An evening at Nobu Barcelona with senior retail leaders, hosted by duvo.ai
Everyone at Shoptalk was talking about AI this week. Every stand, every stage, every coffee queue. So the question we built our evening around was deliberately simple: so what?
What is genuinely changing inside retail businesses? What looks impressive in a demo and dies in the operating engine? And where is AI quietly solving the problems that have been hiding in plain sight for years?
On Tuesday evening, away from the conference floor, we brought a small group of senior retail, digital, commercial and operations leaders up to the 22nd floor of Nobu Barcelona to have that conversation properly. A carefully curated room of leaders from some of Europe's biggest retailers and consumer brands; off the record, with cocktails and Nobu small plates as the sun went down over the city.
The conversation that followed was one of the most honest we've heard on this topic anywhere.
The human API problem
The fireside discussion was led by Giles Smith, senior technology operator and advisor to brands including Burberry, Unilever, Selfridges and John Lewis, in conversation with Tomáš Čupr, founder and Group CEO of Rohlik Group and co-founder and CEO of duvo.ai.
Tomáš built Rohlik into one of Europe's fastest-growing e-grocery retailers, operating across five countries. But the moment that set up the whole evening was what he found when he looked inside his own head office.
"We were paying highly skilled people to be human APIs. Connecting Excel files to emails, to SAP, to different portals, all day long. Instead of finding an amazing assortment, they were the connector between systems that just don't talk to each other."
Every leader in the room recognised it instantly. Processes that live in someone's head. Decisions held together by manual work and constant coordination. Not because anyone failed to fix it, but because nobody ever thought it could be fixed at scale.
Unlimited humans
The idea that landed hardest was Tomáš's way of framing what AI actually changes.
"AI can think, it can click, it can type, it can talk. If I'm oversimplifying: AI can give you unlimited humans to do whatever you want in the company."
And the examples were anything but theoretical. Take supplier price increases. At Rohlik's scale, thousands of price proposal files arrive every month. No buying team on earth researches every SKU in every one of them, so the negotiation becomes a shortcut: the supplier asks for seven percent, the buyer counters with five, everyone moves on. With unlimited capacity, every single increase gets researched and challenged against commodity prices, labour costs and an acceptable supplier margin, every time. That's margin leakage, stopped.
Or freight. Five to twenty percent of freight invoices contain errors, against rate cards that run seventeen bullet points deep. Checking every invoice by hand takes hours, so most never get checked. Duvo is currently running a proof of concept on exactly this for a major US brand, where the problem is worth tens of millions a year. Nobody knew, because nobody could see it.
As Giles put it, this isn't a story about replacing people:
"What this is actually about is doing things we couldn't do before, doing them better, or doing them faster. That's not the same as losing jobs. Nobody could do this work anyway."
Five versions of the truth
The part of the conversation that sparked the most discussion around the room was about why automation projects stall in the first place. It's rarely the technology. It's that nobody actually knows how their processes run.
"You speak to five different operators and they give you five different versions of reality," Tomáš said. "This is super exciting, but you need to speak to Mary, she knows how the process actually runs. Then you speak to Mary, and Mary says, well, I think John knows the rest."
The traditional answer is to pay a consultancy seven figures to interview everyone and put it on slides, then pay someone else to turn the slides into automation. Duvo's answer is to send AI agents to do the discovery instead: operators share their screen, talk through how they actually work, and within half an hour the process is mapped, exceptions, contradictions and all. Then the same context powers the automation.
The questions from the room pushed this into fashion, store operations, returns and creative production, and the same pattern kept emerging. The gap between how a process is documented and how it actually runs is where the margin leaks. Almost nobody can see it.
Start with something painful
If there was one practical takeaway leaders carried up to the rooftop afterwards, it was Tomáš's advice on how to actually begin, without disappearing into a two-year transformation programme.
"Give us something painful, fairly isolated. We prove it works, and then it becomes a very easy conversation about larger-scale deployment. If you solve something really painful in one market, suddenly you're getting benefits in six weeks instead of two years from now."
The discussion could easily have run another hour, but Nobu's rooftop was waiting, and some conversations are better continued over a drink with the city lights coming on below.
Thank you to everyone who joined us, and to Giles for steering the evening with exactly the right mix of curiosity and healthy scepticism. The appetite for honest, closed-door conversations about what's actually working in retail AI is clearly there, and this won't be the last one.
Want to know where margin is quietly leaking in your business? Duvo uses AI to map the real process in days. See where the value sits. Fix it, then run the better version every day. Get in touch with the team.
Frequently Asked Questions
Q: What was the Duvo event at Shoptalk Barcelona 2026 about?
Duvo hosted an executive dinner at Nobu Barcelona during Shoptalk Europe 2026 on the theme "Beyond the Hype: Where AI Is Actually Transforming Retail." The evening brought senior retail, digital and operations leaders together for a closed-door conversation about where AI is quietly solving real operational problems, led by Giles Smith and Tomáš Čupr.
Q: What is the human API problem in retail operations?
The human API problem describes highly skilled employees spending their time connecting disconnected systems — copying data between Excel, email, SAP and supplier portals — instead of doing the strategic work they were hired for. It happens because enterprise systems do not talk to each other, so people become the integration layer.
Q: How much money do freight invoice errors cost retailers?
Between five and twenty percent of freight invoices contain billing errors against complex rate cards. Because most companies lack the capacity to audit every invoice, these errors go undetected. For large retailers and brands, the cost runs into tens of millions per year in overpayments that nobody sees.
Q: Why do most retail process automation projects fail?
Most automation projects stall because nobody actually knows how the process runs. Teams document what they think happens, but the real process lives in screens, emails, spreadsheets and handoffs that no system of record captures. Automating the documented version rather than the real version produces automation that breaks on contact with reality.
Q: How can retailers start with AI automation without a multi-year transformation?
Start with one painful, fairly isolated process. Prove it works in six weeks, then expand. This avoids the classic trap of a two-year transformation programme that loses momentum before delivering value. Scoped pilots that deliver measurable results create the business case for broader deployment.