Visa tested 22 AI agents on shopping for everyday items, and they succeeded only about 57% of the time overall. Even the top-performing agents managed 82%, a useful reality check for anyone expecting software to take the shopping list and run with it.
Watch Desk analysis
What happened
Fortune reports that Visa asked the agents to buy a TV, a laptop and a hair appliance across 15 shopping expeditions per agent. The task included working out a shopper’s needs from incomplete information, rather than simply finding a product that matched a neat specification.
The results varied by model. Google’s Gemini-powered Aragon 4 and Meta’s Wonder agents each succeeded 82% of the time; Claude scored 76%, while an OpenAI agent scored 64%. Visa says the five most expensive models averaged 67% success, compared with 50% for the five cheapest.
Why it matters
Shopping is full of small ambiguities that people resolve through questions: what does “good enough” mean, which feature is essential, and when should a cheaper option be ruled out? Visa’s test suggests agents still struggle to turn that uncertain back-and-forth into a suitable purchase. A wrong recommendation is inconvenient; an agent authorised to pay could make the consequences rather less theoretical.
The results also complicate the idea that stronger models or higher running costs automatically make delegated shopping dependable. Visa’s experiment is one test, not a verdict on every kind of shopping or agent, but its reported success rate gives readers a concrete measure of the gap between a promising demo and a trustworthy errand-runner.
Our read
The interesting benchmark here is not whether an agent can search a catalogue. It is whether it can notice what the shopper has not yet explained. At 57% overall, these agents look more like assistants who need supervision than shoppers ready to be handed the card and wished well.
What to watch
- Whether later tests use more shoppers, products and retailers.
- How success is defined when a product is acceptable but not the shopper’s best fit.
- Whether payment controls and clear responsibility arrive before shopping agents become routine.
Discussion spark: Would you let an AI agent buy everyday items for you at today’s reported success rates, or should it be limited to recommendations until it can ask better questions?
Sources and evidence
- Source update (5 November 2026, 12:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.