Agentic Commerce: AI That Shops and Pays for You

The checkout button is dying. Not because nobody shops online — the opposite. Two competing protocols launched in the last six months have made it technically possible for AI agents to browse, decide, and pay on your behalf without you ever opening a browser tab.

This isn't a prediction. It's already running in production.

The Infrastructure Nobody Talked About

For AI shopping agents to work end-to-end, three things have to be true simultaneously: the agent needs to understand what you want, it needs access to structured product data across merchants, and it needs to be able to initiate a payment without handing off to a human. The first two have been roughly solved since late 2024. The third is what 2025 quietly built — and what 2026 is now deploying at scale.

The plumbing that makes this work sits on a few new infrastructure layers. OpenAI shipped its Agentic Commerce Protocol (ACP) in September 2025 alongside ChatGPT's Instant Checkout feature. Stripe handles the payment processing; a Shared Payment Token means the agent never touches the actual card number. The transaction stays inside the ChatGPT conversation. The merchant charges nothing extra to the shopper, but OpenAI takes a 4% transaction fee from the merchant on every completed Instant Checkout purchase.

Then in January 2026, Google went in the opposite direction — open standard instead of proprietary protocol. The Universal Commerce Protocol (UCP) is an open-source standard developed in collaboration with Shopify, Etsy, Wayfair, Target, and Walmart, and endorsed by over 20 global partners including Adyen, American Express, Best Buy, Mastercard, Stripe, The Home Depot, Visa, and Zalando. The technical bet is that a shared language between AI surfaces, merchants, and payment processors is more durable than any single platform owning the stack.

Both protocols are live. ChatGPT's Instant Checkout has been serving 900 million weekly users since September 2025. Google's UCP is powering a new checkout feature directly on product listings in AI Mode in Search and the Gemini app, allowing shoppers to buy from eligible US retailers without leaving the conversation.

The question isn't whether this technology exists anymore. The question is what it actually changes — and for whom.

How the Protocols Actually Work

The protocols share the same core architecture but make different bets on openness and control.

ACP (OpenAI + Stripe)

ACP is a closed-but-accessible protocol. Merchants onboard through a formal application process (or, if on Shopify, through a simpler admin toggle). Etsy sellers are automatically included through Etsy's Offsite Ads program. Shopify merchants require a paid plan, Shopify Payments enabled, and a US-based location.

At the technical level, the agent never sees the cardholder's payment credentials. Stripe issues a Shared Payment Token — a tokenized authorization that allows a transaction to complete without card-present authentication at every step. The merchant remains the merchant of record. Currently ACP supports single-item purchases only, with multi-item carts, PayPal integration, and international expansion on the roadmap.

UCP (Google)

UCP is architecturally more interesting because it's designed as a protocol layer, not a product. UCP provides a common language for consumer surfaces, business backends, and payment providers, supporting APIs, Agent2Agent (A2A) communication, and the Model Context Protocol (MCP). This means any AI surface — not just Gemini — can theoretically implement UCP and get access to the same merchant inventory.

The payment side runs through Google Pay and Google Wallet credentials, with PayPal mentioned as an upcoming payment method addition. Merchants remain the seller of record, retaining full control over customer relationships and transaction data.

Here's the concrete flow, which Google's own documentation makes explicit:

User query: "Find a lightweight suitcase for an upcoming trip."
     ↓
Gemini / AI Mode in Search
     ↓
UCP: Product discovery request → Merchant catalog (via Merchant Center feed)
     ↓
UCP: Cart creation + discount/loyalty application
     ↓
UCP + AP2: Payment authorization (Google Wallet / Google Pay)
     ↓
Order confirmation → Merchant fulfills

The agent never leaves the surface. The merchant never loses the transaction. The user never enters a card number.

The Payments Layer

Behind both protocols sit the card networks — who have been building their own agent-authentication frameworks. Visa and Mastercard have been racing to forge the systems needed for AI-initiated purchases, with executives describing this as a shift as significant as the move from brick-and-mortar to e-commerce.

Visa introduced a Trusted Agent Protocol with more than 10 partners to help merchants distinguish legitimate AI agents acting on behalf of consumers from unauthorized bots. This matters because fraud detection systems built for human browsing behavior are poorly calibrated for agents — which operate at different speeds, with different concurrency patterns, across many sessions simultaneously.

Stripe is developing mandate-based flows in which cryptographically signed instructions authorize transactions without manual checkout. Fraud detection models need to be retrained to distinguish the velocity signatures of legitimate automation from malicious bot activity.

The fraud problem isn't solved. It's being actively worked around.

Why This Breaks Things (Deliberately)

Agentic commerce doesn't just add a new checkout option. It restructures the entire customer acquisition funnel — and most of it benefits the platform, not the merchant.

Discovery is now decided by the agent, not the shopper. When someone asks Gemini or ChatGPT for "a running shoe under $120 for wide feet," the agent selects what surfaces. The merchant's job shifts from winning a human's attention on a product page to being legible to a machine reading structured product feeds. Structured data, enriched metadata, and clean catalogs now determine whether an agent can understand and recommend a product at all.

This is the SEO problem compressed into a sharper knife. Rank well in traditional search, you get traffic. Be machine-readable in agentic commerce, you get the transaction. Be invisible, you don't get considered at all.

The website becomes optional. The entire shopping journey and decision-making process happens inside the conversation. The user never leaves ChatGPT. There is no click-through to the merchant's website. For brands that have spent a decade optimizing landing pages, A/B testing CTAs, and building retargeting audiences — all of that conversion infrastructure now sits behind a wall the agent doesn't walk through.

Platform dependency deepens. Merchants retain "merchant of record" status under both protocols, which sounds reassuring until you think about what the platform controls: the recommendation, the discovery, the checkout experience, and (in the case of ACP) a 4% fee on every completed sale. Merchant of record means liability; the platform keeps the relationship.

One concern raised by practitioners is whether UCP favors the single most coherent "default" brand rather than creating a marketplace of choices — effectively a "Default Economy" where only one answer is surfaced per query. This is the right question. Search SEO created a brutal competition for page-one results; agentic commerce may compress that to a competition for the single product the agent recommends.

The Counterarguments Worth Taking Seriously

Consumer adoption is the actual bottleneck. In a July 2025 survey of 700 consumers conducted by Forrester, only about a third said they would be willing to complete payment through an answer engine at all, citing data privacy concerns. Handing an AI agent the ability to spend your money requires a level of trust that takes time to build — and one high-profile incident (an agent buying the wrong thing, or a fraud case through an agent-initiated transaction) could slow adoption significantly.

The experience isn't there yet. Amazon CEO Andy Jassy told analysts that most AI shopping agents fail to provide a satisfactory customer experience — lacking personalization and often providing inaccurate pricing and delivery estimates. Generic product recommendations dressed up as AI-driven personalization aren't compelling enough to change shopping habits. This is the "needle in a haystack feeling" problem — until users consistently experience an agent that genuinely understands what they want, adoption will remain shallow.

Amazon is structurally difficult to displace. Amazon already has stored payment credentials, consumer trust, and habitual usage. For an AI company like OpenAI to win, it needs to convince shoppers it's better to buy on ChatGPT than on a retailer's website — which usually requires bigger assortment or lower prices. Amazon has both, plus Prime loyalty. That's not an easy moat to walk around.

Smaller brands may actually benefit. The counterintuitive case: traditional search advertising rewards whoever has the biggest ad budget. Agentic commerce rewards whoever has the best-structured product data. Because UCP is an open standard, a small brand with clean, enriched product data could theoretically be surfaced ahead of a large brand with messy catalogs — without needing the ad spend that would be required in traditional search. This is speculative, but it's the optimistic version of how this plays out.

What Developers and Product Teams Should Do Now

The structural shift is real enough to warrant action, even if mass adoption is 18–24 months away. Here's where to focus:

Audit your product data for machine readability. The agent equivalent of SEO is structured data. Product feeds need clean, enriched metadata — not just price and availability, but answers to common questions, compatible accessories, return policy details, and shipping timelines. If a UCP-enabled agent can't parse your catalog, it can't recommend you.

Understand the protocol landscape before you integrate. Businesses deeply embedded in the Google and Shopify ecosystems will find UCP the more natural fit. Businesses with significant ChatGPT usage among their customer base may need ACP compatibility separately. Most mid-sized retailers will need both. These aren't mutually exclusive — but they require separate integration work.

Think about fraud infrastructure early. If you're building payment flows or working on merchant-side integrations, the agent authentication problem is not solved. AI agents generate transaction patterns that differ significantly from human browsing — particularly in speed and concurrency. Detection models need to be retrained to distinguish legitimate automation from malicious bot activity. Build for this now rather than patching after incidents.

Watch the fee structures. A 4% transaction fee (ACP/OpenAI) is meaningful at scale. If agentic commerce captures even 15% of a merchant's online volume, the economics of that channel look different from traditional e-commerce. Model this before you opt in at scale.

The Takeaway

• Two protocols are live now, not in pilot: OpenAI's ACP (via ChatGPT Instant Checkout) has been running since September 2025; Google's UCP launched in January 2026. Both enable purchase completion without leaving the AI interface.

• The merchant funnel is being restructured: Website traffic, landing page optimization, and retargeting audiences matter less when the purchase completes inside a conversation. Machine-readable product data is the new SEO.

• Consumer trust is the actual constraint: Infrastructure is ahead of adoption. Roughly a third of US consumers are currently willing to let an agent complete a payment. That number will grow — but not as fast as the platforms would like.

• Fraud and authentication are unsolved: Both Visa and Stripe are building agent-authentication frameworks, but AI agents create transaction patterns that break existing fraud detection. This is a real risk for early adopters on both the merchant and consumer side.

• The fee and platform dependency dynamics are real: "Merchant of record" status sounds like control, but ACP charges 4% per transaction and both platforms control what gets recommended. Model the economics before scaling.

---

Tags: agentic-commerce, AI-agents, ecommerce, payments, OpenAI, Google