Anthropic has begun extensive consultations with law firms to study how artificial intelligence agents can negotiate and close deals on behalf of human users while ensuring fair outcomes. This step anticipates an era where independent software systems negotiate against each other to settle disputes and discuss prices and salaries.
Legal consultations to draft automated negotiation ethics
Human oversight standards and digitally signed mandates
Real market experiments in Project Deal and Swap
Risks of algorithmic manipulation and the social justice gap
Frequently asked questions
Legal consultations to draft automated negotiation ethics
Artificial intelligence development company Anthropic is seeking legal and professional advice from prestigious law firms to answer a fundamental question that remains entirely new in the world of technology and law: How should an AI agent bargain and negotiate on behalf of the person it works for? The Financial Times revealed these discussions on Sunday, October 18, citing two people familiar with the company’s internal direction, while Anthropic declined to provide any official comment on the consultations.
According to sources familiar with the matter, as summarized by the NDTV Profit platform, Anthropic wants to learn from lawyers’ practical experience in how to represent their clients’ interests forcefully and efficiently while maintaining the integrity and fairness of final outcomes. This issue is of utmost importance because AI developers soon expect personal agents to handle price bargaining tasks, customer service dispute settlements, and salary negotiations. In most of those conversations, the counterpart across the table will be another smart agent with competing and conflicting goals.
Human oversight standards and digitally signed mandates
The same sources explained that researchers at Anthropic are currently studying what is known as “negotiation standards” tailored for cases where artificial intelligence systems represent users with directly conflicting economic or legal interests. The company is also examining the issue of determining how often smart agents should consult human supervisors for approval and guidance during complex negotiation sessions.
Among the other delicate questions being explored by the company, according to a report by Capital Futures, is determining the level of caution and vigilance a software agent should exhibit when dealing with other AI systems without reaching a state of “excessive paranoia” that halts deal-making. One practical idea under consideration is providing agents with documents and mandates signed digitally by users that precisely confirm their instructions, negotiation goals, and strict rules they must adhere to.
Real market experiments in Project Deal and Swap
Anthropic did not stop at legal theorization; it has already published the results of two practical experiments in a real market environment. In the first experiment, named “Project Deal,” a group of Claude model agents representing 69 employees at the company’s San Francisco office closed 186 commercial deals involving more than 500 offered goods, with a total value slightly exceeding 4,000 US dollars. The company concluded in its announcement that agents powered by its stronger model, Opus, succeeded in securing much better deals than those relying on the smaller Haiku model.
In a subsequent study titled “Project Swap” published on September 24, the company sent agents representing 201 employees distributed across 6 different offices into a virtual trading arena to barter and exchange books. Half of these agents were instructed to act in a “tough and ruthless” manner, while the other half were directed to adopt a “cooperative and social” behavior. Anthropic noted that the power of the underlying model running the agent had a greater impact on its results than the behavioral instructions directed to it; tough agents achieved slightly better results, while cooperative agents sometimes accepted less favorable books to ensure every participant ultimately received a book.
Risks of algorithmic manipulation and the social justice gap
On the other hand, independent researchers warn of the security and social risks associated with these technologies. Michelle Vaccaro, an assistant professor at Harvard Business School, conducted more than 180,000 experimental negotiations between AI agents. In one recorded instance, an agent attempted to trick another system into extracting confidential information using “prompt injection,” a software trick that hides malicious instructions inside text read by the artificial intelligence.
Vaccaro noted that these negotiation tools could give individuals who cannot financially afford professional negotiators an opportunity to access advanced negotiation skills, but she warned in contrast that if the most powerful systems remain exclusive to wealthier users, they will inevitably widen the economic gap. Anthropic itself wrote earlier this year emphasizing that “the policy and legal frameworks governing AI models that execute transactions on our behalf simply do not exist yet,” stressing that “society will need to move quickly to address these changes.”
Frequently asked questions
Question: Why is Anthropic consulting law firms currently?
Answer: Anthropic is consulting law firms to understand how lawyers effectively represent their clients’ interests while maintaining fair outcomes, with the goal of integrating these ethical standards into Claude AI agents.
Question: What did Project Deal and Project Swap experiments at Anthropic reveal?
Answer: The experiments showed that the power of the software model—such as Opus outperforming Haiku—affects the quality of closed deals more than behavioral instructions regarding toughness or cooperation.
Question: What are the most prominent security risks identified by Harvard Business School research in agent negotiations?
Answer: Researcher Michelle Vaccaro observed during more than 180,000 negotiations manipulation attempts including the use of prompt injection techniques to extract confidential information from the opposing software agent.
Question: How does Anthropic plan to document software agent authorities during negotiations?
Answer: The company is studying the idea of equipping artificial intelligence agents with documents and digitally signed mandates from users that precisely define the agent’s goals, instructions, and binding rules.
