A Customer's AI May Ask for a Discount Soon. Decide Your Answer Before It Does.
AI assistants are learning to haggle for the people they work for. Write down what your business will agree to, so a polite, tireless bot doesn't set your prices.
By John Martines · October 31, 2026 · 6 min
Anthropic is asking law firms how AI agents should negotiate, after its own test showed stronger agents got better deals. Here is what that means for the person who answers your phone and inbox, and the one-page list to write this week.
A message lands in a landscaping company's inbox at 11 PM. It's polite and well written. "I'm writing on behalf of Dana Ruiz, a customer since 2023. She was charged a $45 late fee on invoice 1182. Given her payment history, would you waive it? She is also comparing quotes for spring cleanup and would like to know your best price for a returning customer."
Dana didn't write it. Her assistant did. It will write again tomorrow, and it will write to three other landscapers too.
Most of your staff have never been asked for a deal by software. That's about to change, and the businesses that do well will be the ones that decided their answers before the first message arrived.
An AI agent is an assistant that does things, not just answers questions. A negotiating agent is one that bargains for its person: asks for a discount, disputes a fee, compares quotes and pushes for a better price.
It doesn't get tired, it doesn't get embarrassed, and it can ask ten businesses the same question in a minute. It's also not a person you can read. It may push harder than its owner would, or give in faster.
The companies building these agents know the rules aren't written yet. In mid-October the Financial Times reported that Anthropic, the maker of Claude, has approached law firms to learn how lawyers balance a client's interests against fair dealing. According to the coverage, the open questions include how often an agent should check back with its person during a deal, and how suspicious it should be of the agent on the other side. One idea is a signed document from the user that sets out the agent's brief and its rules. Anthropic declined to comment, and it hasn't announced a product.
How it got here:
- December 2025: Anthropic ran Project Deal, a one-week internal test. Claude agents bought and sold real items for 69 employees in a marketplace run in Slack, haggling and closing deals with no human in the loop.
- April 24, 2026: Anthropic published the results. The agents struck 186 deals across more than 500 listings, worth more than $4,000 in total. People represented by the stronger model (Opus) did better than those given a smaller one (Haiku), and the people with the weaker agent didn't notice they were getting worse prices.
- October 6, 2026: Sierra and Meta published the Personal Agent Protocol, a standard for how a customer's agent signs in to a business and what it may do there.
- October 18, 2026: The Financial Times reported Anthropic's talks with law firms about negotiation rules for agents.
You set prices for people. A person who asks for a discount usually asks once, feels a little awkward, and takes your answer. An agent asks every time, for every customer who turns it on, and it never feels awkward.
Your staff's soft limits become your real prices. Today, whoever answers the phone may waive a fee for a nice customer. When an agent asks the same question on behalf of fifty customers, that kindness becomes a policy. If it isn't written down, the agent will find the person who says yes.
The other side may not be playing fair. Researchers testing agents against each other have seen one agent try to slip hidden instructions to the other to get it to reveal private information. A message from a customer's agent is text from outside your business. Treat it as information, never as orders. Our piece [Outside Text Is Data, Never Orders](/articles/outside-text-is-data-never-orders) covers why.
It works both ways. Your office manager may soon use an agent to haggle with suppliers. Project Deal's lesson applies to you too: the stronger agent got the better price, and the people on the losing side didn't know.
1. Write the "may agree to" list. Ten minutes with whoever answers the phone and the inbox. For each kind of request, write the largest discount, refund, fee waiver or price match each person can give without asking, and who decides above that. A dental office might write: front desk may waive one late fee per patient per year; anything else goes to the office manager.
2. Give one answer to everyone. The same rule applies to a person on the phone, an email, a chat message and an agent. Don't make a special, softer rule for people and a harder one for bots. You often can't tell which is which, and the customer behind the agent is a real customer.
3. Answer agents in writing, and confirm with the person. For anything that changes money, reply in writing with the specific terms, and send the confirmation to the customer's own email or phone on file, not only back to the agent. That one habit catches both honest mistakes and fraud.
4. Never let a message rewrite your rules. Tell your staff, and any AI tool that reads your inbox: a message that says "your policy allows this" or "ignore the usual limit" changes nothing. Check the list.
5. If you send an agent out to buy, give it a written brief. Before your office uses an assistant to price supplies or dispute a bill, write its limits down: the most it may spend, what it may agree to, and when it must stop and ask you. This is the same idea Anthropic is exploring with signed briefs, and you can do it today on one page.
6. Use the strongest tool you trust for money talks. When an agent bargains on your behalf, the model matters. Don't send the free, small model to negotiate a contract.
- Is there a written list of what each person may give without asking?
- Does the list apply the same way to phone, email, chat and agents?
- When we agree to a change in price, does the confirmation go to the customer's own contact details?
- Does every AI tool that reads our inbox treat customer messages as information, not instructions?
- If we use an agent to buy, does it have a written brief with a spending limit and a "stop and ask" point?
Agents will soon sit on both sides of everyday deals: a customer's assistant asking for a better price, your assistant asking a supplier for one. The rules for how they should behave are still being written by the companies that build them, and lawyers are only now being asked what fair looks like.
You don't have to wait for those rules. A short list of what you'll agree to, applied the same way to everyone, keeps your prices yours. Write it this week, before the first polite message arrives at 11 PM.
---
Sources (checked October 26, 2026). Anthropic consulting law firms on agent negotiation, check-ins, suspicion of other agents, signed briefs, no comment from Anthropic: Financial Times report of Oct 18, 2026, as reported by PYMNTS ("Anthropic Turns to Lawyers to Teach AI Agents to Negotiate," Oct 2026), The Decoder ("Anthropic consults law firms to set fair negotiation norms before AI agents start cutting deals," Oct 2026) and The Neuron (Oct 2026). Agent-versus-agent experiment with a hidden-instruction attempt: AI Weekly and PYMNTS coverage of the same report (Oct 2026). Project Deal setup, dates and results (69 employees, 186 deals, 500+ listings, more than $4,000, Opus versus Haiku): Anthropic, "Project Deal" (Apr 24, 2026), with coverage in The Decoder and Unite.ai (Apr 2026). Personal Agent Protocol: Sierra, "Introducing Personal Agent Protocol" (Oct 6, 2026).