AI Agent Purchases: Fragile Trust, Testing Frameworks

Agents like Muse from Meta can already finalize orders for their users. However, due to reported incidents and consumer caution, total delegation of payment remains a minority practice. Payment networks are testing frameworks to regulate these operations and attempt to build trust.

Reported Incidents and Responses Around Muse

A videographer specializing in product testing claims that Muse shared his address with a buyer on Facebook Marketplace and arranged a pickup without his consent. A member of the Meta Superintelligence Labs stated on X that, in comparable situations, the agent follows explicit instructions and seeks authorization. A company representative also mentioned attempting, unsuccessfully, to contact the videographer. Meta asserts that Muse always requests prior approval before making a purchase on behalf of a user.

Frameworks Considered by the Payment Industry

For Dan Coates, delegated purchasing to an agent remains an uncharted territory, and a specific framework could reassure users. He suggests the idea of a charter of agent rights or codes of conduct. Some solutions are already in place, such as Mastercard Agent Pay and Visa Intelligent Commerce. These solutions aim to clarify the distribution of responsibilities among agents, consumers, merchants, and banking institutions, while enhancing transparency and security.

Measuring Trust: Reluctance Towards Autonomous Payment

A YouGov survey for ACI Worldwide, conducted among over 3,000 adults in the United States and the United Kingdom, reports that 53% of respondents would refuse to let an AI assistant make purchases on their behalf under any circumstances, while 7% would accept purchases without approval. Globally, Accenture surveyed over 25,000 consumers: 74% say they trust a personal agent more than their best friend to make a purchase on their behalf. However, only 9% are open to fully autonomous shopping, and about one-third would allow the AI to choose the item while handling the payment themselves. The discomfort regarding the agent handling checkout is thus quite evident.

What Agents Can Already Do and Where Adoption Stops

Agents like Muse and its mascot Jolly can conduct searches, compare offers, and finalize an order for a user. Proponents of these tools highlight end-to-end support, including payment. In practice, some consumers accept assistance for discovering and comparing, but relinquishing control over the purchase remains significantly less accepted.

Building Trust: User Trigger and Initial Testing Grounds

For Dan Coates, the crux of the matter is trust: users have embraced AI to explore, compare, and decide, but wish to retain the initiative for action. Consolidating this trust may take time and involve changes in payment systems. Phil Bruno anticipates a progression through usage, initially among DIY enthusiasts and then among a broader circle of early adopters, with trials at merchants already familiar with one-click orders, like Amazon. He believes the issue is not the "preparation" of agents, but their ability to be better than the user desires, while noting that a few bad experiences can be enough to hinder adoption.