What building for agentic commerce actually taught us
Agentic commerce—where AI agents shop on behalf of buyers—promises a future that feels inevitable. Making it work in production with live catalogs, inventory systems, fraud detection, and support queues is another story. Over the past six months, we've launched the Agentic Commerce Protocol (ACP) with OpenAI, introduced the Agentic Commerce Suite, onboarded retailers including Etsy and URBN, and observed how early adopters handle the messy details of agent-driven transactions.
Your product catalog is the first bottleneck
Before any agent can sell for you, it needs to ingest your product data—and every AI agent wants it formatted differently. One requires an SFTP drop, another a custom API, and a third its own feed specification. We've watched brands reformat the same catalog six ways to get listed on multiple agents, which quickly becomes a maintenance burden.
The fix is catalog syndication. Upload your product catalog to Stripe once, and the Agentic Commerce Suite reformats and distributes it across supported agent surfaces, preventing the fragmentation that comes from building custom integrations per agent. But ingestion formats are only the start; the data itself needs to be structured for how agents actually query it.
Real-time inventory is non-negotiable
Always-on agents raise the bar for responsiveness. When a shopper asks an agent about a specific product, the agent must verify availability at that moment—not from data that's 15 minutes stale. One platform asked us whether inventory checks happen down to the millisecond. The pressure is real.
Variants make this exponentially harder. A shirt shoppable by size, color, and custom embroidery, or a sneaker with 14 colorways each having its own size run, requires real-time checks to confirm a specific combination is in stock—or to trigger alternative suggestions. We stress-tested ACP against this complexity with partners like OpenAI. By hosting your endpoint via the Agentic Commerce Suite, you can expose live availability data directly in the checkout API call.
Protocol churn is certain—design for it
Since co-developing ACP in September 2025, we've shipped four releases adding payment handlers, scoped tokens, extensions (starting with discounts), built-in buyer auth, and native MCP transport. That pace means sellers can't rebuild their stack on every release. The Agentic Commerce Suite is intentionally protocol-agnostic, working across standards including Google's UCP, so you aren't betting your roadmap on any single spec.
The businesses we talk to are wary of zombie integrations—something built for a specific AI agent that becomes obsolete after a pivot. Unless you're prepared to staff a team just to track protocol changes, you need infrastructure that absorbs that volatility. Integrate once, and stay compatible as protocols evolve.
Payments are necessary, but not sufficient
Agentic transactions rely on a token layer. Shared Payment Tokens (SPTs) let agents initiate payments with the buyer's permission and preferred method—without exposing credentials. For large enterprises, this scoped, secure token infrastructure is where Stripe adds particular value.
But the full flow spans catalog discovery, checkout state management, shipping, and post-sale concerns like returns and refunds. That's why we play a dual role: an open source contributor alongside OpenAI on the protocol, and a commercial provider of fraud tools, onboarding, and catalog management on top—so businesses can support agent-driven commerce end to end.
Fraud signals shift, but network context fills the gap
A frequent question from sellers: is agentic commerce attracting more fraud? Since launching the Agentic Commerce Suite with retailers like Coach, Kate Spade, and Ashley Furniture, fraud rates have been near zero.
Traditional detection relies on human-typed signals—browser fingerprinting, mouse movements, battery level, window size. Those disappear when there's no person at the frontend. The Stripe network itself provides the alternative. Even if an agentic purchase is new to a specific business, the underlying customer and payment method usually have history in our network, providing immediate risk context.
By using SPTs, Stripe Radar applies its detection to agent-initiated transactions even when authorization happens off-Stripe.
Phase in SKUs and scope
Don't switch your entire catalog on at once. A proven approach: pick a small set of products likely to convert and watch how the channel behaves before expanding. Start with items that ship directly to the home—nothing requiring installation or complex fulfillment—since the frontend user experience is still maturing.
URBN, parent of Anthropologie, Free People, and Urban Outfitters, launched agentic commerce with a focused subset of high-demand items (dresses and denim) rather than its full range of plants and custom furniture. Early phases are about gathering data and understanding which SKUs, payment methods, and fulfillment options matter most. Starting small keeps you positioned to adopt more complex buying experiences—multi-item, multi-business carts included—as they come online.
Agents shift commerce fundamentals, not eliminate them
Physical retail happened in stores; first-wave ecommerce brought it to your site; mobile kept the brand experience intact. Agentic commerce places buying intent on AI surfaces. But discovery requires more than recognition—it's about being the option the agent picks. That work increasingly resembles SEO and performance marketing: being easy to find and easy to choose.
The messy parts of commerce also reappear in new places. If an agent confirms an order but a legacy backend rejects it after a fraud check, how does the customer get notified? If a buyer returns to an AI assistant saying "cancel my order," can the agent route that request to the seller reliably? These are open problems we're tackling alongside sellers and AI partners. Rationalizing agentic commerce as just a new channel misses that the points of friction—and control—have moved.
Identity is the new checkout friction
The logged-in state remains the holy grail for sellers: recognizing customers across sessions, personalizing experiences, applying loyalty benefits. Today, most agentic interactions behave like guest checkout—the agent protects the shopper's identity until the moment they authorize a purchase. Identity signals exist, but honoring loyalty rewards or diagnosing abandoned carts requires manual effort to map them into existing management systems.
The stakes rise as agent recommendations improve: when the decision to buy happens fast, any added friction—extra confirmation screens, form-fills, click-outs—can kill intent. This is why we continue building Link. Returning customers already have shipping and payment details stored, making checkout faster and letting agents complete purchases without exposing personal data. Over time, expect loyalty programs, richer fulfillment, and post-purchase engagement to integrate into the ecosystem.
First-party and third-party agents do different jobs
Sellers often ask whether to build a branded agent assistant or rely on external AI platforms. The two options serve different parts of the journey, so it's less either-or than a measurement question.
First-party agents—NikeAI, Home Depot's Magic Apron, Ralph Lauren's Ask Ralph—focus on engagement. They deepen relationships with known customers, protect brand control, and maintain context like identity and preferences. Third-party surfaces are about acquisition: meeting shoppers where they already are and capturing net-new demand, as Etsy is seeing early on. The strategic play is to design for both—using first-party experiences to drive retention and lifetime value, and third-party channels as a distribution surface that eventually funnels customers back to your owned properties.
Agents paying other agents
Not all agentic commerce flows through a familiar checkout. A growing use case is one agent paying another service directly, mid-task, without any human in the loop. There is no cart, no shipping form, no loyalty step — just a fast, programmatic payment inside an HTTP request.
The payment behavior is different from human purchases. An agent might make thousands of small decisions daily and needs low-latency, HTTP-native rails for pay-per-call or pay-per-task models. Builders say they want to bill agents for tool usage, data access, or automated workflows, but existing tooling assumes a human checkout.
Stripe's early answer is machine payments using stablecoins. A few lines of code through the PaymentIntents API let you charge agents for API usage, MCP calls, or HTTP requests. You specify the amount and currency, and Stripe generates a unique deposit address for each transaction.
The address is handed back to the agent so it can pay programmatically. In an x402 flow, for instance, the protocol passes the address to the agent to indicate where to send funds. Status is tracked via API, webhooks, or the Stripe Dashboard, and settlements land in your Stripe balance. Support currently covers x402 with USDC on Base, with additional protocols planned.
Early examples already exist: charging agents per API call for inventory, pricing, delivery quotes, or pick-up-slot holds, and per task for jobs like fitment checks, bundle building, quote generation, and replenishment. This is rare in conventional retail and is stablecoin-based for now, but it demonstrates where agent-native monetization is headed as the rails mature.
The roadmap
The near-term goal is parity with human checkout: agents, humans, and businesses transacting with the same reliability, but with richer context and stronger controls. For businesses already using Stripe for payments, much of the groundwork is in place.
The Agentic Commerce Suite is evolving along several fronts: more real-time updates, expanding SPT support to more payment methods, stronger fraud signals for new attack vectors, and identity resolution logic so sellers can recognize customers across agentic surfaces. As usage goes global, coverage is widening geographically and into new verticals.
