Lidl's AI-First Support: Accuracy, Scale, and a Clear Path for CX Leaders
At CCW Europe 2025, Lidl delivered a rare thing in CX: clarity. The retailer laid out a problem most support teams feel every day - massive volume, huge topic variety, tight margins for error - and showed how an AI-first model can raise quality and cut cost at the same time.
The centerpiece is Lia, "the face of their customer service." Behind Lia sits an automation engine powered by large language models and Lidl's proprietary system, Optimus. The goal is simple: resolve more with higher accuracy while keeping humans focused on the work that truly needs them.
The grocery support reality: high variety, zero slack
Grocery support isn't a tidy queue of password resets. Lidl's team manages 150 distinct query types: product quality, availability, coupons, deliveries, substitutions, recipes, bottle-return machines, online shop issues, and more. Complexity isn't a side note - it's the job.
Many issues sound similar but require different actions. Coupon complaints are a good example. "Coupons unavailable," "coupons expired," and "coupons didn't apply" are three different cases with three different resolutions. Misclassify the intent and everything downstream breaks - routing, resolution, and customer satisfaction.
Breaking the accuracy barrier: from NLU to LLMs to Optimus
Lidl started with classical NLU at roughly 70% accuracy. An untrained LLM pushed that to 80%. That was the moment they saw a new ceiling for automation.
The real lift came next: training Google Gemini on Lidl's own messages, terms, and product taxonomy to create Optimus. Accuracy climbed to 92% by teaching the model grocery-specific nuance - like why pet food and diapers both live under "food" in Lidl's system, or how private-label names map to internal categories. As they put it: if you don't know the category, you can't automate the process.
Learn more about Google Gemini
What's still hard: multi-intents and rigid flows
Two gaps remain. First, Optimus classifies the first intent and then hands off to deterministic flows, limiting the model's flexibility mid-conversation. Second, multi-intent messages (e.g., partial delivery + return + delivery status) get reduced to one path. "Multi-intents are currently not supported," said Fabian Quast.
Agentic support: Orchestrator + specialists
Lidl's next move is "100 percent agentic." An Orchestrator Agent Zero will be the first face the customer meets. It won't do everything - it will manage a network of specialist agents for products, online shop, loyalty and coupons, and beyond. If a customer brings several issues at once, multiple agents work in parallel.
This is how they'll solve multi-intents and keep quality high at scale. AI does the heavy lifting. Humans handle the edge cases, empathy, and nuance - the final layer that makes experiences feel personal and builds trust.
As Dominik Olejko put it: the winners blend AI-driven intelligence with human intuition to deliver experiences that are efficient, personal, and trust-driven.
Why this matters for customer support leaders
The pattern is showing up across large organizations. Vodafone reports generative pilots cut inbound calls by around 20% while improving satisfaction. DoorDash built a generative contact center on Amazon Bedrock and Amazon Connect that now handles thousands of calls a day and enables far more testing.
Translation: the ROI is real when you pair clean data with domain-specific tuning. For complex taxonomies like retail, telecom, and finance, fine-tuning and tight feedback loops beat plug-and-play setups.
A practical blueprint you can use
- Map the work: build a clear taxonomy of intents, sub-intents, and required actions. Treat "similar-sounding" issues as separate playbooks.
- Measure misclassification cost: track where wrong routing breaks resolution and inflates handle time.
- Start with baseline LLMs: prove lift over NLU on historical data before you customize.
- Fine-tune with your data: use real customer messages, internal terms, and category logic. Protect PII end to end.
- Add guardrails: constrain actions, call trusted tools and APIs, and require confidence thresholds for auto-resolve.
- Introduce an orchestrator: route to specialist agents (refunds, coupons, deliveries, product issues) and allow parallel work for multi-intents.
- Keep humans as the last mile: step in for empathy, policy exceptions, and unresolved or sensitive cases.
- Close the loop: feed outcomes back into training, update taxonomies, and refine prompts and tools weekly.
Metrics that matter
- Intent accuracy and coverage (single vs. multi-intent)
- First contact resolution and repeat contact rate
- Average handle time and time-to-first-action
- Deflection with quality (CSAT on automated vs. human)
- Wrong-automation penalty (cost of misroutes and escalations)
- Containment with safeguards (fallback rate to human)
Build or buy: quick guidance
- Use off-the-shelf where your taxonomy is simple and the cost of a miss is low.
- Invest in fine-tuning where internal terms, categories, and policies drive outcomes.
- Adopt an agentic architecture once single-intent accuracy is stable and you see multi-issue threads in your transcripts.
Team enablement
If you're standing up AI-driven support functions or re-skilling your team for agentic operations, curated training helps. See role-based options here: AI courses by job.
The bigger shift
Lidl's session reframed support from a cost center to a value engine. When generative systems reduce handle time and repeats, you can bank the savings or redeploy agents to higher-value conversations that drive conversion and loyalty.
That's the game: higher accuracy, faster paths to resolution, and humans focused where they matter most.
Your membership also unlocks:
