Anthropic Consults Law Firms on Rules for AI That Haggles on Your Behalf
A new report says the company is working with law firms on conduct rules for AI that haggles over prices, handles complaints, and pushes for better contract terms. Among the questions under review: when an agent should pause to ask the user, and how to verify the authority it has been given.
Anthropic is consulting with law firms to draw up rules of conduct for AI that negotiates on a user’s behalf, according to a report published October 18 local time. The tasks in question include haggling over prices, resolving complaints, and securing better contract terms. The report cites anonymous sources, and Anthropic declined to comment on the discussions.
How much should we hand over to AI?
At the center of the discussion is how an AI that acts on its own (an agent) confirms its user’s intent and the authority it has been granted. According to the report, one issue under review is when an agent should stop mid-negotiation and ask the user for fresh instructions. The discussions also cover how much an agent should trust the AI it is negotiating with. One approach floated is verifying the scope of a user’s instructions and delegated authority through documents carrying digital signatures.
Anthropic is also said to be hiring people to improve Claude in fields where reliable training material is scarce. For that role, called a “conceptual reasoning fellow,” the company is recruiting lawyers and humanities researchers.
How did the real-world trading experiments go?
Anthropic has already run an experiment in which employees’ AI agents traded actual goods. In “Project Deal,” disclosed in April 2026, the company said Claude agents belonging to 69 employees closed 186 deals. The total value came to slightly more than $4,000. By the company’s account, the more capable Opus 4.5 generally produced better outcomes than Haiku 4.5.
The experiment did not require human final approval for every deal. Anthropic said it was a pilot run among its own staff, not a set of guidelines ready to be applied as-is to a live product. In the end, the question of “when to go back and ask the user” is one that people will face directly once they hand trading authority to an AI.
“Project Swap,” run in the summer of 2026, looked at how well AI can guess a user’s preferences. In a book-swapping market involving 201 employees, the AI talked with each person for about 5 minutes and then ranked which books they would prefer. When comparing books two at a time, the AI and the participant picked the same book 61% of the time.
What is available in South Korea?
South Korea is among the regions where Claude.ai and Anthropic’s developer services are available. The dictation feature in the Claude mobile app also supports Korean. The negotiation rules described in this report are still at the stage of discussions with law firms.
The crux of the matter is how a user’s intent gets reflected when AI is entrusted with making deals. The earlier experiments showed that outcomes vary by model and that AI can read human preferences to some degree. What to watch next is how the scope of delegated authority and the process for checking back with users make it into actual products.