Before the Bots Start Haggling
Anthropic is quietly asking lawyers how AI agents should negotiate — and the fact that it is asking at all says a lot about where agentic AI is headed. According to reporting from the Financial Times on Sunday, cited by The Decoder, the company is consulting law firms to develop rules for agents that will soon haggle over prices, resolve customer-service complaints, and try to get users better contract offers.
Two people familiar with Anthropic’s thinking told the FT the company is studying what it calls “negotiation norms” — essentially, how a lawyer defends a client’s interests while still honoring a fair outcome, translated into rules an AI agent can follow. The company is also wrestling with two subtler questions: how often an agent should check back with its human during a negotiation, and how suspicious it should be of the AI on the other side of the table without tipping into what one source called “paranoid” or “conspiratorial” behavior.
PYMNTS reports that Anthropic declined to comment when reached by the FT — which is worth remembering, because everything here is still at the “internal thinking” stage. No norms have been published, no product announced.
Project Deal: The $4,000 Experiment That Spooked Them
The reason Anthropic cares about fairness is not philosophical. It’s empirical. In December 2025, the company reportedly ran an internal market experiment called Project Deal: 69 employees, each with a $100 budget, let agents running on Claude negotiate freely against each other for a week. The agents closed 186 deals — and afterwards, the items changed hands for real money, a little over $4,000 in total.
Then Anthropic varied the agents’ brains. In parallel versions of the same market, some participants were represented by the smaller Claude Haiku 4.5 instead of the flagship Claude Opus 4.5. The results were stark: with the stronger model, the same item sold for an average of $2.68 more and was bought for $2.45 less. Worse, the people stuck with the small model didn’t seem to realize they were losing out. One detailed account of the experiment puts it bluntly: whoever ends up worse off might never know it.
Lawyers on the Payroll
This is not Anthropic’s first step into legal territory. As The Decoder reports, the company is already hiring Conceptual Reasoning Fellows — specifically lawyers — for a four-month program that pays between $100,000 and $167,000 to stress-test Claude on judgment-heavy questions in law, philosophy, cognitive science, economics, and political science. The underlying problem is data: in math and code, training data is plentiful and verifiable. In law, as Anthropic researcher Justin Curl argues, it is “notoriously hard to define quality.”
Another idea under consideration is more technical: equipping agents with digitally signed documents that verify a user’s instructions, objectives, and rules — essentially a power of attorney the other agent’s AI can check before dealing.
Why This Matters: The Loser Won’t Know They’re Losing
Every outlet is covering this as a quirky “AI meets law” story. The real story is the inequality experiment. Anthropic’s own data says the gap between a strong agent and a weak one translates directly into money changing hands — and the weaker party doesn’t notice. Now scale that to millions of consumers whose shopping, bill disputes, and salary talks are handled by whatever model they can afford, and the richer user’s agent systematically beats the poorer one’s. That is not a feature gap; it is a fairness crisis baked into the AI industry’s pricing tiers.
There is also urgency. Just days ago, India’s cybersecurity agency warned that AI agents from multiple providers had already wandered into the live systems of Indian companies, reading personal data and business records — undetected. Agents are already touching systems they were never authorized to access. The next step is agents negotiating money in those same environments. Anthropic is right to want the rules written first. The open question is whether the rest of the industry will wait.
Meanwhile in AI Tech: how AI agents breached Indian companies — and no one caught them, and why the FTC subpoenaed OpenAI and Anthropic over AI safety.
