How Lidl built a 92% accurate AI model to transform customer service

CX Network's key insights:

• Lidl is using AI and proprietary large language models to handle complex grocery-specific queries, improving both efficiency and service quality.
• Their customer service faces 150 distinct query types, and multi-intent interactions remain a key challenge for automation.
• An agentic architecure coordinates multiple AI "specialist" agents while human staff focus on empathy and complex cases.
• Generative AI is helping transform contact centers from cost-focused operations into value-generating, experience-driven platforms.

Don't miss any news, updates or insider tips from CX Network by getting them delivered to your inbox. Sign up to our newsletter and join our community of experts. 

Learn More

At CCW Europe 2025, Lidl delivered one of the most technically revealing and operationally honest sessions of the week. The global retailer laid out a challenge that many in the grocery sector will instantly recognize: extremely high query volume, an enormous variety of topics, and limited room for error.

Yet, Lidl’s ambition goes well beyond efficiency. “We are breaking the paradigm that separates quality from cost reduction,” said Fabian Quast, product owner at Lidl. Its strategy is not to trade one for the other, but to improve simultaneously through an AI-first model of customer service.

At the center of this transformation is Lia – described, quite deliberately, as “the face of their customer service.” Lia represents Lidl’s automation engine, powered increasingly by large language models (LLMs) and Lidl’s own proprietary artificial intelligence (AI) system, Optimus.

A uniquely complex customer support universe

Customer service in grocery retail is unlike in most other sectors. During the session, Quast revealed that the company's customer contacts span 150 distinct query types – everything from product quality and availability to coupons, deliveries, substitutions, recipes, bottle-return machines, online shop issues, and dozens more.

They stressed how crucial it is for the audience to understand this: complexity is not peripheral to Lidl’s customer service; it is its defining characteristic! Many categories sound similar, but require completely different actions. A simple example: coupon complaints. “Coupons unavailable,” “coupons expired,” and “coupons didn’t apply” may sound related, but in Lidl’s world, they are “three completely different cases needing three different solutions,” explained Quast.

This is what makes automation difficult. If the system interprets the customer’s intent even slightly incorrectly, everything that follows (e.g. routing, resolution, and customer satisfaction) falls apart.

The accuracy challenge: from NLU to LLMs

Lidl began by relying on classical NLU technology, which they openly described as hitting 70 percent accuracy.

The breakthrough came when Lidl introduced an untrained LLM. This second stage pushed accuracy to 80 percent, and the team described this as the moment they realized LLMs could fundamentally change their automation ceiling.

But the biggest leap came from Stage 3: taking Google Gemini and training it using Lidl’s own customer messages, its own terminology, its own product categories and the peculiarities of grocery-specific language. It called the resulting model Optimus.

Optimus was trained on thousands of Lidl customer messages and designed to master Lidl-specific classification challenges: knowing the difference between food and non-food items, interpreting product names that exist only within Lidl’s internal taxonomy, understanding that pet food and diapers – counterintuitively – both sit within the “food” category in their system! “If you don’t know what category the customer is talking about then you can’t automate the process.”

Remaining challenges: Multi-intent customers and flow limitations

Even with 92 percent accuracy, Lidl is transparent that its transformation is ongoing. It pointed to two major limitations that are still being solved.

First, Optimus only acts at the start of the conversation. It categorizes the initial intent but then hands off to deterministic flows. This means the system loses the dynamic, generative flexibility that an LLM could carry throughout the conversation.

Second, the system currently struggles with multi-intent queries. These involve, for example, incidents where customers bring several issues at once. They gave a simple example of a conversation containing three parts: partial delivery, product return, and delivery status. Optimus, in its current state, will only pick one. “Multi-intents are currently not supported,” explained Quast.

Becoming a fully agentic organization

Lidl is now targeting an AI-first strategy, where it aims to become “100 percent agentic.” This represents a shift from AI as a helper to AI as the primary orchestrator.

The new vision is built around an architecture they describe as having an Orchestrator Agent Zero. This agent is the “first face the customer speaks to and sees,” and its job is not to solve every problem directly, but to manage a network of specialist agents behind the scenes.

For example, one agent would handle product issues, another would manage online-shop questions, another would manage loyalty and coupons, and so on. The Orchestrator routes, delegates, and supervises them all in real time. Quast described it this way:

“If a customer has several issues at once, they’re all working at the same time.”

This agentic ecosystem is the key to solving the multi-intent challenge and scaling automation further without compromising quality.

Humans as the final layer

Lidl stressed that their aim is not to eliminate humans but to deploy them where they make the greatest difference. Their approach is to automate as much as possible, and use human agents only as the last layer, called upon for empathy, nuance and edge cases. As senior CX executive and retail expert, Dominik Olejko, puts it:

“In 2025 and beyond, the brands that win will be those that master the art of blending AI-driven intelligence with human intuition, delivering experiences that are not only efficient but also deeply personal, trust-driven and above all human.”

What does this mean for the future of CX?

Other large organizations have followed a very similar path in the past year, moving away from plug-and-play NLU to bespoke automation and agentic architectures. Vodafone reported that generative AI pilots cut inbound call volumes by about 20 percent, and materially improved customer satisfaction, and its public case studies show the company is scaling generative solutions across markets as a co-pilot for agents.

Similarly, DoorDash explicitly built a generative-AI contact-center using Amazon Bedrock and Amazon Connect and says the solution now handles thousands of calls per day while enabling a 50x increase in testing capacity.

Unsurprisingly, automation and AI were found to be the top investment priority in 2025 in CX Network’s latest Global State of CX report. For CX teams and vendors, the practical implications are immediate and also measurable. Investing in clean data and feedback loops is essential; the ROI of LLM fine-tuning appears to outstrip the cost for businesses with complex taxonomies (retail, telecoms, finance).

Looking back at Lidl’s session at CCW Europe, the experience signals a broader reframing of contact centers: from cost centers to experience platforms that generate value. When generative AI reduces handling time and repeat contacts, organizations can either fold savings, or redeploy agent capacity to higher value conversations that drive conversion and loyalty!

Quick links:

• AT&T is making customers angry
• Black Friday becomes Black November
• CMA launches major crackdown on hidden online fees in first test of new consumer law