NRF 2025: The Agentic AI Pitch, Assessed

Reading the NRF 2025 coverage, there was a clear vocabulary shift from the previous year. If NRF 2024 was the moment AI stopped being a theme and became the entire conversation, NRF 2025 was where the conversation narrowed to a single, contested word: agents.

The Javits Center in New York hosted the Big Show across 12–14 January, and by the time the post-show analysis came through, the pattern was obvious. Salesforce launched Agentforce for Retail, positioning it as the start of an "Agentic AI Era" where AI stops being something you query and becomes something that acts. Google Cloud announced Agentspace for Retailers, calling it "the present reality, transforming customer experiences from browsing, to buying, and beyond." IBM released research from its Institute for Business Value finding that retail and consumer product companies plan to allocate an average of 3.32% of revenue to AI. Constellation Research's analyst coverage called the show "a retail AI agent parade from enterprise vendors." Accurate, and slightly unflattering in exactly the right way.

"Digital labour" was the phrase that floated around. It deserves some scrutiny.

What Changed Between 2024 and 2025

The honest answer: the underlying models got better, inference costs fell further, and the tooling for connecting AI to external systems matured enough that "agent" became a word you could use in a product demo without it being entirely fictional.

An AI agent, in the sense the retail industry was now using the term, is broadly a system that can reason about a goal, figure out what steps to take, call external tools or data sources, and iterate toward a result without human hand-holding at each step. The standard illustration: your agent notices stock is running low on a fast-moving line, checks your supplier's availability API, raises a purchase order within policy parameters, and alerts you only if something falls outside tolerance. That is genuinely useful. It is also not as far from existing automation as the "agentic era" framing implies.

The backdrop context from Adobe's holiday season recap (which the NRF community was absorbing in the days before the show) showed a 1,300% increase in traffic to US retail sites from generative AI chatbots during the 2024 holiday season. The base was still modest, but the directionality was unmistakeable and gave the NRF narrative something to anchor on.

What was less visible in the post-show reporting was evidence of agents in production at scale, dealing with the messy reality of inconsistent data, edge cases, and the organisational friction that comes from any system that takes action without waiting to be asked.

Two Different Pitches

The most interesting contrast to emerge from the show was the divergence between Microsoft and Salesforce's approaches. Salesforce went hard on transformation: Agentforce agents as autonomous digital workers, restructuring how retail work gets done. The launch detailed pre-built skills for order management, guided shopping, appointment scheduling, and loyalty promotion creation: a productised library of agent capabilities that retailers could deploy without building from scratch.

Microsoft's approach, by contrast, was integration-first: AI that works within and augments existing systems, positioned as an enhancement to what retailers already have. Practical, somewhat unglamorous, and arguably the thing most retailers actually need.

Both are probably right about their respective customer bases. Enterprise retailers running on Salesforce Commerce Cloud will take the Agentforce path, buying into the transformation narrative alongside the platform investment. Everyone else will take something closer to the augment-first approach, whatever they call it.

Where the Serious Use Cases Live

Away from the keynote framing, the use cases getting the most serious attention from retail practitioners in post-show coverage were in category management, promotion planning, and supply chain. These are areas where the data inputs are structured enough, and the decision logic defined enough, that agents have a reasonable chance of doing useful work without constant supervision.

An agent that monitors sell-through by location, identifies early-stage markdown risk, and surfaces recommendations for reallocation before the problem compounds is not glamorous. It has a clear ROI and a clear failure mode, which makes it considerably easier to deploy than open-ended customer-facing applications where the agent's behaviour under unexpected inputs is harder to bound.

For UK retailers, this supply-chain and category-management layer is where the near-term case for agentic AI is strongest. The enterprise platform vendors at NRF are building for global clients, but the use cases they were demonstrating (demand signal processing, automated purchase order management, dynamic promotion allocation) translate directly to the operational challenges that British retailers of any size are wrestling with.

The Question Worth Asking

The word "agentic" will be everywhere in 2025. A lot of things that are not really agents will be described as agents, in the same way that a lot of things that were not really AI were described as AI for the previous decade.

The useful question is not "do you have agents?" It is: "what decisions are you delegating to them, with what guardrails, and how are you measuring whether they are making good ones?" The retailers who will get real value from this technology are the ones asking the second question.

What the NRF announcements established is that the platform infrastructure for agentic retail is now being built in earnest. Whether that infrastructure gets used wisely (with the right governance, the right data, and realistic expectations about where agents can and cannot be trusted to act autonomously) is a question the show, predictably, did not answer.

---

For a one-year assessment of what the agentic AI promises from NRF 2025 actually delivered, see NRF 2026: The AI Reckoning.