For thirty years, every online purchase began with a human: someone browsed, hesitated, and clicked buy. The entire payment stack was built on that quiet certainty. In 2026, it changed.
The buyer arriving today often is not a person at all, but an agent shopping on someone's behalf, and it does not window-shop. Adobe Analytics found that AI-referred traffic to US retail sites not only grew nearly fourfold in early 2026, it converted better than every other channel. These visitors come to buy.
So the question for merchants is no longer whether agents will reach the checkout. It is whether the checkout is ready for a buyer who leaves no trace.
An AI agent shops on a person's behalf: it discovers, compares, and buys inside limits the person set once, often while that person is busy elsewhere.
That small handoff breaks a large assumption. Identity, intent, and authorization were always observable because a human sat at the keyboard, leaving a session to read and a device to trust. An agent leaves none of that. The payment arrives stripped of the very context fraud models were built to read, and the merchant has milliseconds to approve it anyway.
This is where enterprise teams underestimate the work. There is no one standard to adopt. Several landed at once, each from a different layer of the stack.
OpenAI and Stripe released the Agentic Commerce Protocol in September 2025, putting checkout inside ChatGPT. Mastercard launched Agent Pay in April 2025, binding a card credential to a specific agent and consent policy. Visa followed with its Trusted Agent Protocol, now folded into Visa Intelligent Commerce. Google launched its Universal Commerce Protocol at NRF in January 2026.
The point is not which one wins. Most agent flows already touch two or more, and Visa all but admitted it by releasing Intelligent Commerce Connect, a single doorway into four protocols at once. When Juniper Research ranked the leading providers in April 2026, Mastercard, Visa, and Stripe took the top spots for moving early across all of them. Fragmentation is not a phase. It is the floor you build on.
That same Juniper Research study named trust as the biggest barrier to agentic commerce, ahead of every technical hurdle. At the moment of sale, trust comes down to three questions.
Did the buyer actually approve this? Signed consent records tie a transaction back to a real human decision, not an agent improvising beyond its remit. Is this a real agent or a scraper in its clothes? Verified credentials wave the genuine ones through and keep bad actors out. And who owns it when it goes wrong? Chargeback rules assume a person pressed buy, so tokens now carry the agent's identity into the record, letting disputes land where they should.
The first crack shows before the payment even starts. Adobe found that retail product pages are on average only 66% machine-readable, so a third of the content where buying happens is invisible to the models sending the traffic. If an agent cannot read your catalog, the sale never reaches checkout.
The rest is decision making. With session history gone, risk models fall back on transaction-level data, which only works with a unified view of payments rather than signals trapped inside each processor. Routing follows the same logic: an agent paying from an unfamiliar device looks odd to an issuer, and smart routing that retries elsewhere is what protects approval rates.
Underneath it all sits a readiness gap. McKinsey found that 88% of organizations use AI somewhere, but only 7% have fully scaled it. Adopting a protocol and being ready to transact through it are not the same thing.
None of this replaces the networks, processors, or fraud engines you already trust. The new protocols extend their strengths rather than retire them, and as Juniper Research put it, agentic commerce complements traditional checkout for the foreseeable future, it does not replace it. Agents do not break your stack, they stress-test it.
The job is making that stack governable. Orchestration gives a merchant one place to route every transaction in real-time context, human or agent alike, instead of wiring each protocol and provider by hand. That is what DEUNA is built to do. On top of it, Athia, its agentic payments intelligence layer, reads across checkout, payments, and fraud, learning where approvals stall and where revenue leaks so routing keeps improving as patterns shift. When no human is around to catch a failed payment, a stack that adapts on its own is not a nice-to-have. It is the baseline.
