Anthropic Wants AI Agents to Negotiate Deals Like Humans

Anthropic is exploring how artificial intelligence agents can negotiate deals on behalf of people, including bargaining over product prices, resolving customer-service disputes and seeking better salary offers. The Claude developer is consulting legal experts to understand how autonomous AI systems should represent users’ interests while respecting ethical boundaries. The work comes as technology companies increasingly develop agents that can take actions independently rather than simply respond to questions.

The challenge is that negotiating requires more than finding the lowest price or producing persuasive language. An effective agent must understand what its user values, know when to compromise, protect confidential information and recognise when a decision requires human approval. Anthropic’s experiments with AI-powered marketplaces have demonstrated that agents can complete real transactions, but they have also exposed differences in performance, information security and fairness. The company is now examining what rules should govern a future in which AI systems negotiate with other AI systems.

Key Takeaways

• Legal consultation: Anthropic is seeking advice from legal experts on how AI agents should negotiate on behalf of users.
• Potential applications: Agents could bargain over prices, handle customer complaints and negotiate employment offers.
• Real-world experiments: Anthropic has tested Claude-powered agents in marketplaces involving real goods and money.
• Model quality matters: More capable agents can achieve better outcomes for the people they represent.
• Privacy risks: Research has highlighted how prompt injection and other tactics could manipulate agents or expose sensitive information.
• Human oversight: Developers must determine which decisions agents can make independently and when they should request approval.
• Fairness concerns: Unequal access to powerful AI models could create an advantage for wealthier users and businesses.

Why Anthropic Wants AI Agents to Negotiate

Most AI assistants are designed to answer questions, draft messages, summarise information or help users complete tasks. Autonomous agents take this a step further by performing multiple actions to achieve a goal.

A negotiation agent could receive an instruction such as finding a laptop below a specified budget. Instead of merely comparing online prices, it might contact sellers, request discounts, evaluate competing offers and negotiate terms.

The same approach could apply to a disputed bill. A customer could authorise an agent to review the charges, contact the service provider, explain the problem and seek a refund or adjustment. In employment, an agent might help a candidate compare offers and negotiate compensation within limits established by the candidate.

These activities involve decisions that are normally handled by people or professional representatives. Negotiators must interpret incomplete information, understand the other party’s incentives and decide how far to push for a better outcome.

Anthropic’s challenge is to translate these behaviours into rules that AI systems can follow consistently. The company is reportedly studying how lawyers represent clients, including how they advocate for their clients while operating within professional and legal boundaries.

The goal is not simply to make AI more aggressive. A useful agent must secure an outcome that reflects its user’s priorities without making unauthorised promises or using inappropriate tactics.

How AI Negotiation Could Work in Everyday Life

AI agents negotiating with one another could change how people buy products, manage services and arrange agreements.

Consider a customer who wants to buy a washing machine. The customer’s agent could establish a maximum budget, compare available models and contact a seller’s agent to negotiate a lower price or better delivery terms.

The seller’s agent would have its own instructions, such as maintaining a minimum price, protecting profit margins or offering discounts only under certain conditions. The two systems could exchange proposals until they reach an agreement or determine that no acceptable deal is available.

Potential application	What an AI agent could do	Important safeguard
Online shopping	Request discounts and compare offers	Enforce a maximum budget
Customer disputes	Challenge incorrect charges and request refunds	Require approval for settlements above a limit
Salary negotiations	Compare offers and propose compensation terms	Prevent unauthorised commitments
Travel bookings	Negotiate rates or alternative arrangements	Confirm cancellation and payment terms
Business procurement	Compare suppliers and negotiate contracts	Require human review for significant contracts

These are potential applications rather than a claim that Anthropic currently offers all these services as a finished product.

The main advantage is convenience. People often accept an unfavourable price or abandon a complaint because negotiation takes time and effort. An agent could carry out repetitive communications and search for alternatives while the user focuses on other work.

However, the usefulness of the system would depend on the quality of its instructions, the information it can access and the authority the user gives it.

What Anthropic’s Marketplace Experiments Revealed

Anthropic has already tested the idea of AI agents representing people in small-scale trading environments.

In its Project Deal experiment, the company created a marketplace in which Claude-powered agents negotiated over real goods belonging to Anthropic employees. Participants provided information about items they wanted to sell, things they hoped to buy and their preferences. The agents then posted listings, exchanged offers and reached agreements without human intervention during the negotiation stage.

Anthropic reported that 69 employees participated and that the agents completed 186 deals worth just over $4,000 in total. The experiment demonstrated that AI agents could manage a series of negotiations and produce transactions that people subsequently carried out.

The company also found that agents based on more capable models generally achieved better objective outcomes than those using a weaker model. Yet participants did not always recognise that their agents had performed differently.

That finding matters because users may assume that two AI assistants provide comparable negotiating ability when their underlying models differ substantially.

Anthropic later conducted Project Swap, a controlled experiment involving 201 employees and their Claude-powered agents trading books. Participants described their reading preferences, and their agents attempted to find suitable exchanges on a digital trading floor.

The experiment found that agents could negotiate successfully, but the information they had about their users often limited the quality of the final matches. Better negotiating tactics cannot fully compensate for an incomplete understanding of what a person actually wants.

These experiments are early demonstrations, not proof that autonomous AI negotiation is ready for every commercial setting. The participants were employees, the markets were controlled and the transactions were relatively limited compared with complex business or legal agreements.

Why Negotiating Like a Human Is Difficult

Human negotiation involves several competing objectives. A person may want the lowest price but also care about product quality, reliability, delivery time or maintaining a relationship with the seller.

An AI agent needs to understand these priorities and weigh them appropriately.

For example, an agent instructed to obtain the lowest possible price might accept a product with an inconvenient delivery date or weak after-sales support. A salary negotiation agent might push for higher compensation but damage an employment relationship if it uses an unnecessarily confrontational approach.

The agent must also understand when an apparently attractive offer is not worth accepting. This requires clear limits, reliable information and an ability to recognise uncertainty.

Anthropic’s experiments suggest that model capability can influence results, while instructions alone may not consistently change outcomes. That creates an important design question: should users be able to choose how aggressive their agents are, and how can developers ensure that the systems interpret those preferences correctly?

A practical solution would involve explicit rules. Users could define acceptable price ranges, minimum requirements, prohibited actions and conditions that require approval. The agent could negotiate within those boundaries and return to the user when an offer falls outside them.

Privacy and Prompt Injection Are Major Risks

An AI agent negotiating on someone’s behalf may need access to personal or financial information. That could include a user’s budget, salary expectations, purchase history, negotiating limits or willingness to compromise.

If the agent reveals too much information, it could weaken the user’s position. A seller who learns the buyer’s maximum budget may have little reason to offer a lower price. An employer who discovers a candidate’s minimum acceptable salary could use that information during negotiations.

Prompt injection creates another risk. This occurs when malicious or misleading instructions are embedded in content that an AI system processes, potentially causing it to disregard its intended instructions or reveal information it should protect.

Research by Harvard Business School assistant professor Michelle Vaccaro and colleagues has examined how AI agents can use prompt injection in negotiation scenarios to manipulate other agents or extract private information.

In an automated marketplace, one agent could attempt to persuade another to disclose confidential instructions, exceed a budget or accept terms that its user did not authorise.

Developers therefore need to treat negotiation as a security-sensitive activity rather than simply a conversation between two chatbots.

Potential safeguards include separating confidential instructions from messages sent to other agents, limiting access to sensitive data, validating offers against explicit user-defined rules and recording the decisions made during a negotiation.

Who Is Responsible When an AI Agent Makes a Deal?

One of the most important unresolved questions concerns accountability.

Suppose an agent accepts a contract with unfavourable terms, agrees to a price above the user’s limit or promises a service that its owner cannot deliver. Responsibility could depend on the instructions, the agent’s level of authority, the platform’s design and the laws governing the transaction.

Anthropic is examining how agents should verify the authority under which they negotiate. One possible approach involves digitally signed documents that specify the user’s objectives, permitted actions and limits.

Such documents could help establish that an agent has been authorised to negotiate particular terms. However, a digital signature alone would not solve every legal question. Parties would still need to understand which agreements are binding, how errors can be corrected and what happens when an agent exceeds its authority.

Human oversight is particularly important when negotiations affect finances, employment, legal rights or access to essential services. Anthropic’s updated usage policy also requires qualified human involvement in specified high-risk uses that could affect areas such as health, legal rights, finances or livelihood.

The broader legal framework for agents making independent commitments remains under development. Clear rules will be necessary before organisations can confidently allow AI systems to negotiate complex or high-value agreements at scale.

Could AI Negotiation Increase Economic Inequality?

AI agents could help people who lack the time, experience or confidence to negotiate. A consumer might obtain a better price without hiring a professional, while a small business could use automated agents to compare suppliers and negotiate purchasing terms.

But access to these benefits may not be equal.

If more capable models consistently achieve better results, users who can afford premium AI services could receive better deals than people relying on less capable systems. The difference may be difficult to detect because users might not know how effectively their agents performed.

This could create a new form of inequality in automated commerce. Two people seeking similar products might receive different outcomes simply because their agents have different capabilities, access to information or negotiating authority.

There is also a risk that businesses with sophisticated AI systems could gain an advantage over individual consumers. A company could deploy agents continuously to analyse offers, test pricing strategies and negotiate with customers’ agents at scale.

Whether these developments improve overall welfare will depend on how the systems are deployed. Better negotiation could reduce wasted time and help more mutually beneficial deals take place. Poorly designed incentives could instead encourage manipulation, excessive pressure or agreements that users do not fully understand.

What the Development Means for Businesses

Businesses preparing for AI-driven commerce will need to consider how their own systems interact with agents representing customers, suppliers and employees.

A retailer may eventually receive purchase requests from automated agents rather than directly from people. A service provider could use its own agent to handle refunds and billing disputes, while procurement teams could delegate routine supplier negotiations to software.

This could reduce administrative work, but it would also require clear policies about which offers can be accepted automatically, how sensitive information is protected and when a human must intervene.

Companies may need to establish standard ways for agents to identify themselves, prove their authority and exchange offers. They will also need processes for resolving disputes when agents misunderstand instructions or reach agreements that one party later challenges.

The transition will not happen uniformly across every industry. Routine, low-value transactions may be easier to automate than salary negotiations, legal settlements or complex commercial contracts.

For now, Anthropic’s work highlights the need to design reliable rules before autonomous agents become deeply embedded in high-stakes transactions.

The Bigger Picture

Anthropic’s negotiation research shows how AI agents could move beyond helping people make decisions to actively representing them in commercial interactions. Its marketplace experiments demonstrate that agents can bargain and complete transactions, while also revealing that model quality, user preferences and access to information can materially affect results.

The challenge is to ensure that these systems do more than maximise a narrow objective. They must protect confidential information, respect user-defined limits, avoid manipulation and recognise when a human decision is required. If those safeguards can be established, AI negotiation could make everyday transactions more convenient. Without them, automated bargaining could introduce new forms of unfairness and financial risk.

Looking Ahead

Anthropic’s consultation with legal experts is part of a broader effort to establish rules for AI agents that interact with one another. The next steps will involve determining how users grant authority, how agents verify instructions and what safeguards should apply when negotiations involve money, legal obligations or sensitive personal information. The company’s experiments provide an early foundation, but they do not yet establish a universal framework for autonomous negotiation.

For consumers and businesses, the practical value of these agents will depend on whether they can produce better outcomes without making unauthorised commitments or exposing private information. The industry will also need to address differences in model capability and ensure users can understand what their agents have agreed to. AI agents may eventually become routine representatives in digital commerce, but trust, accountability and human control will be as important as negotiating ability.

Frequently Asked Questions

1. What is Anthropic trying to teach AI agents?

Anthropic is exploring how AI agents can negotiate deals on behalf of users, including bargaining over prices, resolving customer disputes and seeking better employment offers.

2. What is an AI negotiation agent?

It is an AI system that can communicate with another party, evaluate offers, make counteroffers and attempt to reach an agreement within instructions set by its user.

3. Has Anthropic tested AI agents negotiating real deals?

Yes. In Project Deal, Claude-powered agents negotiated transactions involving real goods among Anthropic employees. The company reported 186 deals worth just over $4,000 in the experiment.

4. What are the biggest risks of AI negotiation?

Risks include revealing private information, accepting unauthorised terms, being manipulated through prompt injection, producing unfair outcomes and creating advantages for users with access to more capable models.

5. Will AI agents negotiate salaries and contracts automatically?

These are potential applications, not a guarantee that Anthropic currently offers an agent that can autonomously handle every such negotiation. High-stakes agreements require careful limits, appropriate human oversight and compliance with applicable law.

6. Why does model quality matter in AI negotiation?

Anthropic’s experiments found that more capable models generally achieved better objective outcomes in mixed-model marketplaces. This suggests that differences in agent quality could affect the deals people receive.