AI Shopping Helpers Are Already in the Price Fight, and They Err

AI shopping helpers are already part of how online prices get set, and they make mistakes. OpenAI shut down ChatGPT's Instant Checkout on March 20, 2026, about six months after launch. Listings showed wrong prices, shipping costs and stock levels often enough to hurt sales. A separate study found one AI model failed to pick the cheapest item more than 9% of the time.

What went wrong at checkout

Instant Checkout let people buy inside ChatGPT. It launched with Etsy in September 2025. Shopify merchants were expected to follow, and Walmart joined later that year. Walmart executive Daniel Danker told WIRED that purchases made this way converted at roughly a third of the rate of purchases on Walmart's own website. Converted means the shopper actually completed the purchase.

One reason was that the checkout handled a single item. It kept purchases apart from a shopper's existing Walmart basket and stopped the store from bundling related products. OpenAI has since shifted its focus to product discovery and checkout controlled by the merchant. Walmart now runs its own Sparky shopping experience inside ChatGPT, with account linking, synced baskets and loyalty benefits.

Can an AI find the lowest price?

Not always. A study first posted on arXiv in 2025, titled "What Is Your AI Agent Buying?", tested eight otherwise identical listings. When one carried a 1% discount, GPT-4.1 missed the cheapest item 9.3% of the time. Errors dropped sharply when discounts were bigger.

The study also found that agents can pile their choices onto a small number of products. In a quarter of the tested category-and-model combinations, a one-time change to a product description produced a statistically significant gain in how often an item was picked.

Prices may get more fluid

Sellers are building AI tools too. Capgemini forecasts that grocers will develop tools that bid for shoppers' baskets in real time, adjusting prices and offers. In business buying, Walmart uses negotiation agents from a company called Pactum for its many smaller supplier contracts.

Experimental research has found that pricing algorithms working independently can learn to keep prices above competitive levels without communicating. If prices become live negotiations instead of posted numbers, it gets harder for regulators or shoppers to tell tough competition from quiet coordination.

What to do

• Shoppers: compare prices yourself before trusting an AI pick, especially on small differences.
• Shoppers: check shipping costs and stock before you buy.
• Sellers: keep prices, shipping and stock data accurate.
• Sellers: write listings that work for people and for AI shoppers.

What happens next

Data cited by Deloitte says nearly 40% of business buyers already use agentic AI, meaning AI that takes actions for them, somewhere in purchasing. Only about a quarter of suppliers have built agents to meet them. The tools are being built fast, but they are not yet reliable.