Agentic Commerce in 2026: What’s Real & What’s Next

Most of the answers being proposed right now are built on a hazy sense of the market rather than a clear picture of what’s actually deployed, what’s still vapourware, and where the real risk sits. Before deciding what to build, it’s worth being precise about what agentic commerce is today, how big the opportunity genuinely is, and what’s changed in the last year.

Put simply, agentic commerce shifts part of the buying decision away from the shopper and onto software: an AI agent does the searching, the comparing, and (with permission) the paying, cutting out the usual back-and-forth of browsing and clicking through checkout.

McKinsey puts the global opportunity at $3-5 trillion by 2030. It’s also worth noting that OpenAI already had to scale back its own flagship agentic checkout feature this year. Both things are true at once, and this piece sets out why.

What agentic commerce actually is

Autonomy sits on a spectrum, and most of what’s live today falls well short of the far end. Picture someone asking their agent to book a hotel room in Manchester for under £120 a night with free cancellation. The agent checks availability across a handful of booking sites, applies the price and cancellation filters, and either surfaces two or three options for a final decision or, where it’s been given standing authority to book, confirms the reservation outright. No browser tabs and no comparison spreadsheet, because the agent did the legwork that a person would otherwise do by hand.

End-to-end autonomy, where the agent buys without asking, is currently limited to low-stakes, repeat purchases like grocery reordering. That distinction matters because “agentic commerce” gets stretched to cover everything from AI product recommendations to fully autonomous buying. Whenever you see a market-size figure, check which one is being measured.

The opportunity, without the sloppy headline number

Ask five analyst firms how big agentic commerce will be by 2030 and you’ll get five different answers, because they’re not all measuring the same thing.

eMarketer’s narrow measure, direct sales completed on AI platforms, lands around $144 billion. Morgan Stanley puts US agentic e-commerce at $190–385 billion. Bain says $300–500 billion. McKinsey’s $3–5 trillion counts everything an agent orchestrates or influences, globally. That’s roughly a 35x spread on the same term, and the gap is entirely about scope, not disagreement over where the market is heading. 

The protocols, briefly

None of this requires tearing up an existing backend. Most merchants already run payments through infrastructure, Stripe being the obvious example, that agentic protocols like ACP are built to sit on top of, so turning on agent-ready checkout can be a configuration change rather than a rebuild.

Each protocol solves a narrow, distinct problem. AP2 handles authorisation, a spending rule an agent can’t exceed without asking again. ACP handles the transaction itself: order details, payment token, and delivery info passed between agent and merchant without exposing raw card data to the model. MCP is what lets an agent pull live inventory or pricing from a merchant’s system in the first place.

The newest entrant is Google’s Universal Commerce Protocol (UCP), launched in January 2026 with Walmart, Target, Shopify, Etsy and Wayfair on board and endorsements from most of the major payment networks. It’s designed to interoperate with AP2 and MCP rather than replace them, but a coalition that size is exactly why betting your architecture on any single protocol this early would be premature.

It’s worth being precise here, because most of what’s live today sits at the simpler end of this spectrum. When an AI platform “buys” something from a retailer, it’s typically calling a structured checkout API — a well-built shopping cart interface, not agent-to-agent negotiation. 

Genuine agent-to-agent commerce, where a buyer’s agent negotiates directly with a seller’s system, is a different proposition, and one we’ve previously explored. Which model ends up dominating a given category is still an open question, and depends on which protocol merchants and platforms adopt at scale.

What’s actually live in 2026

Agentic checkout has already had a public stumble. In September 2025, OpenAI launched Instant Checkout, letting ChatGPT users buy directly from Etsy sellers, with over a million Shopify merchants reportedly lined up to follow. By March 2026, roughly 30 Shopify merchants had actually gone live, per Forrester analyst Emily Pfeiffer. OpenAI pulled back from native in-chat checkout, shifting to an app-based model that redirects purchases to the merchant’s own site.

This example really highlights the gap between announcement and infrastructure. Payments, fraud checks, tax, and returns are hard problems that don’t disappear because an AI model is involved. Not every retailer wants agents involved at all: Amazon has blocked ChatGPT from accessing its site and sued Perplexity over its browser-based scraping tool.

That fight isn’t going Amazon’s way; in August 2026 an appeals court overturned the injunction, ruling that when a user directs an agent to shop on their behalf, it’s the user doing the accessing. The case continues, but the early signal is that retailers may not get to unilaterally decide whether buyer agents show up. Which makes being ready for them less optional than it looks.

What is growing fast is AI-driven research traffic. Adobe Analytics recorded a 693% year-over-year jump in traffic to US retail sites from generative AI sources over the 2025 holiday season. Grocery and CPG replenishment (low-risk, repetitive decisions) remains the strongest real-world case for actual autonomous purchasing today.

The practical takeaway: build for agent discoverability now. Don’t architect your checkout around any single vendor’s protocol winning this year.

What this means if you’re building a commerce app

Whatever happens with autonomous checkout on your platform, one thing is already worth doing: making your product data legible to agents, the same way you’d optimise it for search engines.

That means structured, machine-readable product feeds; price, availability, variants, delivery windows, returns. Use schema.org Product markup and a real-time API an agent can query directly, rather than scraping your storefront. If your product data only exists as rendered HTML inside CMS templates, an agent has to guess what it’s looking at.

This is also where backend architecture starts to matter beyond user experience. An API-first, headless commerce backend is naturally closer to agent-ready than one where product data, pricing, and inventory logic are tangled into a monolithic frontend. This requires not locking yourself into a backend that will need tearing apart the moment agent traffic becomes material to your sales.

Trust is the real constraint

The biggest obstacle to agentic commerce is trust. Handing an agent payment details and a delivery address takes a level of confidence most people don’t have yet, beyond low-value, repeat purchases.

There’s a builder-side risk too. Gartner predicts more than 40% of agentic AI projects will be cancelled by the end of 2027, citing unclear business value and inadequate risk controls — and flags “agent washing,” where ordinary chatbots get rebranded as agentic with no real autonomy behind them. Worth remembering the next time a vendor pitches an agentic commerce solution.

None of this is a reason to wait and do nothing, but it is a reason to treat agentic commerce as an infrastructure investment with a realistic timeline.

Where to start

Three questions worth answering before investing further:

First: can an agent actually read your pricing, stock, and delivery data today, or is it locked inside frontend templates built for human eyes? 

Second: is your growth better served by showing up inside agents you don’t control (ChatGPT, Perplexity, Gemini) or by building an assistant on your own platform? The two paths pull on different budgets and teams. 

Third: are you building on architecture flexible enough to survive the next eighteen months of protocol churn? Nobody’s placed the winning bet yet, and headless, API-first infrastructure is what keeps that bet optional rather than forced.

If you sell to consumers who are themselves going to be represented by an agent, it’s also worth thinking about what that means for how you retain them. We’ve already separately shared more about what agentic AI means for retail, and about the buyer agents your sell-side strategy needs to be designed against.