AI agents are starting to negotiate: the real issue is no longer price, but authority
An agent can search for suppliers, compare offers and close a deal without getting tired or missing a counteroffer. But when it negotiates for a company, a harder question than getting the best price appears: what may it promise, disclose or give up to reach that goal?

01
From recommending an offer to representing us
On 18 October 2026, the Financial Times reported, citing people familiar with Anthropic’s plans, that the company is consulting law firms on how agents should act when negotiating for users. The questions range from purchasing and customer service claims to agreements with broader financial consequences.
This is not an announced product or an available feature. It is a signal of where agent development is heading: from finding information and drafting proposals to representing interests, making commitments and reaching agreements with people or other agents.
That transition changes the nature of risk. A wrong answer can be corrected; a bad negotiation can reveal a budget, accept a contract term or create an expectation that is difficult to undo.
02
Experiments have already moved beyond the simulator
Anthropic tested the idea with real goods. In Project Deal, 69 agents negotiated the purchase and sale of more than 500 items among employees. They reached 186 deals worth more than $4,000, identifying counterparties, proposing prices, responding to counteroffers and closing transactions in natural language.
The experiment also showed that model capability mattered. Agents based on Opus completed more deals and, when selling the same items, earned an average of $3.64 more than agents based on Haiku. In one example, a lab-grown ruby sold for $65 with Opus and $35 with Haiku.
Project Swap expanded the analysis to 201 participants and a book-trading market. Agents first interviewed each person to infer preferences and then negotiated with other agents. The results exposed two separate problems: understanding what a user wants and respecting it when new opportunities arise during negotiation.
03
A mandate needs more detail than “get the best deal”
Price is rarely the only business criterion. A buyer may prefer a slightly more expensive supplier because of delivery time, financial stability, support quality or cancellation terms. An event manager may accept a higher cost to reduce technical risk. An agent optimising a single number can destroy the very value it was meant to protect.
A mandate should distinguish objectives, limits and preferences. The objective might be to lower total cost; limits could forbid revealing the maximum budget, accepting exclusivity or exceeding an amount without approval; preferences might favour flexibility, proximity or an existing supplier relationship.
It also matters who is accountable when an agent gets it wrong. Traceability should make it possible to reconstruct what information it received, what offers it made, which rules it applied and who approved the irreversible step. Without that record, automation can save time while weakening the organisation’s contractual position.
04
Which business processes could change first
Autonomous negotiation does not need to begin with complex contracts. It may first appear in repetitive purchasing: requesting comparable quotes, asking for volume discounts, checking delivery dates or renewing services within pre-approved terms.
It also fits operations with perishable capacity. Hotels, transport, audiovisual rentals and event venues manage resources whose value changes over time. An agent can quickly explore combinations and counteroffers, provided it cannot commit inventory, margins or terms outside an authorised range.
In customer service, the same architecture could handle compensation or service changes. The agent proposes a solution within a verifiable policy; sensitive, ambiguous or financially significant cases go to a person before the agreement is confirmed.
05
Useful autonomy stops before the signature
Agent negotiation studies show that human dynamics do not disappear. An academic study of more than 180,000 negotiations found that warmth correlated with better overall outcomes, while dominant strategies could capture more value. It also identified tactics specific to AI systems, including injecting instructions during a conversation.
A reasonable business deployment should therefore separate exploring, proposing and committing. An agent can investigate options and negotiate within a narrow scope; a verifier checks policies and data; a person approves exceptions and agreements with significant legal or financial effects.
The advantage will not come from delegating every decision. It will come from turning the company’s mandate into rules a machine can apply and a person can audit. When agents begin negotiating with one another, the quality of those rules will matter as much as the model itself.
Sources
- Financial Times · Anthropic turns to lawyers to help AI agents strike deals, 18 October 2026 ↗
- Anthropic · Project Deal: our Claude-run marketplace experiment ↗
- Anthropic · Project Swap: What happens when agents trade for us? ↗
- arXiv · Advancing AI Negotiations: A Large-Scale Autonomous Negotiation Competition ↗
Where could this apply to your business?
I help teams turn technology into clear, measurable processes, with sensible controls around the work that matters.