Everyone is running AI pilots. Not everyone is moving the P&L. That tension, between the volume of AI experimentation in retail and the relative scarcity of proof that it's actually doing anything, was the starting point for our recent PI Spotlight.
Together, our panel covered a fascinating hour of territory: frontline product knowledge tools to agentic commerce, token costs to sovereign AI, and Nordstrom's clienteling model to synthetic influencers in China selling $7.6M of product in a single afternoon.
🎥 Watch the full discussion below & read on for the full write-up.
The P&L Test: What Actually Makes the Cut
The AI use cases genuinely moving the P&L today tend to be the unglamorous ones. Demand forecasting. Smarter inventory allocation. Better search and recommendations. Not because these are easy, but because they map directly to commercial levers the business already understands and can measure. The ones stuck in pilot purgatory share a common trait: no clear decision they're improving, and no one accountable for acting on what they surface.
The same pattern shows up on the store floor. Associates selling technically complex products often can't match the depth of a customer who's already done their research. AI that functions as on-demand product expertise closes that gap in real time, without months of training. Operationally, platforms are already giving store managers daily, actionable nudges based on heat map and conversion data, without requiring them to interrogate a dashboard.
The through-line, wherever AI is working, is ownership.
"Every AI use case needs a commercial owner, a clear decision that it's going to improve, and a metric that finance will recognise as real value."
— Kimberly Pringle, Second Self RetailOS
What separates organisations that escape pilot purgatory from those that don't is cross-functional alignment across the full product lifecycle. Technology is not just a cost line. It is the layer that sits beneath every workflow disconnect in the business, and until leadership treats it that way, isolated pilots will keep producing isolated results.
"It's actually the layer that is secondary to all of the other workflow disconnects in the process."
— Meg Ball, Centric Brands
Pilot Purgatory, Token Costs, and the K-Shaped Workforce
Most organisations are still deploying point solutions: a chatbot here, a PDF processor there. But the harder question, what enterprise-wide AI actually looks like, is one most have not yet seriously asked. Until AI becomes genuinely ubiquitous within an organisation, the way the internet became ubiquitous, it will keep feeling like something being tried rather than something being used.
The costs of getting there are being systematically underestimated. Token budgets are real, already biting, and about to become a boardroom conversation in organisations that haven't had it yet.
"It's actually more expensive to use AI for some things than it is to use humans."
— Kimberly Pringle, Second Self RetailOS
One example that surfaced during the conversation: an undisclosed company spent half a billion dollars on tokens in a single month. As token costs become visible and budget owners start managing them explicitly, AI access inside organisations is likely to be rationed. That rationing has a shape the industry hasn't fully reckoned with yet.
"We're going to have a K-shaped workforce — people in our organisations that have access to AI, and people that don't."
— Barry McGeough, AmeriCo. Group
When AI tools are allocated only to roles most directly tied to revenue, everyone else falls further behind. Amy Webb's framing of the "tech-enabled human," the idea that AI gives some people superhuman capabilities, cuts both ways.
The inverse is equally real. And the implications extend well beyond the retail floor: whoever reaches artificial general intelligence first, Barry argued, is likely to own the conditions of the technology landscape for the next century. China currently has the cheapest tokens. Amazon is already suggesting DeepSeek as a cost alternative to ChatGPT.
The sovereign AI question, which sounds abstract, is in practice a question about which infrastructure your business is being built on, and who controls it. And it sits alongside a more immediate version of the same problem that every organisation is already navigating: which model do you use, and why? Barry's own preference is Perplexity over ChatGPT, finding it more fact-based and research-grounded. But the broader point is that businesses are now being asked to make consequential decisions about AI infrastructure, ethics, and cost at a pace no governance framework has caught up with.
"We weren't even having these conversations a year ago. And they're being forced on us now."
— Barry McGeough, AmeriCo. Group
Guardrails: The Compliance Layer No One Has Fully Figured Out
Guardrails are not a one-size-fits-all exercise. What works at Walmart and what works at a specialty boutique are different instruments. And in every case, what makes them function is not the technology itself but human ownership: a leader who takes responsibility for the system, not just the rollout.
"No one at CommerceNext said AI in a sentence that didn't also say you needed great leadership, associates, and talent alongside that AI to bring out the best of it."
— Paula Angelucci, WH Smith North America
Most organisations are still thinking about guardrails at the brand level. The regulatory reality is operating at a completely different scale. New York's RAISE Act already requires synthetic performers to be disclosed on any webpage where they appear. Bills in California address LLM citation requirements and name, image, and likeness protections. Gartner is recommending that corporations proactively include NIL clauses in employee contracts now, as a legal safeguard for future AI deployments. These are not hypotheticals. They are current compliance realities most retail organisations have not fully mapped.
Underlying all of it is what Barry described as automation bias: the tendency for humans to trust outputs from authority-projecting systems more than equivalent outputs from a human source.
"ChatGPT is getting a billion queries a day and is responsible for almost two billion dollars of commerce last year. We are using it to make decisions. What are the guardrails?"
— Barry McGeough, AmeriCo. Group
When AI-generated content crosses the uncanny valley, the threshold at which it becomes indistinguishable from real, the question of disclosure stops being an ethical nicety and starts being legal exposure. AI-generated product imagery is already appearing on PDP pages. AI-generated models are increasingly credible. The line is not approaching. For many brands, it has already arrived, and the legal and reputational exposure that comes with it is not theoretical.
Keeping the Brand Voice Alive When AI Is Doing the Talking
Maintaining distinct brand identity when AI is generating content at scale is less a technology problem than a people problem. The brand knowledge that currently lives in a creative director's head, or in a merchant's instinct about a category, is exactly what AI systems need as structured input. If it isn't being fed in deliberately, the output defaults to something generic. Brand differentiation through AI is not automatic. It requires someone with genuine brand authority to take ownership of seeding the system correctly.
"This doesn't start with technology. It starts with the people. Leaders need to take ownership of their brand lanes and apply the knowledge they've been building for years into the data structures that feed the loop back."
— Meg Ball, Centric Brands
Getting teams genuinely bought in requires more than a top-down mandate. One example that resonated: the CEO of EyeBuyDirect introduced AI to her team not through company use cases first, but through personal ones. A leader who hosted regular dinner parties started using AI to plan menus based on what he'd cooked before. His engagement with AI tools at work skyrocketed. The lesson transfers directly: when people experience the value of a tool in their own lives, professional adoption follows without resistance.
"When you add value to someone professionally and personally, you are going to get rapid-fire buy-in."
— Paula Angelucci, WH Smith North America
Getting Personalisation Right at Scale
The best examples of personalisation at scale share a quality that is harder to achieve than it sounds: the technology disappears. Ulta's loyalty programme works because the customer behaviour is identical regardless of channel. Scanning for offers in the app feels the same whether you're at home or standing in the aisle. The AI doesn't announce itself. It just knows.
Nordstrom's clienteling model illustrates a different dimension. An associate who can see what's in a customer's cart before a word is spoken, and can offer to have it pulled, delivered, or shipped without the customer repeating any context, is demonstrating that the AI is invisible, and the associate is more capable because of it.
The structural challenge underneath both examples is that most AI personalisation is still being deployed as a top-down mandate rather than a systems-level change. A directive to use AI, a few approved tools, and a bottom-up hope that it gets executed is not a personalisation strategy. What's needed is something harder: treating the organisation itself as a system with goals, values, constraints, and institutional knowledge, and building AI deployment around that reality rather than around the tools available.
"Companies really are systems. They have goals, values, constraints, personality quirks depending on the leaders. If you're able to convert all of that into institutional knowledge that becomes clearer decision systems, then you can actually educate your company on the best way to use AI in alignment with what it stands for."
— Kimberly Pringle, Second Self RetailOS
When the Agent Decides, Brand Loyalty Doesn't Factor
If a consumer deploys an AI agent to find and buy a pink puffer jacket, the agent will match on price, parameters, and availability. It will not factor in brand affinity, aesthetic loyalty, or the emotional relationship a customer has spent years building with a label.
"If agents are making decisions, they're not saying 'I like The North Face better than Columbia.' They're not saying that."
— Barry McGeough, AmeriCo. Group
This isn't a future risk. It's a structural shift that is already beginning to materialise, and it puts brand preference, the thing retail has invested billions in building, directly in the firing line.
The response brands are reaching for takes two distinct shapes. One is experience: fitness classes, run clubs built around personal data aggregated across devices, community ecosystems. These are investments in the part of the relationship an agent cannot replicate. An agent can find you a running shoe. It cannot give you the run.
The other is the opposite: leaning fully into the synthetic. Two AI-generated influencers in China, visibly synthetic, ran a live selling event for sixteen hours. 17,000 items sold. $7.6 million in revenue. The audience knew they were synthetic and thought it was cool.
The authentic and the synthetic are not mutually exclusive strategies. Brands may find themselves pursuing both, and the ones that do so deliberately will be better positioned than those that drift into one or the other by default.
The Consumer Has Already Moved
The consumer is no longer a passive recipient of whatever retail decides to deploy. Gen Z in particular enters a store having already done more research than most associates have at their fingertips. They know the product. They know the trend. They may know the supply chain. What they're waiting to find out is whether the person in front of them can match their energy.
"The gap is: they've done the research, they're walking in the store, and they're expecting you to be the subject matter expert. And they're expecting an incredibly elevated customer experience."
— Paula Angelucci, WH Smith North America
The elevated in-store experience was one of the defining themes of NRF this year. The brick and mortar store is back, but it has to be better than the phone. AI tools that make associates smarter and more capable in front of that customer are not a future-state aspiration. They are a response to an expectation that already exists and is already going unmet in most stores.
The Bottleneck Is Not the Technology
The tools exist. The data exists. And in most retail organisations of meaningful scale, the technology is no longer the limiting factor. What limits progress is the organisational discipline to deploy it with clear ownership and measurable outcomes, the governance structures to act on what it surfaces, and the cross-functional alignment to stop teams optimising independently at the expense of the shared customer experience.
Underneath all of that is something simpler: the willingness of leaders to be publicly uncertain, to explore AI in front of their teams, to fail visibly, and to model the kind of open experimentation that real adoption requires.
"Leaders need to be open to failing publicly, because the tech is moving so quickly. You need to give your teams the time and the opportunity to explore."
— Meg Ball, Centric Brands
AI in retail is no longer hype. But it is not fully real yet, at least not at the scale or consistency that the investment and the noise would suggest. What it is, right now, is a genuine inflection point.
The organisations building ownership structures, governance models, and institutional knowledge pipelines today are the ones who will be able to move when the technology matures. The ones still running isolated pilots and chasing flashy demos will find, when the dust settles, that they have fallen far behind and are watching the next era of retail from the outside.
We will be continuing this conversation at Retail Rebels NYC on October 20th. Find out more and register your place at the links below.
