Anthropic is consulting with law firms on how AI agents should negotiate on behalf of people: bargaining over prices, resolving customer complaints, and securing better salary offers. As reported by the Financial Times, the company is studying lawyers' experience in balancing client interests with fair bargaining rules and is considering 'signed documents' from the user as a way to limit an agent's authority. This principle has already been tested in practice: in September 2026, Anthropic employees traded books via Claude agents on a virtual platform, while a study by Harvard Business School professor Michelle Vaccaro, covering more than 180,000 negotiations between agents, showed that such systems have a weak spot—vulnerability to prompt injection.
What happened
According to the Financial Times, Anthropic is consulting with law firms on the rules by which AI agents will conclude deals on behalf of people. The topics cover purchase prices, resolution of customer claims, and salary offer negotiations. Lawyers are primarily interested in the principles of their own profession: how they balance client interests with fair rules, and how to transfer these principles to agents. The mechanism of 'signed documents' is also being discussed separately: the user signs a document in advance, fixing the framework of the agent's authority in the deal. A practical test took place in September 2026: in an internal experiment, Anthropic employees used Claude agents on a virtual trading platform to buy and sell books. Agents with a 'hardline' strategy achieved slightly better terms, while cooperative agents agreed to worse options to ensure the deal was closed.
Context
The news does not describe the launch of a new product, a new model capability, or a reproducible benchmark—it is institutional and legal preparation for agentic transactions. The essence of the step is to transfer to the 'agent negotiates for a person' scenario the rules that the legal profession has developed in the human world, before this framework is set by regulators or individual platforms. Negotiating on behalf of a user is being formalized as an independent agentic scenario, and competition is shifting from model quality to the mandate interface: how a person sets limits, authorities, and the agent's style, and how deal terms are verified. A separate background is created by the study of Harvard Business School professor Michelle Vaccaro, based on the analysis of more than 180,000 negotiations between agents: it showed that agents are capable of applying prompt injection, i.e., hidden commands in text, to extract private information from an opponent.
Why this matters for the industry
For the industry, this is the de facto beginning of the formation of norms and a legal framework for agentic transactions: who bears responsibility when an agent 'binds' a user to a deal, and how to verify an agent's mandate. For teams building autonomous trading agents, this work sets de facto security standards—protection against prompt injection and a human confirmation point—and they may be established before formal regulation. The practical conclusion is already clear: incoming messages from other agents should be treated as potentially attacking input, and an agent cannot be given the right to finalize obligations—payment, signing, sending an offer—without user confirmation. Verifiable mandates, authority limits, and protection of negotiations from manipulation, judging by the direction of Anthropic's work, may become a purchasable infrastructure layer, and startups are better off in a neutral trust layer. Such a 'mandate' UX can be prototyped today on top of existing LLM APIs: a form with a budget limit and exit point, a preview of deal terms, and user confirmation.
Why this matters for users
Agents that negotiate for users are no longer hypothetical: scenarios are already being tested—from buying a laptop to salary negotiations. This leads to two practical risks. First, more powerful models may give wealthy users an unfair advantage in bargaining, increasing the inequality of outcomes between those who can afford a powerful agent and those who cannot. Second, an agent is vulnerable to manipulation via prompt injection: the opposing party may use hidden commands to extract private information from it or derail it from the set bargaining line. Therefore, before delegating a deal to an agent, it is worth clearly limiting its authority: fixing the budget and unacceptable obligations, requiring a preview of terms and final confirmation—according to the logic of 'signed documents'—instead of full autonomous right to conclude a deal.
What is still unknown / limitations
The methodology of Anthropic's internal experiment has not been disclosed: the sample size, models used, and evaluation metrics are unknown. Michelle Vaccaro's study is known through retellings, and there is no link to a preprint or publication in available sources, although the sample, according to the description, exceeds 180,000 negotiations. The 'signed documents' mechanism is described at the level of a declaration of authority; it is unclear how it will be formalized and whether its execution is ensured at the system level, not just at the level of intentions. There is no product yet: no public API, no pricing, no timelines—it is about legal consultations and an internal experiment. Predictions about the publication of the methodology or a preprint in the coming months and the formation of industry standards remain interpretations, not established facts.
Sources
Author
Look at AI, editorial team