Visa + OpenAI: Tokenized Payments for Shopping Agents

Agentic commerce payments crossed from experiment to infrastructure on June 10, 2026, when Visa and OpenAI announced at the Visa Payments Forum that tokenized Visa credentials can now power agent-initiated checkout inside ChatGPT and OpenAI's Codex coding agent. The integration routes Visa's tokenization, agent identification, and fraud-monitoring infrastructure into OpenAI's consumer-facing and developer systems.

What makes the timing notable is that Mastercard launched its own framework — Agent Pay for Machines — the very same day at a separate conference. Two of the largest payment networks moved on agentic commerce within hours of each other, without explicit coordination. That convergence, more than any single feature, is the signal: agent-initiated payment rails are no longer a pilot curiosity.

This guide separates what was actually announced from what is already live, walks through how tokenized agentic checkout works, compares the Visa and Mastercard approaches side by side, and lays out a five-layer readiness checklist so merchants can self-assess whether a shopping agent could even find, trust, and transact with them today.

At the Visa Payments Forum in San Francisco on June 10, 2026, Visa and OpenAI announced a strategic collaboration that brings secure Visa payments into OpenAI's AI-agent-driven commerce experiences. The Visa–OpenAI partnership announcement describes integrating Visa's global network — tokenization, authorization, agent identification, and fraud monitoring — into OpenAI's developer and consumer-facing systems.

An important distinction up front: this is the deepening of a relationship that began earlier, not a brand-new product. Visa Intelligent Commerce (VIC) launched on April 30, 2025 with nine founding AI and tech partners — including OpenAI. The June 10, 2026 news is the specific OpenAI integration step, alongside three new merchant tools announced at the same forum: Agent Score (assessing whether agents can navigate a merchant's site), an Agentic Directory of verified agents and merchants, and a Large Transaction Model fraud engine. Visa also showed a Command Line Interface proof of concept for agents paying for digital services in a terminal.

The core mechanism is tokenization. Rather than handing an agent a raw card number, the system issues a secure network token scoped to a specific agent. If any one token is compromised, exposure is limited — and the design is intended to prevent replay attacks, where a captured credential is re-used in a different context.

Crucially, the controls sit with the consumer. User-defined spending limits, merchant-category restrictions, and required-approval thresholds are enforced in real time, before Visa authorizes any agent-initiated payment. On the OpenAI side, the ACP Delegated Payment Spec sends one-time payment requests with spending caps and expiration — meaning a ChatGPT agent cannot reuse a credential or exceed the amount the user authorized for that session. The checkout state and payment processing remain on the merchant's systems; OpenAI is not the merchant of record.

The lineage matters for understanding what is reusable. OpenAI launched ChatGPT Instant Checkout and its evolution with Stripe on September 29, 2025, using the Agentic Commerce Protocol (ACP) co-developed by Stripe and OpenAI as an open standard. Stripe's Shared Payment Token was the first compatible implementation. The Visa integration layers Visa's tokenization and trust infrastructure onto that existing protocol rather than replacing it — which is why the merchant work below is largely additive, not a rebuild.

Most June 10 coverage treated the two announcements as separate stories. Putting them side by side is more useful for a merchant deciding where to invest first. Mastercard Agent Pay for Machines (AP4M) launched the same day at the Mastercard Priceless conference, described as an open protocol enabling AI agents to transact autonomously at machine speed — including micropayments worth fractions of a cent. The clearest structural difference: AP4M stores agent credentials and spending permissions on public blockchains (Polygon, Solana, Base), where Visa uses its own tokenization infrastructure without on-chain credential storage.

For a merchant, the practical read is that you should not wait for a single winner to emerge. The two frameworks are complementary at the integration layer more often than they are mutually exclusive — and our own coverage of Mastercard's verifiable-intent framework for agentic commerce and Stripe's machine payments protocol for autonomous AI agents shows how quickly the standards landscape is filling in. The merchant work — structured data, callable checkout, verifiable trust — is mostly shared across all of them.

The most-cited figure comes from a joint report: a McKinsey/ICSC analysis projecting that agentic commerce could generate roughly $1 trillion in orchestrated US retail revenue by 2030, with a $3–5 trillion range globally. A moderate scenario in that report assigns agents to about 18% of US business-to-consumer spend by 2030. These are forecasts, co-authored by McKinsey and the ICSC — not measured transaction volume, and worth attributing as such.

Demand-side signals are softer than the headline forecast suggests, and should be read with their methodology attached. Visa cites a commissioned survey — 1,000 adults per market across 12 markets, conducted in October 2025 — finding that 47% of US shoppers already use AI tools for at least one shopping task, such as price comparison or recommendations. Because that figure is from a Visa-commissioned survey, treat it as a directional, vendor-stated data point rather than independent research. Reports that a large share of transactions are already "influenced by" large language models exist, but the most widely cited version traces to a single analyst, so we are not putting a specific percentage on it here.

Our reading of the trend: the forecasts are large enough that even a fraction-of-projection outcome would reorder how discovery and checkout work for a meaningful slice of retail — but the uncertainty band is wide, and the demand signals are still early. The rational merchant response is not to bet the business on a 2030 number; it is to make the readiness investments that pay off regardless of which scenario lands, because every one of them (structured data, callable checkout, verifiable trust) also improves conventional search and marketplace performance today.

Vendor press releases rarely publish a single, cross-network checklist a merchant can self-assess against. The table below combines the OpenAI ACP technical requirements, Visa's Trusted Agent Protocol trust signals, and Visa's user-control model into one rubric of five layers — what an agent needs, what most merchants have today, the effort to close the gap, and the standard to use. Read down the layers; the early ones are where most storefronts fall short.

The single most common failure is layer 01. A storefront can look complete to a human while being effectively invisible to an agent that needs structured, queryable attributes — which is precisely the gap Visa's Agent Score is designed to surface. If you want to stress-test where your own catalog sits before committing engineering time, our ecommerce AI agent readiness assessment and our walkthrough of how AI agents navigate the full ecommerce journey from discovery to checkout map the same layers onto a practical audit.

A frequent point of confusion: Visa's Trusted Agent Protocol (TAP) and OpenAI/Stripe's Agentic Commerce Protocol (ACP) solve different problems and are used together, not as alternatives. TAP, introduced by Visa in October 2025, is a trust layer — it lets merchants distinguish legitimate AI agents from malicious bots. ACP is a commerce layer — it defines how an agent reads a product feed, runs checkout, and delegates a bounded payment.

The Visa Trusted Agent Protocol specification is published as an open spec. Its agent signatures include timestamps, unique session identifiers, key identifiers, and algorithm identifiers, and are cryptographically bound to a specific merchant domain and operation — which is what prevents signature reuse and replay across different contexts. Akamai joined the TAP framework to provide behavioral intelligence and fraud controls. On the commerce side, the OpenAI Agentic Commerce Protocol merchant requirements spell out the three flows: product feed, checkout handling, and delegated payment.

That quote captures the whole problem in one sentence. Most stores were architected for a human cursor, not a machine reader. TAP answers "is this agent trustworthy?" and ACP answers "can this agent actually transact here?" — and a merchant preparing for Visa-backed agentic checkout needs both. For a wider view of how the standards are proliferating, see our coverage of Google's Universal Commerce Protocol for AI shopping agents and the broader Q2 2026 agentic commerce platform comparison.

Almost all coverage of the announcement focused on consumer shopping agents. The Visa–OpenAI partnership explicitly names a second category: enterprise use via OpenAI's Codex coding agent, potentially letting autonomous agents purchase developer services, APIs, and cloud inference within user-set spending limits. Visa's Command Line Interface proof of concept, also shown at the June 10 forum, points in the same direction — agents paying for digital services directly in a terminal.

This opens a distinct merchant category that consumer-retail framing misses entirely: API marketplaces, SaaS vendors, and cloud providers. If an agent can be authorized to buy inference or developer tooling within bounded limits, then those vendors face the same readiness question as a retailer — can an agent discover, trust, and transact with you — but with usage-metered, machine-speed purchasing as the norm rather than the exception. Pricing pages built for a human in a billing portal are the developer-tooling equivalent of a storefront that an agent cannot parse.

The readiness work has a natural order, and it front-loads the items that also help conventional channels. Fix the product feed first: it is the highest-leverage, lowest-regret move, because a clean, attribute-complete feed improves marketplace and search performance today and is the precondition for any agent to find you tomorrow. Then expose checkout as a callable session, register for trust frameworks, and wire up delegated payment with bounded caps.

Looking forward, the most likely path is that the demand curve lags the infrastructure by a year or more — capabilities are announced now, but broad consumer adoption builds through the 2026 holiday season and beyond. That lag is the opportunity. Merchants who use the next two quarters to close the five layers will be discoverable and transactable the moment volume arrives; merchants who wait for proof of volume will be registering trust signals while competitors are already in agents' consideration sets. Proof-of-human and verifiable-identity frameworks — covered in our guide to proof-of-human frameworks for agentic commerce — will increasingly sit alongside this readiness work.

The effort profile above is a planning heuristic, not a measured benchmark — every merchant's stack differs. But the ordering is stable: the controls layer is usually configuration, the feed and trust layers are integration work, and exposing checkout as a callable session is the heaviest lift because it touches fulfillment, tax, and payment logic that was historically tied to a browser UI. Scope accordingly. If you want help mapping this to your platform, our ecommerce engagements and AI transformation work start with exactly this kind of readiness audit.