Agentic Commerce Is Coming for Your Checkout: How AI Agents Will Buy From You by 2027

For thirty years, e-commerce has optimized for one buyer: a human with eyes, patience, and a credit card. The next buyer has none of those. Agentic commerce — AI agents that discover, compare, authorize, and complete purchases on a user's behalf — moved from concept to production infrastructure in the past twelve months, and the numbers say it's no longer a futurism topic: it's a 2026 line item that becomes a 2027 channel. This post explains how a machine-executed purchase actually works, what Google, OpenAI, Stripe, Visa, and Shopify shipped while most merchants weren't looking, and what to do about your store before agents start routing around it.

Agentic commerce is the model where an autonomous AI agent acts as the buyer's proxy across the full purchase lifecycle — discovery, cart, authorization, payment, and post-purchase — executing from a *goal* rather than a click. The user says "order trail-running shoes under $150 that arrive by Friday"; the agent evaluates options across merchants, builds a cart, proves it has the user's consent, pays with a scoped token, and reports back. Your storefront's job shifts accordingly: from persuading a human to convincing a machine — with structured data, protocol endpoints, and verifiable trust signals instead of hero banners and urgency timers.

Why "by 2027" isn't hype — the 2026 receipts

The skeptic's version of this topic died sometime in the last two quarters. The evidence:

• The money is already flowing. eMarketer projects AI platforms will process $20.9 billion in US retail spending in 2026 — roughly 1.5% of US e-commerce and about 4× the prior year. McKinsey projects agentic commerce unlocking $3–5 trillion globally by 2030. A 4×-per-year curve on a $20B base is how channels are born.
• The infrastructure shipped. Google, OpenAI, Anthropic, Microsoft, Stripe, Visa, Mastercard, and Shopify have all released production agentic-commerce protocols within the last twelve months. This is not a whitepaper phase.
• Merchants are already enrolled — many without noticing. Over 1 million merchants are live on OpenAI's Agentic Commerce Protocol through Shopify's auto-enrollment alone. PayPal's "Agent Ready" program auto-enrolled millions more for AI-surface acceptance. Microsoft launched Copilot Checkout in January 2026 with PayPal anchoring payments. If you're on a major platform, you may already be sellable-to-agents; the question is whether you're *good* at it.
• The card networks moved. Visa launched Intelligent Commerce Connect in April 2026 as a vendor-neutral on-ramp for verified agents; Mastercard co-developed the Verifiable Intent system. When Visa and Mastercard build agent infrastructure, the networks have concluded machine buyers are inevitable.

The protocol stack, translated

The confusing part of agentic commerce is the acronym soup. The clarifying insight: the standards *stack* rather than compete — a single agent purchase can use all of them, each answering a different question.

UCP — Universal Commerce Protocol (Google, Shopify, 20+ partners): "What can I buy here, and how?" UCP covers the full journey — discovery, cart, checkout, orders, post-purchase. Merchants publish a machine-readable manifest at `/.well-known/ucp` declaring which capabilities they support and where the endpoints live. Instead of an agent learning every retailer's unique checkout, it reads the manifest and transacts. Think of it as robots.txt's ambitious sibling: not "what may you crawl" but "here's how to buy from me."

ACP — Agentic Commerce Protocol (OpenAI + Stripe): "How does the payment execute inside an AI surface?" ACP handles checkout without the user ever leaving the chat. Its key mechanism is the Shared Payment Token — merchant-bound, amount-bound, time-bounded, single-use — so the agent never touches raw card credentials. It's the standard behind ChatGPT's shopping infrastructure, with product feeds, a five-endpoint checkout API, and Stripe as the primary rail.

AP2 — Agent Payments Protocol (initiated by Google, now FIDO Alliance–governed): "Who authorized this?" AP2 is the trust layer: cryptographically signed mandates proving the human actually consented, plus scoped spending limits ("max $200 for hiking gear") and one-time execution tokens that firewall the agent from the user's actual accounts. In April 2026, Google donated AP2 to the FIDO Alliance with sixty contributing organizations — the moment it stopped being a vendor product and became internet plumbing. It's payment-agnostic: cards, bank transfers, even stablecoins via the x402 extension.

MCP — Model Context Protocol (Anthropic): the data layer beneath all of it. MCP is how agents connect to merchant catalogs and systems in real time; ACP and UCP both integrate with it for discovery and capability negotiation.

The one-line summary: UCP says what's for sale, ACP executes the checkout, AP2 proves consent, MCP carries the data. Most merchants will eventually speak several of these — mostly via their platform or payment provider rather than hand-built integrations.

What the last six months taught us (including the failure)

Honest reporting requires the messy part: OpenAI retired Instant Checkout in March 2026 after limited merchant uptake, pivoting to retailer-operated "ChatGPT Apps" — Walmart's Sparky, Etsy, Target, Instacart, Expedia — while keeping ACP alive as the infrastructure layer underneath. Two lessons worth extracting. First, the consumer-facing *shape* of agentic commerce is still being negotiated: fully in-chat checkout versus agent-assisted handoff to merchant-controlled surfaces is an open fight, and big retailers clearly prefer keeping the customer relationship. Second — and more important for planning — the *infrastructure* survived the product pivot. Protocols, feeds, tokens, and mandates all kept shipping. Bet on the plumbing, not on any single storefront experience.

Fees are part of the negotiation too: agent-surface transactions have carried platform fees in the ~2–4% range on top of standard payment processing in early implementations. That economic layer will keep shifting — another reason to build on the open protocols rather than optimize for one platform's current terms.

Selling to a machine: what actually changes

An agent evaluating your store doesn't see your design. It sees:

Structured data or nothing. Complete product schema, accurate feeds, machine-readable specs, real-time price and availability. Every gap is a comparison your product silently loses — an agent won't email you to ask about dimensions.

Programmatic answers to trust questions. Return policy, warranty, shipping cutoffs, and stock status need to be verifiable data, not paragraphs on a policy page. Agents shortlist merchants that answer these questions cleanly and skip ones that don't.

Deterministic checkout. Surprise fees at step four, forced account creation, CAPTCHA walls, and dark patterns don't reduce agent conversion — they end it. A machine buyer treats friction as failure and reroutes to the next merchant. Ironically, everything CRO teams spent a decade fighting for (transparent pricing, guest checkout, clear inventory) is exactly what agents require.

A different loyalty game. Agents don't feel brand affinity; they optimize the user's stated constraints. The durable advantages become being the *remembered default* ("reorder my usual from X"), winning on verifiable attributes (price, speed, return terms), and review/reputation data — which agents weight the way humans weight a friend's recommendation.

The readiness checklist (in order)

1. Get your data spine right — now. Complete Product/Offer/Review schema, clean feeds, real-time availability. This is the same work that wins AI shopping recommendations today (we covered it in our AEO for e-commerce post), which is exactly why it's the no-regrets move: it pays in the current channel while qualifying you for the next one.
2. Know your enrollment status. If you're on Shopify, check your ACP/UCP exposure in admin — you may already be agent-visible. Same for PayPal Agent Ready. Being enrolled without optimizing is how you show up badly.
3. Audit checkout for machine-compatibility. Guest checkout available, full price computable before final step, no interaction-gated information, bot management configured to *verify* legitimate agents (Visa's TAP, agent attestation) rather than block them wholesale. Your WAF settings are about to become a revenue setting.
4. Make policies machine-verifiable. Returns, warranty, shipping SLAs as structured data.
5. Decide your agent strategy consciously. Which surfaces you want to sell through, at what fee tolerance, with what data shared. "We'll wait and see" is a decision too — it just gets made for you by defaults.
6. Instrument agent traffic separately. Agent-driven sessions convert and behave differently from human ones; if your analytics can't distinguish them, you can't manage the channel.

The honest timeline: 2026 is the infrastructure year — protocols hardening, auto-enrollments spreading, single-digit-percentage volumes. 2027 is when agent-initiated purchases become a visible revenue line for prepared merchants, because the curve is already running at 4× annually and the payment networks have finished laying track. The merchants who treat their product data, checkout, and policies as an API for machine buyers will take a disproportionate share of that line — the same way early mobile-optimized stores took a disproportionate share of the last platform shift.

Most of this readiness work is concrete engineering: schema, feeds, checkout flows, bot verification, analytics. If you want to know exactly where your store stands against an agent's evaluation — and what a prioritized fix list looks like — that's a conversation we're already having with e-commerce clients.