Day 120: The Window Before Agentic Commerce Scales

Stripe Sessions is a payments conference. Or it was. This year, roughly half the programming had moved somewhere else entirely — into the infrastructure of a world where AI agents discover, evaluate, and purchase on behalf of the people who used to do those things themselves.

As a fund investing at the intersection of AI and digital commerce, this was the shift we came to understand more concretely. The numbers being discussed were striking: by 2030, US e-commerce could see up to $1 trillion in agent-triggered transactions, spanning everything from a consumer delegating a birthday gift purchase to an enterprise agent managing quarterly procurement. Whether that figure proves accurate or not, the direction of travel felt hard to dispute. The infrastructure required to support that kind of volume doesn't yet exist. And the conversations at Stripe Sessions were, in large part, about what needs to be built, and in what order.

Across the sessions, four layers of the agentic commerce stack kept surfacing. Here's what we heard in each.

On intent and decision-making

Allison Xu from Stripe's agentic commerce team opened her session with a personal story: she used an LLM to plan an upcoming trip, got a full itinerary, a packing list, and ended up buying a dress — all in a single conversation. Her point was that for consumers, agentic commerce is already happening in everyday life. She described the evolution as a five-level spectrum: at level one, you're just telling an agent what you want instead of filling out a form; at level five, the dress arrives three weeks before the trip because the agent knew you'd need it. Most consumer interactions today, she noted, are still at levels one and two. B2B use cases — procurement, software purchasing, API access — are moving faster, already approaching levels where agents execute against a budget and policy with minimal human involvement at each step.

Sam Altman, in conversation with Patrick Collison, described the broader shift in terms that felt useful: moving from "AI that helps you with one task" to "AI that handles all the work you do at a computer." He estimated OpenAI is roughly twenty percent of the way there. The next unlock, he suggested, won't announce itself as a new product category. It'll feel more like friction quietly disappearing from things people had accepted as unavoidable.

On discovery and reach

Jonathan Arena from New Generation AI ran a live demo that made the gap in this layer immediately visible. He searched a standard e-commerce site with a natural language query and walked the audience through what broke. The results were mismatched, buried, or inaccessible — not because the site was poorly built for humans, but because it was built exclusively for humans. His point: "Discovery wins before checkout begins." An agent that can't find or understand a product can't buy it, no matter how good the checkout experience is.

The example he used to illustrate the scale of the problem: a global activewear brand with roughly 160,000 live products, of which around 10,000 were discoverable by AI agents. The other 93% were, from an agent's perspective, invisible.

Three brands described where they are with this in practice, and their approaches were notably different:

Ashley Furniture made an early call to participate in agentic commerce even though a $6,000 sectional is unlikely to close through an LLM anytime soon. Their reasoning, as Kyle from their team explained it: the learning is worth more than the near-term revenue. They'd already sold a meaningful volume of items through agent channels by Black Friday, and they're treating each transaction as data about a channel still being defined. The posture was less "this is working now" and more "we're not going to be the last ones to figure this out."

Tapestry, the company behind Coach and Kate Spade, framed the challenge as one of brand control. How do you tell your brand story when the surface your customer encounters is a conversational interface you don't control? Their answer so far: push your brand data and narrative into conversational discovery channels proactively, and test what works. They've launched a "gift concierge" on katespade.com as an early experiment in shifting from keyword-based to intent-based discovery. Mandeep from their team made a point that stayed with us: 900 million weekly active users on ChatGPT and similar tools are already using these interfaces to research products. If your brand data isn't there, your brand isn't in that conversation.

Woo, a platform serving thousands of SMB merchants, described the problem from the infrastructure side. Their merchants aren't asking philosophical questions about agentic commerce, they're asking where their organic traffic went, and how to get it back. Woo's role, as they described it, is to productize the complexity: make something that would take months of integration available as a single toggle. The two blockers they flagged — real-time data freshness and a new category of fraud that existing systems weren't designed to catch — felt like the most grounded and specific observations in the session.

On transaction, payment, and protocol

This layer was the most technically dense part of the conference, and also the layer where the incumbent advantage felt most pronounced. The protocol decisions being made right now — how agents communicate with merchants, how payment credentials are scoped and transferred, how machine-to-machine transactions get settled — are being driven primarily by Stripe, Visa, Mastercard, and the major hyperscalers. That's worth naming clearly, because it shapes how to think about where new companies fit.

The panel featuring Walmart, Bank of America, and Visa was candid about the pace: things are moving slower than expected. Legacy systems are harder to retrofit than anticipated, and the trust and liability questions haven't been resolved at the ecosystem level. One data point that came up: early in-agent checkout is converting at roughly three times lower than referral traffic to merchant sites. The transition appears to be more gradual and hybrid than the more optimistic projections would suggest — which, from our perspective, means the near-term opportunity is less about replacing existing checkout infrastructure and more about what sits around and between it.

On trust, fulfillment, and risk

This is the layer the ecosystem is least prepared for, and the one that determines whether everything else can function at scale.

The same Visa/Walmart/BofA panel put the problem clearly: payment fraud systems have spent 25 years operating on one assumption: machine-initiated transactions are suspicious. Agentic commerce requires flipping that assumption entirely. Now the infrastructure needs to distinguish legitimate shopping agents from malicious bots, verify that a transaction actually reflects what the customer intended, and establish who is liable when something goes wrong.

The liability question is the one that stayed with us. If an agent buys the wrong item, picks the wrong shipping option, or misinterprets a preference — who is responsible? At what point does that shift from the agent platform to the merchant to the payment network? These aren't hypothetical edge cases. They're the framework that every participant in agentic commerce needs to agree on before fully autonomous transactions can happen at scale. The panel's honest answer was that standards are still forming, and the market may align on principles before it aligns on specifications.

What Stayed With Us

Based on what we heard across the sessions, not every layer of the agentic commerce stack is equally open to new companies. Some of what needs to be built will be defined by platforms that already control the infrastructure. But several areas still look like genuine startup territory, and here is where we think early-stage founders may have room to build.

The catalog and discovery layer looks like the most accessible entry point. The gap between what merchants have for sale and what agents can actually find is large, measurable, and unresolved. Fixing it requires work at the data and commerce operations layer — semantic cataloging, real-time inventory signals, agent-readable product information. This is a layer where deep knowledge of how merchants structure and maintain product data matters more than payment network scale. Incumbents don't have the same structural advantage here that they have further down the stack, which is part of why we think it remains open.

Agent identity and trust is a coordination problem as much as a product problem. Verifying which agents are legitimate, what they've been authorized to do, and who bears liability when mistakes happen requires coordination across agent platforms, merchants, payment networks, and issuers. No single incumbent is well-positioned to drive that unilaterally. We think this creates space for early companies that can help build the standards and tooling across parties that don't naturally work together — not by owning the whole problem, but by making the coordination easier.

The B2B intent layer is moving faster than the consumer narrative suggests, and the tooling to support it is largely unbuilt. Procurement, API access, software purchasing, media buying — these are early candidates for agent-delegated execution. A corporate agent operating with budget authority, vendor policies, and approval workflows needs infrastructure that enterprise software incumbents aren't set up to move quickly on. Startups with deep workflow expertise in specific verticals may have a window here before that changes.

The handoff between discovery and conversion is the area that received the least attention at the conference, and may be worth the most attention from founders right now. Early data suggests in-agent checkout converts poorly relative to referral traffic to merchant sites. If that holds, the near-term opportunity may not be replacing checkout but improving the handoff — making it faster, cleaner, and more attributable. Neither agent platforms nor merchants have strong incentive to solve this on their own. That's often where early companies find room to work.

Patrick Collison framed this moment as "Day 120 of the Singularity", the early phase of an acceleration that is, for now, still running at human speed. Whether or not you take the framing literally, the window to help define how agentic commerce works feels genuinely open. The signals we saw at Stripe Sessions were stronger than they were when we first started paying attention to this space. The infrastructure is being assembled in real time, and many of the choices being made now will be harder to undo later.