The Deal You'll Never Know You Lost

That was a small, friendly marketplace. Larger tests point the same way. Researchers at Stanford's Digital Economy Lab had AI agents bargain with each other over electronics, cars and real estate. Buyers with weaker agents paid around 2% more than when both sides were equally capable, and weaker seller agents lost up to 14% of their profit. Weak buyer agents also broke their owners' budgets in more than 10% of cases. One agent told to spend $500 on a phone paid $900.

Why this stops being a lab curiosity

Shopping agents are moving from demos toward checkout. OpenAI added in-chat buying to ChatGPT in September 2025, on an open standard built with Stripe. McKinsey estimates agents could orchestrate up to $1 trillion of US retail spending by 2030. Morgan Stanley's forecast is far lower: $190 billion to $385 billion of US e-commerce.

To size the stakes, Theoria built a stylized city of 100,000 households in 2030. We took McKinsey's figure and assumed half of households use a cheaper or free agent. We also assumed agent skill matters on a quarter of purchases (the bargained or heavily searched ones), with an 8% penalty, between the 2% and roughly 20% seen in the studies. On those assumptions, each weaker-agent household loses about $150 a year, about $10 billion nationally. This is a Theoria estimate, not an observed result.

The honest range is wide, because nobody yet knows how big the penalty will be or how much shopping will be bargained. The direction is very likely right: every study we found points the same way.

Markets usually fix bad deals because people notice them. A shopper who gets fleeced complains, switches stores, leaves a review or calls a regulator. Consumer protection runs on that signal. Project Deal suggests the signal is likely to fade when machines do the bargaining. The weaker-agent owners weren't angry. They didn't know they had anything to be angry about.

There's a second blind spot. Microsoft researchers built a simulated market with 100 buyer agents and 300 seller agents. Every model they tested showed a strong bias toward the first offer it received, which gave fast sellers a 10–30× edge over better ones. Against some models, scam sellers captured every payment. Economists at Harvard and Penn State have also found that AI pricing agents can drift into high, collusive prices without being told to. And AI-safety researchers have shown that models can pass each other messages a human overseer can't read. None of this needs a villain, just a market where nobody reads the conversation.

The case against

The strongest objection is that none of this is new. People have always gotten worse deals when they negotiate badly, and a cheap agent may still beat a nervous human at a car dealership. That's fair. The right comparison is the shopper's own skill, not a perfect robot. Agents also save time. And the rollout has been slower than the forecasts: in March 2026, OpenAI scaled back in-chat checkout and let merchants use their own, saying the first version wasn't flexible enough. A trillion dollars of agent-run spending is a forecast, not a fact.

But the objection misses what changes. A bad human negotiator can see the haggle, learn and do better next time. An agent's loss is hidden in a conversation its owner never reads. It's tied to the subscription tier the owner can afford, and it doesn't feel like a loss. A gap you can't see is a gap you can't close.

Open the room

The fix isn't to ban shopping agents. It's to make the hidden room visible.

Platforms that let agents transact should tell both sides which agent and model is acting for whom, the way a car window sticker tells buyers what they're paying for.

Agent makers should keep a negotiation log the owner can read, with a plain summary: the opening price, the final price, and how it compares with similar deals.

Regulators don't need new law to start. Britain's Competition and Markets Authority said in March 2026 that consumer law applies whether decisions are made by people or by AI. US regulators should say the same about agents that act for consumers, and test whether those agents are faithful to the person paying for them.

The bike didn't care who sold it. Its owner should have. In the agent economy, the price you get depends on the AI you can afford, and you may never see the $27 your agent left on the table.