AI agents excel at product discovery, but they struggle to close the deal. While 132 million U.S. adults have used AI to assist in a retail purchase, according to PYMNTS Intelligence, the actual conversion rate for these autonomous systems remains stuck. Only 3% of transactions currently involve agents, revealing a massive chasm between finding an item and completing the purchase.
The bottleneck is the approval screen. It is the point where the industry’s vision of seamless agentic commerce collides with the reality of risk. When an agent presents a final purchase for confirmation, it is not just asking for a click; it is asking the user to trust a black box with their money and data. If the interface is too complex, the user abandons the cart. If it is too simple, the user feels exposed.
OpenAI learned this the hard way. In March 2026, the company retreated from its Instant Checkout feature. Data from Walmart, as reported by WIRED, showed that products purchased directly inside ChatGPT converted at one-third the rate of those requiring a click-out to the retailer’s own site. Even with the convenience of an in-chat experience, users were clearly more comfortable completing the transaction on a familiar, merchant-controlled domain.
The industry is currently testing three distinct approaches to this approval friction. Meta’s Muse, launched in September 2026, relies on an insurance-backed model: a one-time-use virtual card powered by Stripe Link and a refund guarantee underwritten by Cover Genius. At the other end of the spectrum, GrokBot adopts a “buyer beware” stance, placing full liability on the user with an as-is disclaimer. Between them, Google and Mastercard are pushing a technical standard: the Agent Payments Protocol (AP2) and Verifiable Intent, which use cryptographically signed mandates to prove exactly what a user authorized.
Trust remains the binding constraint. Only 14% of consumers trust AI to execute purchases without manual verification, and 79% express concern over data privacy, according to Visa’s Earning Consumer Trust report. This skepticism is rational. According to the Quantum Metric 2026 AI Experience Benchmark, AI-referred customers are twice as likely to abandon a purchase after a single friction event. Once a user has a bad experience, 81% will not return.
Merchants are equally wary, which creates a feedback loop of added friction. Because 93% of merchants believe the AI provider should bear the financial loss for incorrect purchases, they are hesitant to open their inventory, per PYMNTS Intelligence. Currently, only 28% of merchants are willing to offer agents their full product range on the same terms as other channels. They are effectively gatekeeping their catalogs until the liability framework matures.
The Google-Mastercard approach attempts to standardize this through Verifiable Digital Credentials. By creating a tamper-resistant record of the user’s intent – what they authorized, for how much, and for which items – they aim to provide the portable evidence needed to satisfy both merchants and consumers. It is a move toward infrastructure rather than interface, replacing blind trust with cryptographic certainty.
The industry faces a difficult balancing act: excessive friction at the approval screen kills conversion, yet removing it entirely destroys user trust. The market for agentic commerce is projected to grow from $8 billion in 2026 to $1.5 trillion by 2030. Future growth hinges on whether developers can design an approval process that functions as a genuine safeguard rather than a deterrent.