Omar opened the day not with a vision statement but with data. AI-driven web traffic has tripled year-over-year. The share coming from autonomous agents has grown by 7,851% in a single month. And 63% of all autonomous traffic is currently headed toward retail and e-commerce.

He then shifted to the supply side, describing how commerce has become structurally more complex over the past decade — retailers and brands now operate across stores, marketplaces, social, and increasingly LLMs simultaneously. His thesis: agents are becoming the last-mile connection that removes that friction, in the same way fiber upgraded the internet's copper-wire bottleneck.

The panel opened with a genuine debate: is AI-driven discovery genuinely new, or just the latest version of a familiar story? Rob Garf argued for continuity — the buy button has moved upstream before, from stores to e-commerce to mobile to social. Chris Silver took a different angle: what makes this moment distinct is the character of the new customer. AI agents are extremely demanding, want structured factual information rather than aspirational copy, and penalize brands that can't provide it.

Jarid Lukin brought the practical reality check: the traffic numbers are dramatic in percentage terms but small in absolute terms. The work to be done now is preparation, not reaction. His team at Mars had built what he called AEO/GEO — treating it as SEO 2.0, including a new Reddit strategy, since Reddit has become a significant feeder of LLM training data.

Chris Silver described Just Food for Dogs' approach in detail: scrubbing the web for every product review, pulling in vet citations and health claims, mapping HTS codes to establish provenance, and syndicating all of it to all major LLMs simultaneously. His framing: it is like SEO but the iceberg goes much deeper, and almost no one is actually doing it yet.

Three very different business models produced three very different answers to the same question. One retailer described growth as coming primarily from the higher-end consumer, with the value segment largely absent from the market. That team uses real-time AI to determine promotional depth after just four clicks — resulting in hundreds of basis points of margin improvement — while leapfrogging traditional marketplace investment to focus directly on LLM surfaces and data supply chains.

Hemang Upadhyay from LG described marketplaces and owned channels as complementary rather than competitive: marketplaces attract and introduce; the brand website converts and retains. His team's focus has shifted from traffic acquisition to content quality and personalization, with the insight that conversion rate gaps usually signal a content problem, not a traffic problem.

Joshua Rockoff brought the most unusual perspective: relaunching four formerly bankrupt retail brands with no physical stores, fighting against years of content declaring the businesses dead. His discovery that 1996-era SEO tactics are delivering results today was one of the day's more memorable observations. His team is 61% AI-operated in core workflows including customer service.

The session opened with Christine Russo's organizing premise: LLMs are the new storefront, and the rules of visibility have fundamentally changed. You are no longer optimizing for a human typing into a search bar. You are optimizing for a machine deciding what to surface to whom and why.

Dan Cook offered a nuanced take on democratization: LLMs do lower the barrier to entry for smaller brands, because they reward precise structured data over raw budget. But visibility isn't fully democratized, because trust, reviews, and credibility signals still matter. The difference is that AI typically surfaces one to three options, not ten. If you're not in that narrow set, you simply don't exist.

The conversation turned to the art of browsing, and whether AI-driven discovery kills it. Ken Pilot argued that two ends of discovery will coexist: highly articulate prompting for people who know exactly what they want, and occasion-based navigation for everyone else. E-commerce sites that create genuine browse experiences will hold their value. He also called out the store associate opportunity: every associate should have expert-level product knowledge available in the moment, and almost none do.

Gary Haas opened with a counterintuitive argument: AI readiness is not a state you can achieve before you start. It is a state you reach by doing. Companies that use "not being ready" as a reason to wait will never be ready. His practical advice: empower the people closest to business problems to experiment, fail, learn, and hand off what works to IT.

Ravikanth Raparla introduced a framework from his work at David's Bridal, where the team built the world's first AI wedding planner covering more than 300 decision steps. He identified three non-negotiable elements for AI success: a strong data foundation, an organizational structure capable of acting on AI recommendations, and a feedback loop that allows models to improve over time.

Gireesh Sahukar addressed the widely-cited MIT statistic that 95% of AI projects fail, clarifying its origin: those projects were never intended to reach production. They were experiments. Experimentation failing is expected. The question is whether the organization has the process to take experiments that work and move them into production use cases that deliver measurable value.

Ashye Marcus opened with a candid framing: we are in a period of agentic fog. Anyone who tells you they have a definitive, clear view of where this is going should be treated with skepticism. What Stripe does have is direct visibility into the infrastructure being built by the AI companies themselves — which gives the session a different kind of credibility: not prediction, but evidence.

She described the pace of change in retail as unlike anything she had seen across previous shifts to digital, mobile, or social. Each of those gave the industry years to adapt. This one is not. The brands that leaned in early to digital are the ones with 40–50% digital penetration today. The same pattern will play out with agentic commerce.

Bobby Jones walked through Stripe's Agentic Commerce Suite: a one-to-many connector that lets a merchant integrate once and then toggle on access to a network of AI agents including OpenAI, Gemini, Meta, and Copilot. He gave the example of Urban Outfitters, which started with a ChatGPT integration and is now live with Meta and Gemini with zero additional engineering effort. A real-world product catalog insight emerged: having an item listed as "denim" rather than "jeans" meant missing every search for wide-leg jeans entirely.

Janet Sherlock brought the organizational lens that was largely absent from the morning sessions. She opened by naming the structural reasons enterprise AI fails: unrealistic expectations, data fragmentation, conflicting department priorities, and — most critically — poor organizational clarity. When everyone is experimenting in their own direction with no coordination, the collective effort moves sideways rather than forward.

She introduced her CFD model: a center of enablement (centralized in IT, responsible for data, platforms, integration, and governance), federated data science (core expertise living within each business area), and democratized data and insights. Her argument against creating a Chief AI Officer was pointed: in most organizations, it creates a power center that impedes rather than enables the federated model that actually works.

Drawing on her experience building Ask Ralph at Ralph Lauren, she identified the practical risks most companies are not thinking about: new channel complexity, ownership friction (who actually owns agentic commerce?), the dual workload dilemma of running existing operations while building new capabilities, and machine-readable content gaps that are larger than most organizations realize.

Jason Franklin delivered the day's most self-critical — and therefore most credible — testimony. He described American Furniture Warehouse's AI journey with unusual honesty: the company built impressive-looking AI experiences before the data foundation was ready, saw real metric improvements in conversion and purchases, and then watched the operational infrastructure buckle because the underlying data was wrong. They outsold what they were capable of producing. Logistics dependency on humans increased rather than decreased.

His definition of AI readiness: an operating foundation strong enough for AI to make trusted recommendations across product data, inventory, identity, governance, security, compliance, and integration. Not an AI strategy. An operating foundation. The strategy is built on top of that.

He also described building trust with machines: training AI on company values, customer service transcripts, and operational context so that when an agent is asked whether American Furniture Warehouse will stand behind its warranty, the answer is grounded in truth rather than inference.

Sue McMahon introduced Microsoft's framework for the companies pulling ahead in AI adoption: the frontier firm. These are organizations that structure themselves around on-demand intelligence, powered by humans and the AI they supervise, in a way that genuinely augments and extends the intelligence already in the organization.

She described AI adoption as a continuum across three phases: human with an assistant (tools like Copilot or Claude used by individuals), human-led agents (specialized agents embedded in workflows and back-end systems), and autonomous agents (operating within guardrails to take real action and eliminate repetitive tasks). Different parts of an organization will be at different stages simultaneously. The goal is not to reach phase three everywhere. It is to keep moving.

She also highlighted a breakthrough that is still underappreciated: unstructured data is now usable. Institutional knowledge trapped in PDFs, SharePoint folders, and image files can now be brought into AI workflows without needing to be structured first. And she closed with a challenge: she has worked with retailers for 30+ years and has never once heard someone say their data is in great shape. Waiting for perfect data before starting means sacrificing where you could already be good or great.

The conversation opened with a sharp contrast: a Business of Fashion headline calling AI "a thief of everything," set against the World Retail Congress consensus that AI is the doorway to a creativity renaissance. Sharonda Weatherspoon landed closer to the optimistic end, but with an important qualifier: the benefit depends entirely on whether organizations invest in enabling the people who do the work, not just the technology.

Her most consistent theme was the frontline workforce. She described her frustration attending conferences where every AI conversation is about digital and no one talks about store associates. Her argument: the customer who makes the deliberate choice to visit a physical store deserves an exceptional experience. AI should be making associates smarter and more confident — not leaving them with less knowledge than the customer who just spent an hour researching the product on their phone.

She described Michael Kors's approach in detail: a company-wide hackathon using Microsoft tools where the HR team — not engineering, not finance — won by building an AI agent to help district managers answer employment law questions. The insight: the teams with the least analytical infrastructure often have the most to gain from AI, because they have been doing the most manual work. She is now taking the same hackathon format to 400 store managers at the August conference.

Vishal Joshi offered the day's most concrete demonstration of what an agentic product actually looks like at scale. Joy processes 600,000 events per year across multiple markets, with double-digit market share in every market it has entered — and spends nothing on marketing. The platform moved from a tool (which expects users to know how to use it) to an agent (which understands the user and returns solutions).

The distinction between tool and agent is the key conceptual contribution of the session. A tool requires the user to do the cognitive work of breaking down a need into steps. An agent receives the intent, performs the decomposition internally, and presents a recommendation. Joy's AI event planner takes a request like "plan a 500-person event in this city for this budget" and returns venue recommendations, hotel selections ranked by proximity, schedules, and even draft RFPs — all without the user touching a search bar.

He described the trust model Joy uses to manage the anxiety that comes with AI-driven decisions at high stakes: consent (the user explicitly authorizes the agent to act), context (the agent builds a deep model of the user's needs), and control (the agent presents its reasoning and returns final decisions to the human). The system can draft a contract and prepare to send it. It stops short of sending without the user's click. His closing advice for retailers: stop optimizing for flashy AI experiences visible to boards and investors, and focus on eliminating pain in back-end operations.