Where an AI Agent Creates Value in E-Commerce, and Where a Rule Still Wins

“Agentic e-commerce” gets sold as one idea: an AI agent that shops, decides, and pays on a customer’s behalf, with no human in the loop. In 2025 and 2026, the two companies best positioned to build exactly that, OpenAI and Shopify, actually tried it, and the part that removed the human from the purchase failed inside six months. What survived is narrower and more useful: agents that find and match products reliably, running on top of a payment layer that was rebuilt to bound an agent’s authority, not expand it. That split, what an agent is actually good at versus what still needs a human or a fixed rule, is the shape agentic e-commerce has today.

Direct Answer

Agentic e-commerce is not one capability but two: an agent that discovers and matches products, and infrastructure that lets an agent authorize a payment. OpenAI’s 2025 attempt to remove the human from checkout was killed within six months. The discovery layer, running on Shopify’s Agentic Storefronts, kept scaling in the same period. The two layers are earning trust at very different speeds.

Key Takeaways

• Agentic e-commerce splits into two layers that mature at different speeds: product discovery and matching, and payment execution.
• OpenAI’s Instant Checkout removed the human from the final purchase step in September 2025 and was killed by March 2026, only around 30 Shopify merchants ever went live on it.
• Google’s AP2 protocol answers the same trust problem by making an agent’s payment authority narrower and provable, signed Intent and Cart Mandates, not by making agents more autonomous.
• Discovery works because matching a customer’s request to a catalog is genuinely variable. Payment execution is closer to deterministic, which is exactly why a bounded rule beat an autonomous agent there.
• The useful question for a store isn’t “should we add an AI agent.” It’s which of these two layers the decision actually belongs to.

What Does “Agentic E-Commerce” Actually Mean?

Agentic e-commerce describes software that acts inside a purchase rather than only describing it. Instead of returning a list of products for a person to click through, an agent can search a live catalog, compare options against a stated need, and, in the more ambitious versions, complete the purchase itself.

The infrastructure behind it split into two separate efforts in 2025. Google announced the Agent Payments Protocol (AP2) on September 16 with more than 60 partners (Mastercard, PayPal, Coinbase, Salesforce, and Etsy among them) to standardize how an agent proves it has the authority to spend someone’s money. Two weeks later, OpenAI launched Instant Checkout, built with Stripe on their own Agentic Commerce Protocol, so ChatGPT could complete a purchase directly inside a conversation, starting with U.S. Etsy sellers on September 29, with over a million Shopify merchants, including Glossier, Skims, Spanx, and Vuori, queued to follow.

Both protocols answer the same underlying question, can a merchant trust that an agent’s purchase request is real, from opposite ends. One built the checkout experience first. One built the trust layer first. What happened to each over the following six months is the actual evidence for where the value sits.

What It Is Not: What OpenAI’s Own Retreat Proved

By March 2026, OpenAI killed Instant Checkout, five months after launch. Only around 30 Shopify merchants had ever gone live on it, and the real-time inventory synchronization the feature depended on was never fully built. OpenAI’s own explanation was that it is “evolving how we approach commerce in ChatGPT… prioritizing making ChatGPT search and product discovery great, with ACP serving as the infrastructure that connects users to merchants across the full shopping journey”: in plain terms, back to being a research tool, not a storefront.

The stated reason wasn’t a smarter model. It was two separate problems: shoppers didn’t actually want to pay inside a chat window (they researched there, then completed the purchase somewhere they already trusted) and the technical work of keeping prices and inventory correct in real time across merchants proved harder than the demo suggested. Why Adding an AI Agent to a Broken Workflow Usually Fails applies almost exactly here: the data and behavior underneath were never redesigned for an agent to own the last step, so removing the human from it surfaced every gap at once.

Forrester’s own read adds the harder number underneath the retreat: just 23% of Gen X, 32% of Millennial, and 35% of Gen Z US online adults had used ChatGPT to search for products at all in the past month, per Forrester’s December 2025 Consumer Pulse Survey and completing a purchase inside that chat was already the least-adopted part of that smaller behavior.

Agentic e-commerce, in other words, is not yet “an agent that buys things for you, unsupervised.” Where it actually held up in production was narrower than the pitch.

Where the Agent Earns Its Complexity: Discovery, Not Checkout

While the checkout experiment retreated, the discovery side of agentic commerce kept expanding. Shopify activated Agentic Storefronts by default for eligible merchants on March 24, 2026 (opt-out, not opt-in) making products from 5.6 million stores discoverable inside ChatGPT, Microsoft Copilot, Google’s AI Mode, and the Gemini app. Tellingly, Shopify’s own description of how a sale actually completes still routes through the merchant: users find products through Shopify Catalog, then complete purchases via an in-app browser, and on desktop ChatGPT links out directly to the merchant’s own store. The agent finds and matches. The merchant’s own checkout still closes the sale.

That’s the actual boundary, not a guess at one:

What an Agent Actually Needs to Work Reliably

The lesson isn’t “agents are good at discovery and bad at payment” as some fixed law. It’s that an agent’s judgment is only as good as the data it’s reasoning over, and a purchase is the exact point where a stale number stops being an inconvenience and becomes a wrong charge. An agent that recommends the wrong product wastes a click. An agent that completes the wrong purchase creates a support ticket, a chargeback, and a dispute over who actually authorized what.

That’s why AP2 didn’t try to make agents more trusted. It made their authority narrower and provable. Google frames the protocol around three questions a merchant needs answered before it will accept an agent’s request: whether the user actually authorized this specific purchase, whether the request reflects the user’s real intent, and who is accountable if the transaction turns out to be fraudulent or wrong. Picture the actual sequence: a shopper tells an agent “reorder my usual filter coffee when it drops under €12.” That instruction becomes a signed Intent Mandate: a record of exactly what was authorized, and under what condition. Weeks later, when the price crosses the threshold, the agent doesn’t just charge the card. It generates a Cart Mandate (a secure, unchangeable record of the exact item and price) and only that document, matched against the original Intent Mandate, triggers the payment. A merchant can verify the transaction matches what the person actually authorized, not what the agent decided to buy on its own. That isn’t autonomy. It’s a rule with cryptographic enforcement, wearing an agent’s front end.

An agent working at the discovery layer needs the same discipline in a smaller form: a catalog feed that’s actually current, a way to say “I’m not sure, here are three options” instead of guessing, and a boundary past which it hands the decision back rather than completing it. AI Agent vs Chatbot: The Difference Is Write Access, Not Intelligence is worth reading here: the moment an agent gets write access to inventory or a payment method, the standard it has to meet changes completely.

When a Fixed Rule Still Wins

Most of what a store needs “automated” isn’t a judgment call at all. A reorder point, a shipping-cost calculation, a rule that swaps a sold-out size for the next one up: these have a fixed, enumerable set of conditions, and a rules engine handles them for a fraction of the cost, with a complete audit trail and no risk of a model inferring the wrong thing on a busy day. What Is AI Automation? draws this line in general; agentic commerce is where it gets tested with real money on the table. An agent earns its complexity specifically where the condition can’t be enumerated in advance, an open-ended shopper request, and costs more than it’s worth everywhere else. AI Automation Cost is worth reading before assuming the more sophisticated option is automatically the cheaper one.

What Should Stay Human

Even inside AP2’s bounded model, someone still has to set the boundary: the spend cap on a delegated Intent Mandate, the decision to dispute a transaction, the call on a refund that falls outside policy. What Should Stay Human When You Embed AI Into Operations? is the general version of this question; the e-commerce answer is specific: a human sets the limits an agent operates inside, and reviews what happens when a transaction falls outside them. The agent doesn’t get to decide its own ceiling.

Why This Split Is Likely to Hold

It’s tempting to read the Instant Checkout retreat as a temporary limitation: better models, cleaner data pipelines, and the autonomous version comes back. That misses what actually changed between the two layers. A bad product recommendation costs a click: the shopper closes the chat and looks elsewhere, and nothing irreversible happened. A wrong purchase costs money, creates a chargeback, and raises a real question of who is liable for it. That asymmetry doesn’t shrink as the underlying model gets better at language. It’s a property of what a purchase is, not of how good the agent reasoning it out happens to be. That’s the actual argument for keeping payment execution bounded by a rule rather than an open-ended decision, independent of next year’s model release.

Summary

“Agentic e-commerce” will keep being marketed as one product, because “an agent that shops for you” is a simpler pitch than “an agent that matches products, running on top of a payment rail built specifically to limit what it can do.” The two companies that actually shipped this in 2025 handed back a more useful answer than the pitch: build the agent where the request is genuinely open-ended, and keep the money-moving step bounded by a rule, whether that rule is a human clicking confirm or a cryptographically capped mandate. The businesses skipping that distinction are the ones re-learning it in production, with a customer’s card on the line.