At NRF this year, two announcements cut through the noise.

First, Google unveiled the Universal Commerce Protocol (UCP), an open standard designed to let AI agents participate directly in shopping journeys, from discovery through checkout and post-purchase support. Then, almost immediately, Walmart and Google announced a new experience built on UCP, embedding Walmart’s full assortment, pricing, and fulfilment directly into Google’s Gemini.

The headlines are understandably excited.
Agent-led commerce has arrived. Shopping is becoming conversational. Discovery, decision, and checkout are collapsing into a single flow.

All of that is true.

But there’s a deeper shift happening underneath the announcements, one that matters far more than any single protocol or partnership.

Agentic Commerce Changes Where Decisions Are Made

What UCP and similar initiatives signal is not just a new interface for shopping, but a new decision surface.

Instead of customers navigating websites, comparing tabs, and responding to static offers, AI agents now:

• interpret intent in real time,
• surface relevant products dynamically,
• and increasingly, make or mediate decisions on the customer’s behalf.

In Google’s own examples, brands can offer a discount in the moment while a user is asking an AI for recommendations. Walmart’s integration goes further — blending discovery, personalisation, fulfilment, and membership benefits directly into an agent-led flow.

This is a fundamental departure from traditional eCommerce and retail media models, which were built around:

• batch segmentation,
• delayed attribution,
• and post-hoc optimization.

In an agentic world, the moment of influence is live. If you miss it, there is no retargeting safety net later.

The Industry Is Celebrating the Interface, But the Real Shift Is Architectural

Most commentary on agentic commerce focuses on:

• LLM capabilities,
• conversational UX,
• or protocol interoperability.

Those are important, but they are not the constraint.

As models converge and agent frameworks proliferate, the real differentiator becomes what the agent can learn from, and how quickly it can act.

In other words:
AI doesn’t start with agents.
It starts with data flow.

If your systems only understand behavior after it has been logged, warehoused, and analysed, then your AI — no matter how advanced the interface — is operating on a delay. You are explaining outcomes, not shaping them.

Agent-led commerce collapses the timeline. It demands intelligence that can:

• observe intent as it forms,
• update belief continuously,
• and intervene while the outcome is still mutable.

That requires a very different foundation.

Why the “First Mile” Becomes Mission-Critical

In an agentic environment, decisions are made upstream — before sessions end, before carts are abandoned, before spend is wasted.

This is where first-mile intelligence matters.

From a server-side position in the data flow, it becomes possible to:

• observe clean, high-fidelity behaviorial signals,
• anchor those signals to identity in a privacy-safe way,
• and update intent models in real time.

This is not about building yet another model.
It is about where learning occurs relative to behavior.

When inference and learning happen in the flow, not days later in a warehouse, AI systems can move beyond prediction and into policy:

• Should this user see a discount right now?
• Should spend be suppressed because this journey looks non-human?
• Should the agent recommend complementary items, or hold back?

These decisions cannot wait for batch retraining cycles. They must be made mid-flight.

Agentic Commerce Exposes the Limits of Legacy AI

The irony of the current moment is that while agents are becoming more capable, many enterprise data stacks are still operating in what I often describe as data archaeology.

They dig through the past sessions, conversions, attribution logs, trying to infer what might work next time.

But in an agent-led world, next time may never come.

If an AI agent has already mediated the decision, the only moment that mattered was the one you missed.

This is why protocols like UCP are necessary — but not sufficient.

They standarize interaction.
They do not solve learning latency.

From Tools to Systems That Learn in the Moment

One of the most interesting implications of agentic commerce is that AI can no longer be treated as a feature or a tool.

It must behave more like a real-time system:

• continuously learning,
• continuously evaluating,
• and acting only when confidence thresholds are met.

Think less “model deployment” and more AI general manager, one that adjusts policy incrementally, validates decisions against live outcomes, and compounds learning across use cases.

This is where behavioral foundations matter. A single, continuously updated representation of behaviour can support:

• bot detection,
• conversion likelihood,
• brand or SKU affinity,
• lifetime value estimation.

Different questions.
Same underlying understanding.

The NRF Signal Is Clear, The Work Starts Now

The announcements from Google, Walmart, Shopify, and others are not experiments. They are signals.

Agentic commerce is no longer theoretical.
It is entering production.

But success in this new landscape will not be determined by who integrates the most agents or adopts the latest protocol first. It will be determined by who can learn and act fastest at the moment intent is revealed.

As interfaces become conversational and models commoditise, first-party behaviorial data, captured and activated in real time, becomes the scarce asset.

AI doesn’t start with agents.
It doesn’t start with models.

It starts with data flow, and with the ability to shape outcomes while they are still unfolding.

‍