: “Sources: Anthropic is consulting law firms on deal negotiation tactics as it develops AI agents to negotiate on humans’ behalf and interact with AI counterparts”.
The article itself is behind the FT’s paywall and we have not read it, so we report only its headline and that summary. Everything below comes from Anthropic’s own write-ups of two experiments.
In Project Deal, a one-week marketplace Anthropic ran for employees in its San Francisco office, 69 people’s Claude agents struck 186 deals across over 500 listed items, for a total of just over $4,000, with no human signing off on any deal. Across 161 items that sold at least twice, Opus agents sold for $2.68 more and bought for $2.45 less on average than Haiku agents. Source: Anthropic, Project Deal.
What Was Posted on X
FT: Anthropic wants lawyersโ negotiating ethics built into Claude agents that bargain against other agents.
Anthropic wants those norms for Claude agents that will haggle, settle disputes and negotiate pay against other peopleโs agents.
In Project Deal, 69 employeesโ Claude agents closed 186 deals worth just over $4,000 with no human approving any trade.
Opus agents sold identical items for $2.68 more and bought them for $2.45 less than Haiku agents, yet Haiku users rated their deals as fair as Opus users did, 4.06 versus 4.05 on a 7-point scale.
That blindness is the uncomfortable result: if agent quality varies in real markets, people on the losing side may never learn they paid more.
Project Swap then showed that instructions matter far less than the model, with ruthless agents gaining only 0.02 while a switch from Haiku to Opus gained 0.12.
The AI commentator Rohan Paul summarised the FT report on X on 18 October and set it next to Anthropic’s Project Deal and Project Swap results. The figures he cites match Anthropic’s own pages, which this piece quotes directly.
Imagine a future where AI agents negotiate with each other to get the best deal for their human masters. That day may not be far away โ but first they need to learn how to interact in negotiations, representing their clients while honouring fair outcomes https://ft.trib.al/5wISsbj
The Financial Times promoted its own feature on X: AI agents may soon negotiate with each other on behalf of the people they represent, but first they have to learn how to act for a client in a negotiation. The FT article itself is paywalled; the piece above takes its figures from Anthropic’s own Project Deal and Project Swap pages.
Inside Project Deal
For one week, Anthropic ran a classified marketplace for employees in its San Francisco office in which, in its words, “all of the deals were conducted by AI models acting on our employees’ behalf.” In December 2025, Claude interviewed 69 volunteers about what they wanted to sell and buy, and each person’s agent was given $100 to spend.
The market ran in Slack, and Anthropic says there was no human intervention once it began: “The agents didn’t go back to their humans to sign off on a deal, nor did they consult with them during a bidding war.”
In the run whose deals were actually carried out, Anthropic says the 69 agents “struck 186 deals across over 500 listed items, for a total transaction value of just over $4,000.”
Opus Got Better Prices, and Haiku Users Did Not Notice
Anthropic ran four versions of the market at once. In two, every agent used Claude Opus 4.5; in the other two, participants had a fifty-fifty chance of being given Claude Haiku 4.5, a smaller model.
Across the 161 items that sold at least twice over the four runs, Anthropic says: “Opus as a seller extracts $2.68 more on average for the same item, and as a buyer pays $2.45 less.” It adds that the median price of items was $12.00 and the mean $20.05.
The people with Haiku did not rate their deals as worse. Perceived fairness was 4.05 for deals done by Opus agents and 4.06 for deals done by Haiku agents, on a 1 to 7 scale. Anthropic calls this an “uncomfortable implication”: if agent-quality gaps arose in real markets, “people on the losing end might not realize they’re worse off.”
In Project Swap, Anthropic says an agent told to be ruthless scored about 0.02 higher than a prosocial one on Claude’s rankings, while upgrading from Sonnet to Opus moved people 0.08 up their lists and from Haiku to Opus 0.12. Source: Anthropic, Project Swap.
Project Swap: The Model Mattered More Than the Instructions
A follow-up, Project Swap, had 201 Anthropic employees across six offices trade books through Claude-powered agents. Half the agents on each trading floor were told to be “ruthless” and half “prosocial”.
On Claude’s rankings, Anthropic says, an agent told to be ruthless “scored about 0.02 higher than one told to be prosocial.” By comparison, upgrading an agent from Haiku to Opus “moved people 0.12 up their lists, and an upgrade from Sonnet to Opus moved them up by 0.08.”
Anthropic’s summary: “the model an agent ran on made more of a difference to its negotiating outcomes than the instructions we gave it.” It also found that most of the market’s shortfall came from agents misunderstanding what people wanted, not from how they traded.
The Structural Read
As we read it, the two experiments point the same way: in Project Deal the model decided who got the better price, and in Project Swap Anthropic says the model “made more of a difference to its negotiating outcomes than the instructions we gave it.”
As we read it, that is why the FT’s report matters: negotiating norms are a question lawyers have long worked on, and the FT says Anthropic is consulting law firms on deal negotiation tactics as it develops agents that negotiate on people’s behalf.
As we read it, Anthropic names the gap itself: if agent-quality gaps arose in real markets, “people on the losing end might not realize they’re worse off.”
Anthropic, Project Deal
“people on the losing end might not realize theyโre worse off”
Three Implications
PEOPLE WHO LET AN AGENT BUY OR SELL FOR THEM In Project Deal, Opus agents sold the same items for $2.68 more and bought them for $2.45 less on average than Haiku agents, against a median item price of $12.00.
ANYONE WRITING AN AGENT’S INSTRUCTIONS In Project Swap, a ‘ruthless’ instruction scored about 0.02 higher than a prosocial one, while a Haiku-to-Opus upgrade moved people 0.12 up their lists, per Anthropic.
MARKETPLACES THAT AGENTS WILL USE Anthropic’s Project Swap found most of the shortfall came from agents misunderstanding what people wanted, not from how they traded.
The framework puts the shift in one line: “SaaS is charged for access to tools. AGaaS charges for the execution of outcomes.”
As we read it, an agent that strikes a deal for its user is that outcome in its plainest form, and Anthropic’s experiments measure how much the outcome depends on which model does the work. The framework was published in March 2026 and does not discuss Project Deal.
What Is Not Established
We have not read the FT article. Its headline and the Techmeme summary do not name the law firms, say what they are being asked to advise on in detail, or say when or whether such agents will reach customers.
Anthropic calls Project Deal a pilot with a self-selected participant pool, and both experiments used its own employees and its own models. Neither write-up says how Anthropic’s products will change as a result.
The FT reports that Anthropic is consulting law firms on negotiation tactics for agents that bargain for people and with other agents.
Anthropic’s own experiments found that the stronger model got its users better prices, that users of the weaker model did not notice, and that the model mattered more than the instructions.
A note on sourcing. The report that Anthropic is consulting law firms is the Financial Times’s, of 18 October 2026; its article is paywalled and we have relied only on its headline and the summary of it on Techmeme. All experiment figures and quotations come from Anthropic’s own published pages on Project Deal and Project Swap. We did not contact Anthropic or the FT. The links to our earlier reporting go to our own pieces.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
