French e-commerce just crowned its startup of the year, and it’s betting on AI “agents” that actually do the work

A French startup called Lemrock just took top honors as “e-commerce startup of the year” in a major industry challenge, and the win signals where online retail in Europe thinks the next battle will be fought: AI that doesn’t just chat, but acts.

The award came at the 10th edition of the Start Me Up Challenge, run by Fevad, France’s main e-commerce trade group, alongside consulting and audit giant KPMG. This year’s through-line was “agentic e-commerce,” a fast-emerging approach where software agents handle more of the shopping journey, from advising customers to executing tasks like follow-ups, returns, and operational workflows.

A French industry contest turns into a bellwether for what retailers will buy next

Start Me Up has become a kind of proving ground for French and European commerce tech, less science fair, more “can this plug into a real retailer and move the numbers?” Organizers say they’re looking for products that can be industrialized and rolled out at scale, not flashy demos that collapse outside a lab environment.

That focus reflects the pressure e-commerce teams are feeling across Europe and the U.S.: paid acquisition costs keep climbing, delivery expectations are unforgiving, and competition increasingly comes down to personalization and execution. Retail leaders want tools that reduce friction, lift conversion, and cut support costs, without requiring a months-long rebuild of their tech stack.

KPMG’s involvement adds a hard-nosed filter: governance, compliance, and the ability to scale. In practice, that means startups are judged not only on what their AI can do, but on whether it can do it safely, predictably, and in a way that fits into the messy reality of enterprise systems.

What “agentic e-commerce” means, and why it’s suddenly everywhere

Generative AI in retail started with text: product descriptions, chatbots, recommendations. “Agentic” systems go further. They chain actions together, spot a need, compare options, propose an optimized cart, trigger a payment reminder, open a logistics ticket, or orchestrate a customer-service response, often semi-autonomously, under rules set by the business.

For merchants, the pitch is straightforward: move a chunk of repetitive, high-volume work from humans to supervised agents, then measure the impact in everyday metrics, faster handling times, higher first-contact resolution, fewer avoidable returns, cleaner product information.

But the risks are just as real. A hallucinated answer can become a broken promise. A pricing mistake can turn into a margin hit. A confident but wrong delivery commitment can ignite customer backlash. That’s why serious “agentic” products are increasingly built with guardrails: business rules, approvals for sensitive actions, permissioning, and audit logs that show exactly what the system did and why.

Why Lemrock won: turning AI into execution, not just conversation

Lemrock’s win reflects a shift in what retailers are rewarding. The new bar isn’t “does the AI sound smart?” It’s “can it take operational action, and can we trust it?”

Retailers evaluating agentic tools tend to demand two things at once: a smoother customer experience (consistent answers, relevant recommendations, seamless handoffs across channels) and tighter risk control (preventing hallucinations, pricing errors, or unsafe access to sensitive data). The startups gaining traction are the ones that treat safety and oversight as core product features, not afterthoughts.

Integration is another make-or-break issue. Merchants run on a patchwork of platforms, Shopify, Salesforce, homegrown systems, plus specialized product information management (PIM) and order management systems (OMS). Tools that can connect via APIs and prebuilt connectors, without a heavy rebuild, have a major advantage. If “agentic automation” takes months to deploy, much of its value evaporates.

There’s also a quieter organizational shift underway: marketing and e-commerce teams spend less time manually pushing buttons and more time defining rules, scenarios, and alert thresholds, then monitoring performance. In other words, the work moves from doing tasks to supervising systems.

Three startups recognized as retailers demand practical, near-term ROI

Organizers said three startups were recognized this year, underscoring how fragmented retailer needs have become. One company may be focused on customer support automation, another on product search, another on inventory, fraud, or pricing.

The common thread is that AI and automation are becoming modular, tools that can be deployed for specific, measurable use cases rather than sweeping “we do everything” promises. Retailers are increasingly shopping for narrow solutions that can be tested quickly, prove value on real data, and meet compliance requirements.

Agentic systems also force companies to answer uncomfortable governance questions: Who approves an automated refund? Who signs off on a price change? What logs are retained? How do you roll back a bad action? Startups that can’t offer clear controls often find that “autonomy” stays theoretical, because businesses won’t delegate without visibility and a kill switch.

The bigger picture: productivity gains, new risks, and a looming vendor-dependence fight

Agentic e-commerce is attractive because the productivity upside stacks. An agent that handles routine “where’s my order?” requests can reduce contact-center load. Agents that normalize catalog data can improve search and reduce returns. Agents that adjust merchandising based on seasonality and availability can lift conversion. At scale, those incremental wins can add up fast.

At the same time, compliance and privacy concerns are rising. Europe’s GDPR, its sweeping data-protection law, pushes companies to be explicit about data use, retention, and transparency when AI is involved. That pressure is likely to shape product design in ways American retailers will recognize as U.S. state privacy laws expand and enforcement tightens.

Then there’s vendor dependence. Many agentic tools rely on third-party AI models, which raises questions about variable costs, uptime, and changing terms of service. Retailers are increasingly drawn to hybrid setups: keep business rules in-house, plug in interchangeable AI components, and avoid rebuilding the whole system every time a model provider changes.

Lemrock’s win, and the contest’s focus, point to a market that’s moving past AI experimentation and toward operational AI, systems expected to perform under real-world constraints of margin, speed, and quality. The next phase won’t be about who has the cleverest demo. It’ll be about who can automate safely, integrate cleanly, and deliver results retailers can measure week after week.