In April 2026, Stripe launched Link's wallet for agents: a way for an autonomous AI agent to pay on a person's behalf without that person ever handing over their card number. A month earlier, together with OpenAI, it had published the Agentic Commerce Protocol (ACP), the open standard that now powers "Instant Checkout" inside ChatGPT. And they're not the only ones betting on this: Coinbase is pushing x402 in parallel, a protocol for agents to pay directly in stablecoins, already backed by Visa, Mastercard, Ripple, Google, AWS and Stripe itself.
Three large companies building payment infrastructure for machines at the same time isn't noise: it's a signal of where they believe commerce is heading. But it's worth separating what already works from what's still a bet.
How it actually works
Stripe's wallet doesn't hand the agent your card: the person authorises access via OAuth from the web or their phone, and the agent receives either a one-time-use card or a "Shared Payment Token" (SPT) with amount, currency and merchant limits already set before the token ever reaches it. Each spend is reviewed and approved from the app before it executes. It's the same logic as a corporate card limit, applied to an agent instead of an employee.
ACP works differently: it's the protocol that connects a buying agent to a merchant's catalogue to complete a purchase without the person ever leaving the chat. That's how "Instant Checkout" works in ChatGPT today: US users buy from Etsy sellers and, increasingly, Shopify merchants (Glossier, SKIMS, Spanx and Vuori are already on board), right inside the conversation. OpenAI charges a 4% fee on every purchase completed this way.
It's not just cards: the other bet runs on stablecoins
x402 solves a different problem: micropayments between machines, with no card or account involved. It repurposes the HTTP 402 status code ("Payment Required," part of the protocol since the '90s and never actually used until now): an agent requests a resource, the server responds asking for payment, the agent signs a stablecoin transaction (USDC on Base, Coinbase's network) and retries the request with proof of payment attached.
Adoption numbers as of late April 2026: roughly 69,000 active agents using the protocol, more than 165 million transactions, and $50 million in total volume -about 30 cents per transaction on average, which is exactly the micropayment range the protocol was built to solve.
How much this is worth, depending on who's measuring
Worth being honest here: forecasts vary wildly because every firm defines "agentic commerce" differently. McKinsey estimates up to $1 trillion in AI-orchestrated US retail revenue by 2030, and $3 to $5 trillion globally. Bain puts the US market alone at $300-500 billion by 2030 (15-25% of total e-commerce). Morgan Stanley places it between $190 and $385 billion. Gartner projects that 20% of digital commerce transactions will run through AI platforms by that date.
None of these numbers is a certainty: they're the range an industry agrees on when it agrees on little more than "this is going to grow."
The reality today, not the promise
The infrastructure is being built fast -that's a fact, not a projection. Real demand is still uneven: for x402 specifically, multiple industry outlets noted in March 2026 that volume remains low relative to the protocol's ambition. Thirty cents per transaction on average is proof the technology works, not proof there's already a mature market paying at volume.
That's not a reason to ignore it -the companies building this (Stripe, OpenAI, Coinbase, Visa, Mastercard) don't usually bet together on something they don't believe will happen- but it's not a reason to rush out and integrate every protocol this week either. Most of them are still stabilising their own spec.
What's worth doing now
What's worth solving now, regardless of which protocol ends up winning: making sure an external system -AI or not- can query your catalogue, your pricing, and confirm a purchase without depending on a person filling out a form or approving something by hand. An agent can't pay you if it can't first understand what you sell and at what price, and it can't complete the purchase if the last step depends on someone on your team reviewing an email.
That means a catalogue with structured data, clear and programmable pricing rules, and infrastructure that responds reliably when whoever's asking isn't a browser with a person behind it. That's exactly what it takes to scale any sales or billing process, whether or not this particular wave pans out.
If your business still depends on billing or sales processes a person resolves by hand today, it's a good time to automate them before an agent asks you to. We cover how we do that in our AI process automation solution. If you also need the infrastructure to keep that running reliably in production, that's how we solved it for an identity platform in our cloud infrastructure and DevOps solution. If you want us to look at your case, get in touch.
