10 Lessons Learned Building the First Generation of Agentic Commerce

Buying through AI agents may look natural on the surface, but in actual operation it has to connect live catalogs, real-time inventory checks, fraud prevention, and customer support. Over the past six months, Stripe has rolled out the Agentic Commerce Protocol (ACP) and Agentic Commerce Suite, integrating with merchants like Etsy and URBN, and it summarizes what it has learned about the early operational realities of agentic commerce.

The core issue isn't a simple payment API, but infrastructure that reliably ties together the entire transaction process. Through ACP, Stripe lets buyers, AI agents, and sellers transact via API, and emphasizes that uploading a catalog to Stripe allows deployment to supported agents, reducing duplicate integrations and format rework.

In practice, the biggest challenges are accurate product data and immediacy. Each agent demands SFTP, custom APIs, or dedicated feed specs, requiring the same catalog to be reworked in multiple ways, and inventory needs to be checked as of right now rather than 15 minutes ago. Variant products such as size, color, and embroidery are even trickier, making real-time validation at checkout critical to confirm whether a combination is actually purchasable.

The protocol itself keeps changing. After co-developing ACP with OpenAI, Stripe has added payment handlers, scoped tokens, extensions, buyer auth, and native MCP transport, and it believes merchants need a layer less vulnerable to standards changes rather than betting everything on one particular spec. That's why the Agentic Commerce Suite was designed as a standard-agnostic commerce layer that also works alongside other standards such as Google's UCP.

Payment is only part of the overall flow. With Shared Payment Tokens (SPTs), Stripe lets agents pay using the buyer's authorization and preferred payment method without exposing credentials, and it also handles post-purchase stages like checkout status, shipping, returns, and refunds. At the same time, it explains that it uses Radar and network signals to maintain fraud detection at a level comparable to existing standards even for agent purchases.

The recommended launch strategy is to start narrow rather than go broad from day one.

• Start with a small number of SKUs that have high conversion potential
• Exclude products requiring installation or complex fulfillment early on
• Expand in stages from popular product categories, as URBN did

Agentic commerce is also not simply a new distribution channel. Like search and performance marketing, discoverability needs to be redesigned from scratch, and as agents insert themselves into the middle, operational issues like order cancellations, refunds, and routing have to be solved anew at different points.

One of the biggest bottlenecks is identity resolution. Currently things mostly work like guest checkout, with the buyer's identity revealed only at the last moment, making it hard to handle loyalty, discounts, and conversion attribution at existing levels. To address this, Stripe is improving Link to leverage stored shipping and payment information, aiming to let agents complete payments faster without directly exposing personal information.

Finally, Stripe sees early signs of agents-to-agent payments as well — not human checkout, but agents making small, low-latency payments to other services within HTTP calls. To support this, Stripe has previewed stablecoin-based machine payments, currently supporting x402 with USDC on Base, opening up new models such as billing per API call, tool use, or unit of work.