How far can a grocer take AI?

Before the first pull request written by an agent, people across Rohlik were using chatbots like ChatGPT and we were teaching colleagues how to get real value from them.

Today, Devin works as both an engineer and an on-demand data analyst, with permissioned access to Rohlik’s data and company context. Our own Rohlik Notetaker turns meetings into secure knowledge that authorised agents can use. We also have access to tens of NVIDIA B300 GPUs to train models that inspect food, pack bags and, soon, power robots.

We got here by moving one useful step at a time, making plenty of mistakes and building each missing layer when the work demanded it.

Our first job was to make people comfortable using AI.

We ran sessions showing colleagues how to use ChatGPT and Gemini in their actual work. Teams tried it for research, writing, analysis and repetitive tasks they understood better than any central AI team could.

As people used the tools, they began spotting their own opportunities and testing them instead of waiting for a company-wide AI plan.

Any company can start here: give people access, teach the basics and let them use AI on problems they know well. The useful cases appear quickly.

At Rohlik, this created demand for tools that could take responsibility for a full job.

In May 2025, we spent three hours at Cognition’s Silicon Valley office onboarding Devin. The following week, our engineers merged more than ten pull requests fully generated by it.

Coding was a good first proving ground. The queue was visible, the output could be tested and a bad result could be rolled back.

We kept going. By June 2026, Devin and Claude Code were present in 196 repositories. A 30-day GitLab scan classified 85% of contributions as agentic. Merged pull requests were up 38% month-over-month, while review wait was down 19% against the 90-day baseline.

Devin has since moved beyond coding. We now use it as an on-demand data analyst, connected to Rohlik’s data and company context. A colleague can ask a business question and let the agent investigate across data, code and internal knowledge instead of waiting for another report to be built.

The AI was now doing work with a clear owner, a measurable output and consequences when it got things wrong.

Company context quickly became the constraint.

Important knowledge lives in dashboards, Slack and Drive. A huge amount also lives in meetings, where teams explain why a decision was made, what changed, which customer raised a problem and which approach already failed.

We built Rohlik Notetaker ourselves to capture and structure that knowledge. It stays secure inside Rohlik and becomes available to authorised agents according to permissions. Onyx gives agents cited access to Slack and Drive. The Notetaker adds the spoken context that would otherwise disappear into personal notes or memory.

This gives agents continuity. Devin can understand the discussion behind a metric. An operational agent can find the decision that shaped a process. New work starts with what Rohlik already knows.

We then moved agents closer to customers and operations.

Maia searches products, builds baskets and handles common claims across Rohlik’s five markets. It can select a delivery slot and take payment with a saved card inside chat or voice. Our public MCP server gives compatible assistants access to the same catalogue, basket and order tools.

The mistakes became more expensive. Some agents stalled halfway through a process. Some produced a good answer when the business needed completed work. Others looked great in a demo and struggled with real permissions, exceptions and handovers.

Duvo.ai came out of those mistakes. It gave us an operating layer for agents that cross systems, preserve state, handle exceptions and verify the result. Duvo is one step in the wider Rohlik journey, helping turn early experiments into repeatable operations. It also wasn’t a smooth journey, but that’s the pattern in this story. You have to start early to learn and improve.

Every failure created useful data: where the process broke, when a human stepped in and what the next version needed to learn. We are now on a journey to truly autonomous operations.

In June 2025, we spun our grocery platform out as Veloq. At launch, it was running Rohlik across five countries, supporting €1.1 billion in annual revenue and more than 1.3 million orders a month. It now supports 2 million orders a month.

In October, we announced the rollout of 24 Sereact robots across Berlin and Vienna. Veloq and Sereact have since formed a global partnership combining warehouse software, AutoStore and AI-driven robotic picking.

A warehouse produces the evidence physical AI needs. Every pick has an image, product, location, action and outcome. Failed grips, damage, delays and human corrections show the model where it was wrong. Customer outcomes reveal whether the order was genuinely good.

Captured properly, millions of physical actions become training data.

Rohlik now has access to many tens of NVIDIA B300 GPUs. The compute matters because we can connect it to Rohlik’s operating data and train models against real grocery outcomes.

Quality control on every order. Vision and world models can reason over images and video. Rohlik data can teach them to spot damaged packaging, poor fresh produce, wrong products and incomplete bags before an order leaves the warehouse.

Vision-language-action models can be adapted using demonstrations and corrections from experienced packers. A grocery bag can contain soft fruit, glass bottles, frozen products and thousands of awkward combinations. Packing it well requires judgment.

Humanoid robots that can move between tasks in spaces designed for people are the next stop. This comes later, after narrower picking, quality and packing systems have proved themselves in production.

The learning loop can outlast any one model: run real work, record the outcome, capture the correction, evaluate the next model and deploy it when it performs better.

Rohlik has already changed how people buy groceries. Across five European markets, a household can order a full weekly basket, including genuinely fresh food, and have it delivered quickly, at a precise time, with no delivery fee and at supermarket prices.

The next opportunity is much larger than improving the efficiency of today’s operation.

Maia, our voice-first agentic shopping assistant, can understand what a household needs and build the basket. Agents can plan purchasing, availability and delivery around real demand. Vision models can check quality continuously. Robots can pick and pack more of the order. Each delivery can improve the system that creates the next one.

This can make grocery more personal, affordable and less wasteful. Veloq can take the operating model to other grocers and markets, multiplying the impact beyond Rohlik.

Then we can look beyond groceries to prepared food. Today, preparation is the expensive part because labour costs much more than the ingredients. That is why people buy groceries and cook at home. But what if prepared food cost the same as groceries do today? Or delivered food became cheaper than food from a supermarket? People would order more of it, and online’s share of food would go way up.

Rohlik shows that customers change how they buy food when the experience becomes dramatically better. AI and robotics now give us a chance to change how the entire system works behind that experience.

Most of the value is still ahead of us: rebuilding how food moves from base ingredient producers to households.

Most of the value is still ahead of us: rebuilding how food moves from base ingredient producers to households.