Shopify, Zooplus and Otto show how retail AI is moving from model calls into operating systems

New European examples connect tailored language models, catalogue search and warehouse simulation with the systems retailers already run.

Retail AI is increasingly being assembled around a particular job: querying a merchant's store, retrieving the right product or testing a warehouse layout. NVIDIA's 21 October European retail update describes projects at Shopify, Zooplus and Otto Group that connect open models and simulation tools to those operating workflows.

The examples span online discovery and physical fulfilment. Their shared feature is the integration work between a general technology and the retailer's own information, systems and measures of success.

Shopify specialises a store-data subagent

Shopify has post-trained Nemotron 3.5 Lightning for a GraphQL subagent within Sidekick. The subagent retrieves store data for features including Sidekick Pulse, which examines a merchant's business and suggests improvements. NVIDIA describes Shopify using Megatron Bridge and NeMo RL in the adaptation process.

GraphQL provides a useful lens on the project. A merchant's natural-language question has to become a query against an actual store schema, and the returned records have to support the answer. A specialised component can focus on that translation while the broader assistant manages the conversation and presents the result.

The project's cost and performance figures come from Shopify's internal comparisons reported by NVIDIA. For another retailer, the relevant evaluation would use its own schema, permissions and representative queries. Product variants, refunds and inventory locations can make apparently simple business questions surprisingly specific.

Zooplus combines catalogue enrichment and semantic search

Zooplus worked with Infosys and NVIDIA technology on a product-discovery prototype inside its AWS environment. The described stack combines Nemotron Retriever with a retail catalogue-enrichment blueprint to create attributes that make products easier to find.

This addresses a different problem from generating a fluent shopping response. Search depends on what the catalogue actually records: ingredients, sizes, compatibility and other product properties. Enrichment can help convert inconsistent supplier material into searchable attributes, while retrieval connects a customer's question to candidate products.

That structure gives a retailer several places to inspect results. It can examine an extracted attribute, the ranked product list and the final answer separately. A successful change may come from better catalogue data rather than a larger language model, and the application can preserve that distinction when reporting its results.

Otto tests physical operations in a digital twin

Otto Group extended its autonomous-warehouse initiative, developed with Reply using Omniverse libraries on Google Cloud, to its largest fulfilment centre in Haldensleben. The update describes logistics teams testing and optimising operations in simulation before physical changes.

A warehouse adds time, space and equipment constraints to the information problem. A proposed workflow needs to account for where goods move, where people work and how different robot fleets share the building. A digital twin gives planners a place to test those interactions against a defined layout.

The result is a more varied retail technology stack than a single customer-facing chatbot. A language-model component may help query store data, retrieval may improve discovery, and simulation may support fulfilment planning. Each serves a distinct part of the retailer's business.

For technology leaders, the immediate lesson from these current projects is architectural: begin with the operating task and connect the appropriate model, data and interface to it. The October examples make that integration increasingly visible, from the first product query to the movement of an order through a real fulfilment centre.

Sources

- [NVIDIA Transforms European Retail, From Digital Discovery to Physical Operations](https://blogs.nvidia.com/blog/european-retail-open-models-robotics/) — NVIDIA