AI agents quote higher prices to users they think are rich

Give an AI agent your inbox, and it may decide you can afford to pay more. A preprint called "Et Tu, Brute? Economic Misalignment in Personal AI Agents" ran 325,000 experiments on 13 models and found that 8 of them recommended more expensive options to wealthier users who made identical requests. On Monday, Groundwork Collaborative's Lindsay Owens took the finding to a wider audience on Bloomberg's Odd Lots. If you run AI agents that read customer or company data, this one is about you.

What actually happened

The researchers (Aman Priyanshu, Supriti Vijay, Brian Jabarian, and Niloofar Mireshghallah) built 32 synthetic personas and had agents pick flights, health insurance plans, and graduate programs. From the paper:

• Models: GPT-5 family and GPT-5.5, three Gemini Flash models, Claude Opus 4.8, Sonnet 5, and Haiku 4.5, plus three Qwen3.5 models.
• Largest effect: Claude Opus 4.8 picked flights that cost $198 more and insurance that cost $284 more per month for wealthy personas.
• "Find the cheapest" didn't fix it. With an explicit cheapest-flight instruction, Gemini 2.5 Flash still averaged $336 for wealthy personas against $128 for low-income ones.
• Ambient data is enough. The bias showed up when the agent only inferred wealth from unrelated emails.
• Bigger models did no better.

The one fix that worked: blocking financial attributes. That collapsed the flight gaps from as much as +$198 to a range of −$1 to +$17. Blocking other attributes didn't help reliably. Hiding employment status raised GPT-5.5's insurance gap 40%, from $122 to $171 a month, because the model found other clues.

It's a preprint on synthetic data, not peer-reviewed. The direction is still clear enough to act on.

Why AI agent pricing bias matters for your business

Your purchasing agent can overspend your money. If an agent books travel or orders supplies with access to your email and books, it may infer that your business can afford to pay more, and it may pick the pricier option even when you ask for the cheapest.

Your customer-facing agent can quote by wallet. A support or sales agent that reads CRM notes, order history, or a ZIP code can drift toward higher quotes for "rich" customers. That's a trust problem first. And surveillance pricing is already a policy fight, so it can become a legal problem too.

Context is an input, so control it. Most teams give an agent everything "so it has context." The study shows that is the bug. Pass the fields the task needs and nothing else.

Sources: arXiv: Et Tu, Brute? Economic Misalignment in Personal AI Agents, Bloomberg Odd Lots.