As personal AI assistants like Muse and Claude become more integrated into daily routines, their use for shopping has increased. However, these agents face several challenges in executing shopping tasks effectively.
A pioneering study by Visa evaluated 22 different AI agents on their ability to purchase everyday items such as a TV, a laptop, and a hair appliance. Results showed that the overall success rate was approximately 57%. This study is documented on a new Visa website, which displays the outcomes of 15 distinct shopping tasks performed by each of the 22 agents.
Agents utilizing advanced AI models achieved better outcomes. The Aragon 4 version of Google’s Gemini and the Wonder version of Meta’s Making demonstrated the highest success, completing successful purchases 82% of the time. A Claude agent had a success rate of 76%, while an OpenAI agent ranked sixth with a 64% success rate.
The performance diagram from Visa indicates the success rates of all 22 models. The vertical axis represents the agents’ success rate, while the horizontal axis reflects the cost of each shopping task. Visa highlights that the five most expensive models had an average success rate of 67%, whereas the cheapest six achieved 50%.
One major issue is that the motivations for purchases are complex and consumers often lack fully detailed information. Visa’s study addressed this by providing a list of desired attributes for TV shoppers, such as compatibility with a child’s gaming console, a 30-day return policy, and Dolby Vision capability. Agents, like real consumers, did not always have complete information, necessitating interactions to extract necessary details.
To succeed in the experiment, AI agents needed to uncover missing information efficiently. Often, they failed by purchasing unsuitable products or not suggesting appropriate options. Visa’s Global Head of Growth, Rubail Birwadker, stated that although trust and technology are currently lacking, advancements are being made. He cited product discovery as just one component of this evolving commerce form and hinted that agent-driven shopping is likely to differ greatly from current online shopping methods.
Smooth and secure payment processes remain a hurdle for mainstream AI shopping. Visa and competitors like Mastercard are working on tools and education for merchants to facilitate agentic commerce. Additionally, unresolved legal issues persist, such as liability for rogue purchases or fraud committed by AI agents.
Birwadker suggested these issues would eventually be resolved, akin to the evolution of traditional ecommerce. He noted that agent-based shopping would present a distinct experience compared to contemporary online shopping methods.
Accompanying its experiment results, Visa launched a website that allows consumers to interact with the AI agents evaluated in their study.

The Daily Routine of Klook’s CEO Focused on Fitness and Efficiency
The Jensen Huang Controversy: Silicon Valley’s Divided Opinion
California Lawsuit Against OpenAI: Key Details and Implications
Updates on Key International and Domestic Events
E.l.f. Cosmetics Breaks Makeup World Record at TwitchCon
Understanding Credit Freeze, Fraud Alert, and Identity Monitoring