Impel swaps a third-party LLM for fine-tuned Llama in its dealership Sales AI and reports 20% improved accuracy

The problem

Impel sells AI software to car dealerships. Its Sales AI answers prospective buyers by email and text around the clock, to keep them engaged until they book a showroom visit. The company relied on a third-party language model. Impel says that provider's per-token pricing became too costly as volumes grew, and that limits on fine-tuning stopped it from fully using its own data to improve results.

What they built

Impel replaced the outside model with a Meta Llama model fine-tuned on its own data and hosted on Amazon SageMaker AI. Llama is an open-weight model, which means its files are published and a company can adapt and host it itself. The model powers three Sales AI features: summarizing past conversations to work out intent, writing follow-ups that stop deals stalling, and personalizing replies to a dealer's messaging.

How they did it

Impel fine-tuned the model on anonymized customer interaction data using LoRA, a low-cost way of adapting a large model. It shrank the model with a method called AWQ so it runs faster, and served it on SageMaker with automatic scaling to handle traffic peaks. Impel's R&D team worked with AWS account, strategy and SageMaker teams over several sprints before launch. Staff graded the incumbent and the new model on each use case.

Results

In Impel's evaluation, personalized replies improved from 73% to 86% accuracy, conversation summaries from 70% to 83%, and follow-up messages from 59% to 92%. Impel and AWS also state a 20% accuracy increase across core features. Impel lists predictable hosted costs, tighter control of training, and data processed inside its own AWS account as further benefits. It plans more in-house models and retrieval and agent workflows next.