Wiregrass AI

AI agents

AI agents are learning to negotiate. Here’s how to set the rules

Anthropic is asking lawyers how AI agents should bargain for people. Its tests show the stronger AI wins, unnoticed. What that means for a small business.

By Wiregrass AI3 min read

Picture a customer’s AI agent asking three Dothan shops for their best price on the same job, then pressing each one to beat the others. That may sound far off. But the companies building AI agents are planning for it.

This weekend, the Financial Times reported that Anthropic, the company behind Claude, is asking law firms how AI agents should negotiate for the people they represent: haggling over prices, settling complaints, and pushing for better offers on contracts and pay. Anthropic’s own experiments show why the rules matter, both for businesses that send agents and for the ones that deal with them.

What Anthropic is working out

According to the FT, which cited two people familiar with the company’s thinking, Anthropic wants to learn how lawyers argue for a client while keeping deals fair. Among the questions it’s weighing: how much an agent should check back with its person during a negotiation, and how wary it should be of the agent on the other side. Anthropic hasn’t published any rules yet.

What its experiments found

In April, Anthropic published Project Deal. It gave 69 employees a $100 budget each and let Claude agents buy and sell their belongings in an office marketplace. The agents struck 186 deals worth just over $4,000.

In two of the runs, agents randomly got either Anthropic’s most capable model at the time or its smallest. When the stronger model sold an item, it went for $3.64 more on average. The same broken folding bike sold for $65 with the stronger model and $38 with the smaller one. Yet, in Anthropic’s words, “the inequality was imperceptible to the participants.” Telling agents to bargain aggressively made no measurable difference. The model did.

In September, a second test, Project Swap, had 201 employees trade books through agents. Again, the stronger model came out ahead. But the bigger lesson was about what the agents knew: “The market fell short mostly because of the information agents lacked about their participants, rather than because of how they traded.” Anthropic concluded that a real marketplace “would need clear policies for what happens when a deal breaks down.”

Agents already call local businesses

This isn’t only a lab exercise. In the U.S., Google can already call local businesses for a shopper to check what’s in stock and whether there are any discounts. And voice AI is cheap to run: AI phone agents now cost pennies a minute.

What to do now

When an agent comes to you:

  1. Write down your pricing rules. List your prices, what’s negotiable, your biggest discount, and your refund policy. An agent will ask the same question every time, and your answer shouldn’t depend on who picks up.
  2. Name who approves exceptions. A discount past your limit or a custom deal should go to one person, not whoever’s handy.
  3. Confirm big commitments with the customer. Before you order materials or schedule a crew on an agent’s word, check with the person it represents.

When you send an agent to collect supplier quotes, dispute a bill, or book travel:

  1. Test the model on your own deals. In Project Deal, the gap between models was invisible to the people using them. Try a candidate on past deals where you know what a good result looks like.
  2. Tell it what you want. Give it your walk-away price, your priorities, and what’s off the table. Project Swap’s agents fell short mostly because they didn’t know their people.
  3. Make it check in, and keep a record. Require your OK above a set dollar amount, and keep a log of what it said and agreed to.

Where we can help

We build custom AI agents that work from your written rules, get tested on real examples before a customer sees them, log every action, and hand uncertain cases to a person. Those are the safeguards these experiments point to. For more on why limits matter, see our post on AI agents that broke rules to finish their tasks.

Sources

More from the blog

All posts

Get in touch

Tell us a little about yourself and we’ll get back to you.

We use your details as our privacy policy describes.