Lippert’s call volume was climbing fast, straining support teams and leading to inconsistent answers across a vast, complex product catalog. At the same time, warranty and quality data sat locked in disconnected silos across plants, vendors, and dealers, so problems in the field surfaced only after they had already cost real money.

Lovelytics partnered with Lippert to build two connected GenAI systems on Databricks: an AI-powered virtual agent that answers customer and agent questions in real time, and an eight-layer early warning engine that uses Anthropic’s Claude models to reason over warranty data and flag emerging risks before they escalate.

Lippert builds the parts that keep RVs, boats, and vehicles on the road, a catalog wide enough that even experienced teams can’t hold all of it in their heads. That scale was showing up in two different corners of the business at once: call centers were fielding a growing volume of repetitive questions, and warranty data scattered across plants, vendors, and dealers was too disconnected to catch quality problems early.

Lippert set out to solve both with generative AI, and brought in Lovelytics to build it.

Chapter One: An AI Virtual Agent for Customer Support

Long wait times and inconsistent answers were wearing down customer satisfaction and straining agent capacity. Leadership faced a familiar tradeoff: hire more agents to keep pace, or make the team they already had more effective. 

With headcount growth off the table, Lovelytics built a virtual agent using retrieval-augmented generation and vector search, then embedded it directly into Microsoft Teams so customers and agents both get accurate, real-time answers pulled straight from Lippert’s own product manuals and case logs. 

Three things made that possible:

• Retrieval-Augmented Generation & Vector Search: the assistant pulls exact answers from Lippert’s own knowledge base and holds context across multi-turn conversations, so a follow-up question doesn’t mean starting over.
• Automated Data Pipelines: product manuals and case logs feed into the system automatically, keeping answers current without manual updates.
• Source-Cited Answers: every response comes with a citation, so agents can verify it in one click instead of taking it on faith.

Chapter Two: Catching Supply Chain Failures Before They Happen, with Claude

Behind the scenes, a single analyst was manually recoding warranty claims by hand, reading through inconsistent and often incomplete technician notes to assign each one to the right issue category, and falling further behind every week. Lovelytics built a classification engine that automates that recoding for Lippert’s highest-volume product categories, standing in for the judgment of an experienced RV quality engineer.

Getting there meant teaching the model to work through the same judgment call an experienced RV quality engineer makes on every claim:

• Evidence Hierarchy: it weighs technician notes when their quality is good, cross-checks them against the claim’s part description and failure code, and scans the valid recode list, resolving any conflicting or ambiguous evidence with a defined evidence hierarchy and data-quality rules instead of guessing.
• Most-Specific, Validated Classification: it works toward the most specific classification the evidence actually supports rather than defaulting to a generic bucket, then checks that the result matches one of Lippert’s allowed recodes exactly.
• Explainable, Confidence-Scored Output: it extracts the component involved, summarizes the repair action the technician performed, assigns a confidence score, and explains its decision by ranking the strongest evidence behind it, not just a label.

Rather than trusting a single model’s read on a noisy note, four different LLMs, including two Claude models, each work through that process independently and generate their own classification and reasoning for every claim. Snorkel, a weak-supervision framework, then resolves the four outputs into a single, higher-confidence label. 

That ensemble approach lifted classification accuracy from the low 70s into the 90 to 93 percent range, and the misses that remain are almost always a symptom of thin source data rather than a shortcoming in the models’ reasoning.

The same system doubles as an early warning capability, true to the name: because claims are recoded automatically instead of sitting in a manual queue, safety and quality issues surface far sooner than the old process allowed. 

As it classifies each claim, the ensemble also determines a safety category directly from the technician notes, deliberately independent of the recode classification, so flagging a hazard never skews which issue category a claim lands in, and vice versa. Claims that point to genuine product risk, such as an electrical wiring fault or a door assembly failure, get rated on a green-yellow-orange-red scale so quality teams know at a glance which items need urgent corrective action. 

Because all four models, including both Claude models, have to agree before a claim gets flagged red, those findings carry the kind of consensus that can hold up in a legal review.

Lippert needed a partner who could operationalize generative AI in a way the business could actually trust: accurate, source-backed, and ready to scale across more than one part of the company. Lovelytics’ experience pairing Databricks with Claude, and building both a customer-facing assistant and a business-critical reasoning system on the same platform, is what let Lippert grow from one AI initiative into two.

On the customer support side, accuracy alone wasn’t enough to earn agents’ trust. An assistant that’s right most of the time still gets ignored if people can’t tell when to double-check it, so a source citation on every answer becomes non-negotiable. 

On the supply chain side, the team learned that sending every anomaly straight to an LLM wastes both cost and attention. Filtering the statistical layers down to only the top-ranked, highest-risk claims before Claude ever sees them kept the reasoning sharp without inflating inference cost, a design choice that now shapes how Lovelytics builds LLM reasoning into layered systems more broadly.

Within months, Lippert had two GenAI systems running in production: one handling half of all incoming support calls, the other catching warranty and quality issues while there’s still time to act.

Together they’ve reshaped how two different teams work day to day, and Lippert has already reported measurable savings from both.

• 50% of incoming call volume now handled by the AI assistant
• $2.1M in projected annual savings from the virtual agent
• 90% increase in support accuracy
• 106,000+ work hours saved across the organization
• Supply chain detection lag cut from months to days
• $1M–$3M in annual cost avoidance reported by Lippert from the Early Warning System

Lippert and Lovelytics are continuing to build on this foundation. The virtual agent is expanding into EMEA to support multilingual customer service, and the Early Warning System is scaling into more issue types as Lippert classifies it “proven, expanding now.” Powered by Databricks and Claude, the goal is a GenAI foundation Lippert can keep building on across the business.