How Experian Automates 35% of Customer Emails with Databricks Mosaic AI
Experian, the multinational data broker holding data on 1.1 billion people across 32 countries, built “Latte” — a GenAI email automation system — on Databricks Mosaic AI by fine-tuning a Llama 8B model. The system now handles 35% of incoming contact center emails autonomously, reduced model fine-tuning time 11x from 86 hours to under 8, and lifted customer NPS by 8%.
Impact
35%
Customer emails automated
8%
NPS improvement
11x faster
Fine-tuning time improvement
1,000+
Daily emails handled by AI
Challenge
Experian’s contact center faced escalating email volumes without a scalable AI solution — Llama model fine-tuning took 86 hours, existing cloud infrastructure lacked native LLMOps capabilities, and regulatory traceability requirements for a global consumer credit company made deploying AI with stitched-together third-party tools impractical.
Solution
Experian built “Latte” on Databricks, fine-tuning a Llama 8B model with Mosaic AI and DBRX-generated synthetic data, powering email understanding through a Vector Search RAG pipeline, and satisfying regulatory compliance via Unity Catalog and MLflow governance — reducing fine-tuning time 11x and automating 35% of incoming customer emails.
Tools & Technologies
What Leaders Say
“During early testing, Databricks delivered faster LLM response times and six times the initial token throughput at the same hourly cost, compared to their cloud services. Even after optimizations, Databricks maintained a 3x performance advantage, all without compromising model accuracy or compression.”
“Once we could prove our GenAI models were stronger for text-based email automation, we could repurpose the initial call-based AI respondent for the customer support center through a more scalable, text-forward approach.”
“Since employees can now focus on more complex customer needs, internal job satisfaction has improved and, more importantly, without any loss of human jobs.”
Sign up to read complete case studies, access detailed metrics, and unlock all use cases.
Full Story
Experian operates globally across 32 countries with more than 22,000 employees, holding credit and financial data on 1.1 billion people and 150 million active businesses. The company’s contact center fields thousands of inquiries daily — questions about credit scores, credit freezes, account status, and financial literacy — from consumers navigating major financial decisions. As a business that processes sensitive personal data at global scale, any AI deployment had to operate within a fully governed, private environment without exposing customer data to public model providers.
The contact center was under sustained pressure from growing inquiry volumes. Existing systems couldn’t scale to meet demand without degrading response times. When the team attempted to address this with AI tooling, the infrastructure proved inadequate: fine-tuning a Llama model took 86 hours, cloud services lacked native LLMOps capabilities, and there was no unified framework for model governance, synthetic data lineage, or regulatory traceability. Building a production-ready GenAI pipeline from stitched-together third-party tools was both time-consuming and difficult to govern.
Experian chose Databricks as the unified platform to build, fine-tune, deploy, and govern Latte. Using Mosaic AI, the team fine-tuned a Llama 8B model on instruction datasets initially generated with MPT models, then shifted to DBRX — Databricks’ open source LLM — for richer, more accurate prompt handling. Databricks Vector Search powered a retrieval-augmented generation pipeline that enabled Latte to answer questions like “How do I freeze my credit?” and “How can I lock my report?” by semantic intent, not keyword matching. Unity Catalog and MLflow provided full traceability across the model lifecycle — from synthetic data generation through fine-tuning inputs to production outputs — satisfying the regulatory audit requirements a consumer credit company must meet. The system launched with human-in-the-loop validation before progressively transitioning to autonomous operation.
The production results were clear across every measure. Fine-tuning time dropped from 86 hours to under 8 hours — an 11x improvement — with some production runs completing in under an hour at approximately $100. Latte now handles more than 35% of incoming customer emails without human involvement, a figure that has grown steadily as confidence in the system’s accuracy increased. The NPS lift of 8% confirmed that automated responses were improving the customer experience rather than degrading it. Contact center teams, freed from routine email responses, shifted focus to more complex customer issues that genuinely require human judgment.
Experian’s roadmap targets 50% automated email resolution and plans to expand Latte to other customer service channels. Beyond the contact center, the company has documented dozens of GenAI use cases for future delivery. Latte has become a template: a proof that AI-driven automation can directly contribute to company-wide OKRs — from stronger customer satisfaction scores to employee retention — without displacing the workforce that makes it possible.
AWS