Anthropic is courting lawyers to teach AI agents how to negotiate and strike deals like humans

Anthropic is reportedly exploring how artificial intelligence agents should negotiate deals on behalf of their human boss. This includes bargaining over prices, resolving customer complaints and even securing better salary offers. The company is seeking legal advice to understand how these agents can represent their users without crossing ethical or legal boundaries. The challenge is becoming more important as AI companies work towards a future where their agents can talk to one another and make decisions with little human involvement. However, when both sides have an AI agent trying to get the best outcome for its user, things could get complicated.According to a report by the Financial Times, Anthropic has been speaking to law firms to understand how human lawyers negotiate on behalf of clients while keeping negotiations fair. The company is also studying how much freedom AI agents should have and when they should ask their users for guidance.How AI agents could negotiate on our behalfImagine asking an AI agent to buy a laptop at the lowest possible price. Instead of simply comparing deals, it could contact a seller's AI agent, bargain over the price and try to secure a better offer. The same idea could apply to rental agreements, customer service disputes and employment contracts.But there is a catch. The agent representing a buyer may have goals that directly conflict with those of the seller's agent. Anthropic is exploring what it calls “negotiation norms” to help agents handle such situations. It is also considering ways to verify an agent's instructions, including signed documents from users that spell out what the agent can and cannot do. In September, the company revealed in a blog post that it tested AI agents in real-world-inspired bargaining exercises. Employees used agents to buy and sell books on a virtual trading floor. Agents instructed to be “ruthless” performed slightly better at securing favourable deals, while more cooperative agents sometimes accepted less-preferred books to ensure everyone got something.Anthropic is not alone in examining this problem. The report Harvard Business School assistant professor Michelle Vaccaro conducted more than 180,000 negotiations between AI agents involving scenarios such as job offers and rental contracts. Her research found that agents could use tactics such as prompt injection to trick other agents into revealing private information. There is also a bigger concern: inequality. More capable AI agents could help people negotiate better deals without needing professional help. But if the strongest models remain accessible mainly to wealthier users and businesses, they could gain an unfair advantage.For now, the legal and policy rules governing AI agents that negotiate and transact on people's behalf remain underdeveloped. As these systems become more independent, deciding when they should push hard, compromise or ask a human for help could become just as important as making them smarter.- Ends