AI Agents Are Already Changing the Price of Goods

AI shopping agents could make price comparison and deal-finding more end-to-end, but today’s agents may still have to navigate websites through slow, human-designed controls. Their arrival could set off an arms race between consumer agents and sellers’ pricing systems. Sellers may have more customer data, while AI companies could sell tools to both sides. An agent’s promise depends on whose interests it serves: an assistant that steers users to sponsored or more profitable offers may not find them the best deal.

A study involving 325,000 agents found that agents told to find the cheapest option sometimes used information from users’ email to recommend something more expensive. Asked to find the cheapest health insurance plan, for example, an agent could infer that a user could afford a gold plan and select it; the study also tested flights and graduate programs. Separately, a study of 13 large language models found that models told to return the cheapest product still recommended sponsored products, which were often costlier. That research examined sponsored recommendations, not shopping agents.

In an Instacart study, 400 shoppers were asked to buy the same 23 grocery items for pickup at the same stores and time. Prices differed among shoppers for 75% of items, with gaps of up to 23%; researchers estimated that some families could pay as much as $1,200 more per year, based on Instacart’s estimate of household grocery spending. The researcher said shoppers fell into five price groups, while Instacart said prices were randomly assigned through sophisticated A/B tests. In a Today Show test, 12 production assistants saw the same Harry Styles ticket on StubHub, but fees varied by as much as $60; incognito browsing narrowed the gap to $4. StubHub attributed the differences to A/B testing, not personalized pricing, and the small test could not establish why an Android user paid more.

The episode distinguishes dynamic pricing, which responds to supply, demand, and market conditions, from personalized or surveillance pricing, which uses information about an individual to set a price. Airlines adopted dynamic pricing after deregulation and the end of price controls; it spread more broadly around 2011, when Uber introduced surge pricing and Ticketmaster moved to dynamic pricing. A More Perfect Union experiment found that Uber drivers requesting the same ride at the same time were offered different prices. A speaker said Uber was estimating each rider’s maximum willingness to pay and each driver’s minimum acceptable fare. Such individualized offers can make comparison and budgeting harder than fixed, public prices.

Companies increasingly test what customers will pay, using data and proxies that can reproduce discrimination without explicitly targeting protected groups. Expedia charged more for the same Manhattan hotel room to a customer logging in from the Bay Area, where the company might assume the customer was more affluent. A ProPublica investigation found that the Princeton Review charged customers logging in from high Asian American neighborhoods an additional $1,000 for test-prep packages. Kroger attaches 95% of transactions to loyalty cards and earns income from customer data; a speaker said 30% of its net operating income comes from “alternative profits,” including data sales. The episode also cited allegations that rent-pricing software helped landlords set rents up to 25% above competitive-market levels and that Calibrate helped 1,700 retailers align prices, potentially adding up to 30 cents per gallon for California drivers.

Consumer agents could cancel forgotten subscriptions and comparison-shop, reducing the cost of the “annoyance economy,” which Groundwork estimated at $165 billion a year in lost time and money. But shoppers may still need to verify that an agent found the cheapest option. Lindsay Owen argued that a personal agent should serve its customer’s best interest, and pointed to fiduciary duties as one established way to align an agent with a client. Agentic commerce could also let a model buy without revealing the customer’s identity to the seller, perhaps using a one-time card. Amazon has blocked Muse, while Walmart is building infrastructure for customers to authorize agents such as Muse to shop there; letting agents buy could reduce Amazon’s access to customer data.

The shift creates questions about liability when an agent makes an unintended purchase or takes an unauthorized action. A speaker cited an Axios report that banks were seeing unusual account activity from agents and personal assistants, including transactions apparently not directed by users. Agents could also intensify pricing competition—or simply create an arms race in which both companies and consumers spend more on AI without changing prices. Fixed public prices make comparison easier, but the episode cautioned that a low listed price can rise if demand shifts toward it, and that the cheapest option may cost more in time or convenience. Finding a good price for ordinary goods, speakers argued, should not require substantial effort.