Visa asked 22 AI agents to buy a television, a laptop, and a hair appliance across 15 attempts for each product. The overall success rate was only around 57%, with some agents purchasing unsuitable products or failing to suggest a suitable option. Agents powered by the most advanced models, including Google Gemini Aragon 4 and Meta Making Wonder, completed 82% of purchases successfully, compared with 76% for a Claude agent and 64% for an OpenAI agent. The five most expensive models achieved an average success rate of 67%, while the five cheapest reached 50%. Visa says commerce driven by AI agents still requires advances in trust, secure payments, and clarification of legal responsibility for fraudulent or unauthorized purchases.
Source: Read the original article
