Agentic retail is here

When I wrote my previous articles on trust and payment in agentic retail and how it will actually arrive in reality before Christmas, I didn’t really expect to see so many of the building blocks I talked about drop into place quite so quickly. I probably should have known better.

Since then, a number of ideas that still felt abstract or speculative have started to harden into real systems: protocols, integrations, and production checkout flows that allow agents to participate directly in commerce. Not hypothetically — but in ways people can actually use.

Google’s announcement of the Universal Commerce Protocol (UCP) at NRF last week, alongside Microsoft’s Copilot Checkout launch the week before, has created significant waves — and for good reason.

UCP: real progress, not theatre

The Universal Commerce Protocol (UCP) is a meaningful step forward for agent-assisted commerce.

It’s an open, implementable specification that models the full consumer commerce lifecycle: discovery, cart construction, checkout sessions, identity linking, order state, and post-purchase events.

It’s designed to interoperate with AP2 and adjacent agent protocols, and it gives merchants a structured, policy-aware way to expose their systems to automated intermediaries without rewriting their commerce stack.

That alone would be noteworthy. What makes UCP more significant is that this isn’t just a paper exercise.

Google has already shipped working implementations across its own AI surfaces, wiring real merchants, real catalogues, and real checkout flows into production systems.

 Anyone who’s been around long enough knows the difference between a plausible spec and something that survives contact with payments, identity, refunds, and edge cases.

UCP has crossed that line.

What this enables in practice

Taken together, UCP and AP2 define a clear, usable model for how agents can participate in commerce today.

Agents can reason about what’s available, assemble carts, negotiate offers and pricing rules, initiate checkout, track orders, and handle post-purchase events — all through explicit, machine-readable interfaces rather than scraping or guesswork.

Delegated execution is part of this model. Through signed mandates, agents can act in human-not-present scenarios, provided the user has authorised a specific intent in advance. Authority is explicit, attributable, and auditable — whether approval happens synchronously or ahead of time.

The result is a system where:

• Agents can meaningfully act, not just recommend

• Transactions can complete without constant human supervision

• Responsibility is clearly expressed, not implied

From a user’s perspective, this already looks and feels like a complete purchase journey. From an agent’s perspective, the mechanics of commerce are largely solved.

What varies is not capability, but how that capability is invoked.

Mobile first

It’s also worth noting where this is most likely to show up first. Agentic retail maps unusually well to mobile.

Small screens favour intent over browsing, background execution over step-by-step flows, and quick confirmation over prolonged decision-making.

Payments, identity, and authentication are already native, and users are accustomed to delegating simple tasks to apps they trust.

In that context, agent-assisted — and increasingly agent-executed — commerce feels less like a leap and more like a continuation.

What happens next

At this point, most of the foundational pieces are at least proven in concept.

The protocols exist.

The payment rails exist.

The mandate models exist.

And some of this is already running in production.

Usable retail agents can now exist — not as demos or assistants, but as systems that can reason, negotiate, and complete real transactions under clearly defined rules.

From here, progress will be shaped less by what the technology allows and more by what markets choose to adopt. Different categories, cultures, and regions will make different decisions about how much authority they’re comfortable delegating, and under what conditions. Adoption is unlikely to be uniform or linear and that’s probably a good thing.

If there’s one near-term pattern to watch, it’s low-friction, low-engagement shopping on mobile, the kinds of tasks where delegation feels helpful rather than risky. Simple reorders, household basics, and time-constrained purchases are natural candidates.

Commuter shopping may become an early proving ground: busy professionals briefing agents on the way into work and letting them handle the routine tasks in the background. Not everywhere, and not for everything, but enough to establish new habits.

What UCP represents isn’t the arrival of a single, fixed model of agentic retail, but the moment where speculation gives way to practice — and where the market starts deciding, context by context, how far it actually wants to go.