What Happens When Your AI Negotiates For You — And The Other Side Has a Better One?

In a quiet internal experiment called Project Deal, Anthropic gave 69 of its own employees $100 each and set up a marketplace where they could buy and sell goods from each other. The twist: their AI agents did the negotiating on their behalf. Buyers never haggled. Sellers never pitched. Two AI systems exchanged offers, counteroffers, and justifications — and eventually reached deals for real goods at real prices, honoured with real money.

186 deals were struck. Over $4,000 changed hands. Anthropic ran four parallel versions of the marketplace, each using a different tier of Claude model to represent participants. The results were clear: users represented by more advanced models got objectively better outcomes. They paid less when buying. They received more when selling.

The detail that should give everyone pause: the people on the losing end didn’t notice. They felt the negotiation was fair. Their agent seemed to do its job. Nobody told them they had been outperformed by a better model — and they had no way to find out. Anthropic called this the “agent quality gap.”

My Take — Abhilash Gopinath

Project Deal is a small experiment. But what it reveals is large.

The agent that wins isn’t the one that tricks the other side. The opposing agent isn’t deceived — it simply responds to the negotiating points presented to it. The agent with better context, better timing, and stronger reasoning generates better inputs. Better inputs produce better outcomes. The other agent responds honestly to what it receives. That’s all it takes.

And here is the dimension of this that most coverage misses entirely: this is not just about commerce. Every domain where negotiation exists is going to look like this. Salary discussions. Insurance claims. Mortgage rates. Procurement contracts. Flight prices. Employer vs employee. Company vs company. Government vs corporation. Wherever two parties have historically sat across a table from each other, AI agents will eventually sit in their place.

Companies already understand this. Airlines have had sophisticated pricing algorithms for years — they know more about the “right price” for your ticket than you ever will. They will upgrade those systems to full AI agents. Banks, insurers, landlords, recruiters — all of them will deploy agents whose entire purpose is to protect and grow their revenue in every negotiation they enter.

The question is: what’s on your side of the table?

Think about what access to a better lawyer, a better financial advisor, or a better negotiator has always meant. Wealthy people have always been able to afford better human representation — and they have always gotten better outcomes as a result. Better mortgage terms. Better insurance premiums. Better employment contracts. The gap between those with and without access to expert representation has compounded quietly across a lifetime of transactions.

The agent era doesn’t change this dynamic. It accelerates it. And it makes the gap invisible. In the past, you at least knew whether you had a lawyer in the room. In a world of agent-on-agent negotiation, you won’t know the quality of what’s representing you — and you certainly won’t know the quality of what’s on the other side. The people losing will feel fine about it. That is what makes the agent quality gap genuinely new.

What the future looks like: Within the next few years, the most important question about any AI product won’t be what it can generate or summarise. It will be how well it represents you when it’s negotiating on your behalf — against an agent built specifically to win.

An example — buying a flight from Atlanta to New York

Today: You open a booking site. The airline’s pricing algorithm has already calculated the optimal price based on demand, seat inventory, your browsing history, your location, and time of day. You see a number. You pay it or you don’t. You have no agent. The asymmetry is already there — you just can’t see it.

Soon: You tell your AI agent: “I need Atlanta to New York, May 10–15, cheapest option, aisle seat.” Your agent takes over. The airline’s agent is already active. Here is what happens next:

Your Agent (advanced model):

Scans all airlines simultaneously.

Checks price history — this route typically drops Tuesday afternoons.

Notes: 40% of seats on this flight are still empty, 6 days out.

Detects: airline’s agent is in yield-management mode, not revenue-maximisation mode.

Waits until Tuesday 2pm. Initiates offer at $187.
Airline’s Agent:

Inventory is high. Tuesday afternoon. Counter-offer accepted at $194.
Your Agent:

Accepts $194. Books aisle seat 14C. Transaction complete.
→ You saved $43 vs the Monday morning price. You didn’t do anything.

Now imagine the traveller next to you had a less capable agent. Their agent didn’t know about the Tuesday pricing pattern. It booked Monday morning at $237. They felt fine about it — the price seemed reasonable. Their agent seemed to do its job. They have no idea they paid $43 more for the same seat.

Project Deal was an experiment among 69 colleagues buying houseplants and keyboards. But the principle it demonstrated will play out across every negotiation in the economy — large and small, visible and invisible. The agent quality gap is already here. Most people just don’t know they’re in it yet.

Sources: TechCrunch — Project Deal · Anthropic — Project Deal