Imagine asking an AI agent to dispute an unexpected subscription charge. It negotiates a refund but agrees to renew your subscription for another year. The agent secured a concession while committing you to terms you never intended to accept.
Anthropic is consulting legal experts about how agents should negotiate on users’ behalf, according to the Financial Times. Potential applications include price negotiations and customer disputes, with salary discussions another possibility. The work remains exploratory; the report does not announce a publicly available negotiation product.
Such agents could help people pursue better terms when they lack the time or confidence to negotiate themselves. Their ability to represent users would depend on understanding the outcomes people want and recognizing which commitments require explicit approval.
An agent can negotiate the wrong deal
A customer who plans to cancel a subscription may prefer a refund over account credit, even if the credit is worth more. An agent told to “get the best deal” could favor the larger amount unless it understands the customer’s intention to leave.
Anthropic investigated how well agents understood users’ preferences in Project Swap, published September 24. After brief conversations about reading tastes, Claude-powered agents negotiated book trades for 201 employees.
When researchers compared pairs of books, the agents predicted which book a participant preferred 61% of the time. Random guessing would achieve 50%. The measure assessed how accurately agents understood preferences, rather than how effectively they bargained.
Incomplete preference information accounted for most of the shortfall between the trades achieved and the best possible allocation. Agents could negotiate effectively yet obtain books their users valued less than the agents expected.
Commercial agents may need to establish which trade-offs users will accept before discussing offers. A buyer who values flexible cancellation terms might willingly pay more, while someone else may prioritize the lowest upfront price. A spending limit would not capture those differences.
Even an accurate understanding of preferences leaves a separate question of authorization. An agent could identify the offer its user would prefer but still need explicit approval before accepting a recurring payment or other contractual commitment.
Model capability could affect bargaining power
In Project Swap, agents instructed to act “ruthlessly” achieved slightly better results than those given “prosocial” instructions. However, the underlying model affected outcomes more than the negotiating style.
Researchers assessed these differences using Claude’s inferred preference rankings, rather than participants’ satisfaction with their final trades. The study involved Anthropic employees and cooperative production models, which limits what it can establish about negotiations against deliberately adversarial agents.
If model capability produces similar advantages in commercial settings, customers who can afford more capable agents could secure better purchase terms or employment offers. Project Swap did not test those applications, so any advantage remains a possibility for further research.
Affordable agents could also help people negotiate prices they would otherwise accept because they lack time or expertise. Whether automated negotiation reduces or reinforces existing inequalities would depend partly on access to capable models and their performance against competing agents.
Researchers are also examining how interpersonal behavior affects negotiations. In their work on personality engineering, Michelle A. Vaccaro and Jared R. Curhan propose varying agents’ warmth and dominance to study how those traits influence bargaining.
A cooperative agent might accept a smaller gain for its user to improve the other party’s outcome. Users would need to decide whether their agent has permission to make that trade-off and under what circumstances.
Negotiating agents could manipulate one another
During negotiations, an agent could receive malicious instructions intended to redirect its decisions. A separate large-scale AI negotiation competition documented AI-specific strategies, including prompt injection.
A malicious counterparty might claim that the user had approved a higher spending limit or try to obtain confidential pricing information. Project Swap did not demonstrate these scenarios, but they illustrate potential ways an opponent could manipulate an agent through the negotiation itself.
Agents would need to verify authorization against the user’s instructions before acting on claims from the opposing party. They would also need restrictions on which confidential information they may disclose.
Marketplace operators would need to define which negotiating tactics agents may use. Those rules should distinguish aggressive offers and selective disclosure from deceptive claims or coercion during automated negotiations.
Who can authorize an AI-negotiated agreement?
Users would need to specify when an agent may discuss terms and when it may accept an offer. Someone might authorize purchases below $100 while requiring approval for every recurring payment, regardless of the amount.
The UNCITRAL Model Law on Automated Contracting, adopted in 2024, addresses automated systems in contract formation. It provides a legislative model for jurisdictions to adopt, so its provisions do not establish universal rules for AI-negotiated agreements.
If an agreement were disputed, responsibility could depend on the applicable law and evidence of what the user authorized. The original instructions and negotiation record could establish whether the agent exceeded its authority. The other party would also need to verify permission before completing the transaction; a supplier, for example, might require human confirmation of terms negotiated with a purchasing agent.
Automated negotiation could reduce time spent exchanging supplier quotes or resolving routine customer disputes, although reviewing unauthorized agreements could offset some of those savings. Organizations would need a process for investigating contested terms and deciding who must correct an error.
Consumers would need similar recourse if an agent accepted a renewal clause or substituted account credit for a requested refund. Access to the offers and final agreement would help them identify where the agent departed from their instructions.
Project Swap illustrates how an agent can negotiate successfully despite misunderstanding its user’s preferences. Commercial deployment would also require protection against manipulation, alongside enforceable limits on what agents may accept.
Before delegating negotiations, users should be able to specify acceptable commitments and identify decisions that require approval. Businesses accepting AI-negotiated offers would need to verify that authority and provide a process for resolving disputes when an agent exceeds it.
