Checkout Is Not a Feature

Ron Reynolds · 2026-03-13 · 8 min read

OpenAI just killed Instant Checkout inside ChatGPT.

The feature that was supposed to let 700 million weekly users buy products without leaving the chat — gone. After partnerships with Walmart, Target, Shopify, Instacart, and Etsy. After co-developing the Agentic Commerce Protocol with Stripe. After months of "agentic commerce" headlines.

Of the millions of merchants on Shopify, roughly a dozen actually integrated. Not a dozen thousand. A dozen. People browsed. They didn't buy. And as of February, OpenAI still hadn't built a system to collect and remit state sales taxes.

The most capable AI company in the world tried to build commerce and couldn't close a transaction.

This isn't a failure of intelligence. It's a failure of architecture. The Checkout Illusion

The tech press treated "buy it in ChatGPT" as a UI problem. Put a checkout button in the chat. Connect it to Stripe. Done.

But checkout isn't a button. Checkout is the last step of a chain that starts with inventory availability, runs through tax calculation, shipping options, fraud detection, payment processing, order creation, fulfillment routing, and confirmation — then continues into returns, refunds, customer service, and accounting reconciliation.

Every link in that chain needs live data. Not a product feed that syncs every few hours. Live. Is this item in stock right now, in the warehouse closest to the customer's shipping address? What's the sales tax in their jurisdiction — not just the state rate, but the county, city, and special district overlays? Is this payment method valid? Is this transaction pattern consistent with the customer's history or does it look like fraud?

OpenAI had none of this. They had a product feed and a payment button. The distance between those two things and an actual commerce operation is the distance between a brochure and a business. Why Alibaba Works and OpenAI Doesn't

There's one company doing conversational commerce at scale right now. Alibaba's Qwen app completes purchases inside the conversation — in China, with hundreds of millions of users.

Why does it work? Because Alibaba owns the AI model, the marketplace, the payment rails, and the logistics network. It's one system. When a customer says "order this," the model doesn't need to call out to a dozen disconnected services. Inventory is right there. Payments are right there. Fulfillment is right there. Tax is handled. Fraud is handled. Returns are handled.

Alibaba didn't bolt commerce onto a chatbot. They had the commerce system first and gave it a conversational interface.

OpenAI tried the reverse — take a conversational interface and bolt commerce onto it. Every retailer was a separate integration. Every product feed was a separate sync. Every checkout was a bridge between systems that don't share context, don't share data, and don't share state.

The architecture dictated the outcome before a single line of checkout code was written. The -Ilities

Early in my career, I took a workshop on what software engineers call "the -ilities" — the non-functional requirements that determine whether a system works in production, not just in demos:

Reliability. Will this system process the 400th order of the day with the same accuracy as the first? When inventory changes mid-checkout, does it catch the conflict or sell something that's already gone? When Stripe returns an ambiguous response, does it retry correctly or create a duplicate charge?

Maintainability. When tax laws change — and they change constantly, across thousands of jurisdictions — how quickly can the system adapt? When a payment provider updates their API, how many integration points break? When a merchant changes their return policy, does that propagate to every agent that needs to know?

Extendability. When a new payment method appears, how much work to support it? When a new shipping carrier launches, how deeply can it integrate? When a new AI protocol emerges — and they're emerging quarterly — can the system speak it without a rewrite?

Adaptability. Can the same system serve a merchant selling handmade candles and one selling industrial equipment? Can it handle subscriptions and one-time purchases? Can it adapt to a merchant who wants full autonomous operation and one who wants to approve every decision?

These aren't features you add later. They're architectural decisions made on the first day that either compound into a resilient system or collapse under real-world complexity. OpenAI skipped them. They built the demo. The demo looked great. Then reality showed up.

Twelve merchants. No tax compliance. People browse, they don't buy.

That's what happens when you treat checkout as a feature instead of what it actually is — a consequence of everything underneath working correctly. The Silo Problem

Shopify's Harley Finkelstein explained why checkout was so hard: "That is subscriptions, that is the inventory, that is shipping, taxes, all the different merchandising options."

He's right. And he's describing why Shopify's own architecture makes this nearly impossible.

Shopify is an app store. Inventory is one app. Shipping is another. Tax is a plugin. Email marketing is Klaviyo. Returns are Loop. Reviews are Yotpo. The merchant's business is scattered across a dozen vendors, none of which share context.

When an AI agent tries to execute a purchase in that environment, it needs to coordinate across systems that weren't designed to coordinate. The inventory app doesn't know about the return policy. The tax plugin doesn't know about the subscription terms. The shipping integration doesn't know about the customer's loyalty tier.

The agent isn't dumb. It's blind.

Forrester's research confirms this: organizations without an integration platform had a zero percent success rate running AI workflows across five or more data sources. Not low. Zero.

You can't do agentic commerce on a siloed system. Not because the AI isn't smart enough. Because the architecture won't allow it. The agent can only reason across what it can see — and in a siloed system, it can't see across anything. What an Operating System Changes

There's another model. Not integration — unification.

When inventory, pricing, orders, fulfillment, customer data, support history, return policies, marketing, and analytics all live in one system, the agent doesn't need middleware to see the full picture. It already has it.

An agent processing a checkout already knows the inventory is live — because the inventory service is the source of truth, not a synced feed. It already knows the tax — because the tax engine is a service, not a plugin. It already knows the customer's history — because the customer service, the order service, and the support service share the same data layer.

When a return happens, the inventory updates, the customer record updates, the financial ledger updates, and the analytics reflect it — not after a nightly batch job reconciles three systems, but immediately, because it's one system.

This is what "operating system" means. Not a marketing term. An architectural reality. The same way your laptop's operating system lets every application share memory, files, network, and hardware without middleware between them — a commerce operating system lets every business function share data, context, and state without integration layers hoping they agree.

The -ilities are built in, not bolted on: Reliability from shared state — no sync conflicts, no stale data, no phantom inventory Maintainability from single responsibility — update the tax service once, every agent sees it Extendability from protocol-native design — new protocol support plugs in, merchants don't notice Adaptability from merchant-controlled autonomy — same OS, different configuration, different business The Market Just Told You What It Needs

The most well-funded AI company in history partnered with the largest commerce platform in the world and couldn't reliably sell a pair of shoes through a chatbot.

That's not a setback. That's a signal.

Agentic commerce isn't failing. The attempt to build it on top of architectures designed for humans clicking between tabs is failing. The Forrester analyst was right: the market is "still learning how consumers will actually adopt conversational commerce journeys." But the Alibaba example proves the concept works — when the architecture is right.

53% of US consumers have already made purchases based on AI recommendations. Generative AI traffic to retail sites grew 693% year-over-year during the 2025 holiday season. The demand is proven. The intent is there. What's missing is the infrastructure to fulfill it.

Not another integration layer. Not another middleware platform. Not another partnership between companies whose systems don't talk to each other.

An operating system. One system where agents can see everything, reason across everything, and act on everything — reliably, maintainably, extendably, adaptably.

The companies that build on architecture will serve that demand. The companies that bolt features onto silos will keep writing press releases about partnerships that produce twelve merchants and no tax compliance.

The -ilities don't care about your funding round. They care about your architecture.