The use of Large Language Models (LLMs) in eCommerce is still in its infancy. Many companies are already experimenting with chatbots, automated product recommendations, or smart search – but in practice, the limitations quickly become apparent: New products appear that a model doesn't yet know about. Customers ask questions in many different ways that the system isn't prepared for. And for nearly every task – from recommendation to sentiment analysis – specialized models are needed. The result: complexity, high costs, and little flexibility.

This is exactly where a new research project comes in: eCeLLM by Bo Peng and colleagues.

"We aim to build a generalist model that can effectively generalize across diverse e-commerce tasks by leveraging large-scale, high-quality instruction data."— Bo Peng et al., 2024

The goal of the research project is to train LLMs so they better address the real requirements of retail. Instead of deploying one model for product recognition, another for attribute extraction, and a third for customer inquiries, eCeLLM pursues a generalist approach. This becomes possible through a specially developed dataset – ECInstruct – which contains tens of thousands of carefully curated instructions for ten central eCommerce tasks.

The results are remarkable: In tests, eCeLLM significantly outperforms established models like GPT-4 or specialized state-of-the-art approaches – on average by more than ten percent. Particularly exciting is that the models remain robust even when confronted with previously unknown products or new formulations of user queries. For dynamic product ranges and a diverse customer base, this is a decisive advantage.

Figure 1 – Overall schema of the eCeLLM instruction optimized with ECInstruct (Peng et al., 2024)

For eCommerce companies, this means two things. On one hand, fragmentation into many small AI components could soon be a thing of the past. A well-trained cross-task eCommerce LLM can handle multiple tasks simultaneously – and saves resources in development and operation. On the other hand, time-to-market shortens because new products don't need to be laboriously "trained into" the model. Customers will notice this through more flexible responses, more precise recommendations, and overall better customer experience.

Of course, eCeLLM remains a research result for now. But the project also clearly shows where the open challenges lie: User profiles and context data are hardly integrated so far, explanations for decisions ("Why this product?") are missing, and the transfer from laboratory conditions to the productive shop has not yet been solved.

Our assessment at foobar Agency

For us, eCeLLM is above all a signal: Research in the field of AI is noticeably moving toward real practical relevance. It's no longer just about impressive demos, but about models that can actually make a difference in dynamic, complex retail environments.

In many companies today, specialized (AI) systems work side by side: one for product recommendations, one for search, one for support. This is exactly what the composable commerce approach stands for, which we pursue at foobar Agency: independent but interconnected components that together create a strong overall picture.

What eCeLLM shows in this context is the next stage of development. AI models no longer need to be trained in isolation for individual tasks, but can use knowledge across boundaries. In a modular, open architecture, this can be perfectly combined: specialized tools remain in place, but are complemented by cross-task eCommerce LLMs that understand connections and make data flows more intelligent.

However, for companies to leverage these potentials, they need more than just a powerful model. They need a clean data foundation, clear interfaces, and a flexible system architecture. This is exactly where we come in at foobar Agency: We create the technical and organizational prerequisites so that new technologies can be seamlessly integrated. Our experience shows that composable approaches in particular are crucial – they give companies the freedom to integrate new models like eCeLLM as soon as they're market-ready, without being tied to monolithic systems.

In short: eCeLLM is a glimpse into the future, but the groundwork is being laid today. Those who get their data, systems, and processes into the right shape now will benefit faster and more securely from such developments in the coming years – and will clearly set themselves apart in competition.

Dominik Thalmeier, AI Architect at the agency foobar, summarizes:

"eCeLLM is not a finished product, but a research impulse. But it clearly shows that the future in eCommerce lies in robust, generalist AI systems – and that companies that invest in composability and data quality today will be among the winners tomorrow."
— Dominik Thalmeier

Do you want to know how your company can lay the groundwork today to successfully use AI like eCeLLM in the future?

👉 Talk to us – we guide you from strategy to implementation: Get in touch

‍

-----------

‍

FAQ: AI & eCeLLM in eCommerce

How do classical LLMs differ from cross-task eCommerce models like eCeLLM?

Classical LLMs are trained on general knowledge and quickly hit limits in retail. eCeLLM, on the other hand, was instructed with real eCommerce data and can specifically address tasks like product recognition, recommendation, or question answering.

Why are clean product data and composability critical to successfully implementing AI in eCommerce?

Only with consistent, high-quality data can models like eCeLLM work reliably. Composability additionally ensures that new technologies can be flexibly integrated – without dependence on monolithic systems.

Can a single model really handle multiple tasks like recommendation, search, and customer service simultaneously?

Research results show that cross-task eCommerce models can share knowledge between different areas – and thus often deliver better results than specialized individual models. We're excited to see further results and practical studies.

What benefits does eCeLLM bring for retailers with large, dynamic product ranges?

The model is more robust against new products and formulations. Retailers can thus react faster, provide more precise recommendations, and improve the overall customer experience.

When does it make sense for companies to invest in AI-powered eCommerce solutions like eCeLLM?

eCeLLM is still a research project. However, companies should begin preparing their data foundation and system architecture now. Those who invest in data quality and flexible architectures now can use new models immediately once they're market-ready.