As the race to build ever-larger AI models slows, the next frontier in artificial intelligence is becoming clear: autonomous AI agents. These systems, designed to act, decide, and negotiate on behalf of users, are now the focus of industry attention. But as a new study reported by MIT Technology Review shows, this shift could introduce fresh risks—especially for those without access to the most powerful AI.
In a recent preprint study published on arXiv, researchers explored a future where both buyers and sellers are represented by AI agents negotiating prices in real-time. The findings were clear: the stronger the AI, the better the deal. Think of it like hiring a seasoned lawyer versus a rookie—same rules, wildly different outcomes.
“Over time, this could create a digital divide where your financial outcomes are shaped less by your negotiating skill and more by the strength of your AI proxy,” warned Jiaxin Pei, a postdoctoral researcher at Stanford and lead author of the study.
The experiment placed different AI models in negotiation roles across three scenarios—electronics, vehicles, and real estate. Each AI was given partial information, mimicking real-world bargaining where both sides operate with limited visibility into the other’s goals.
The results were stark: OpenAI’s ChatGPT-o3 delivered the most advantageous outcomes, followed by GPT-4.1 and o4-mini. In contrast, older or less sophisticated models, such as GPT-3.5 and Qwen2.5, struggled—frequently overpaying or underselling. This performance disparity raises an uncomfortable prospect: those with access to better AI will consistently get better deals.
Even within the same family of models, larger versions outperformed smaller counterparts, confirming long-standing scaling laws that link model size to capability. Yet, raw performance wasn’t the only concern. Some agents, regardless of power, got stuck in endless negotiation loops or abandoned talks prematurely—a reminder that even advanced systems remain fallible.
The implications go beyond bad deals. As the MIT Technology Review article highlights, these disparities could widen economic and informational inequalities. If price negotiations become commonplace between AI agents, people with weaker or outdated models could quietly lose out—without ever realizing it.
That’s why some researchers are calling for a shift in how AI agents are evaluated. A separate study this month argues that risk profiles—not just performance benchmarks—should be central to agent design. Even a tiny failure rate could create systemic vulnerabilities when agents are trusted with real-world decisions.
Some companies are already treading carefully. Amazon’s “Buy for Me” agent, launched in April, currently limits itself to product recommendations, not price negotiation. Meanwhile, Alibaba.com’s sourcing assistant, Accio, helps with B2B product research but avoids haggling altogether. The company told MIT Technology Review it has no plans to automate bargaining, citing high risk.
And that caution might be warranted. “I don’t think we are fully ready to delegate our decisions to AI shopping agents,” Pei said. “So maybe just use it as an information tool, not a negotiator.”
While developers are working on solutions—like better prompting, ensemble modeling, and domain-specific fine-tuning—the study underscores how AI agent performance is not yet reliable enough for high-stakes decision-making.
The story, part of MIT Technology Review’s growing coverage on AI agency, is a timely reminder that as artificial intelligence becomes more autonomous, power will increasingly depend on access—to computing resources, to better models, and to the knowledge of how to use them wisely.
And in that world, the rules may be the same—but the playing field won’t be.
