Customer support · Production

Impel enhances automotive dealership customer experience with fine-tuned LLMs on Amazon SageMaker

The problem

As Impel's Sales AI transaction volumes grew, their third-party LLM provider's per-token pricing became cost-prohibitive, and restrictions on fine-tuning prevented them from leveraging proprietary customer interaction data to improve model quality.

First attempt

Impel's existing third-party LLM provider charged per-token pricing that became unaffordable at scale and did not allow sufficient fine-tuning on Impel's proprietary automotive data.

Workflow diagram · grounded in source
1
Customer inquiry via email or text
trigger
“Sales AI uses generative AI to provide instant responses around the clock to prospective customers through email and text”
2
Summarize past engagements
ai_action
“Summarizes past customer engagements to derive customer intent”
3
Personalize response
ai_action
“Personalizes responses to align with retailer messaging and customer's purchasing specifications”
4
Generate follow-up messages
ai_action
“Provides consistent follow-up to engaged customers to help prevent stalled customer purchasing journeys”
5
Lead to showroom or sales team
output
“This maintained engagement during the early stages of a customer's car buying journey leads to showroom appointments or direct connections with sales teams”
Reported outcome

By deploying a fine-tuned Meta Llama model on Amazon SageMaker, Impel achieved a 20% improvement in accuracy across core features, with personalized replies rising from 73% to 86%, summarization from 70% to 83%, and follow-up generation from 59% to 92%, alongside cost predictability and greater operational control.

Reported metrics
Overall accuracy improvement across core features20%
Personalized replies accuracy73% to 86%
Conversation summarization accuracy70% to 83%
Follow-up message generation accuracy59% to 92%
Show all 5 reported metrics
overall accuracy improvement across core features20%
personalized replies accuracy73% to 86%
conversation summarization accuracy70% to 83%
follow-up message generation accuracy59% to 92%
cost predictabilityCost predictability through hosted pricing, mitigating per-token charges
Reported stack
Amazon SageMakerMeta LlamaSageMaker LMI containersAmazon SageMaker StudioPyTorchtorchtuneLoRAAWQawscurl
Source
https://aws.amazon.com/blogs/machine-learning/impel-enhances-automotive-dealership-customer-experience-with-fine-tuned-llms-on-amazon-sagemaker?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

By deploying a fine-tuned Meta Llama model on Amazon SageMaker, Impel achieved a 20% improvement in accuracy across core features, with personalized replies rising from 73% to 86%, summarization from 70% to 83%, and f…

What tools did this team use?

Amazon SageMaker, Meta Llama, SageMaker LMI containers, Amazon SageMaker Studio, PyTorch, torchtune, LoRA, AWQ, awscurl.

What results were reported?

Overall accuracy improvement across core features: 20%; Personalized replies accuracy: 73% to 86%; Conversation summarization accuracy: 70% to 83%; Follow-up message generation accuracy: 59% to 92% (source-reported, not independently verified).

What failed first in this deployment?

Impel's existing third-party LLM provider charged per-token pricing that became unaffordable at scale and did not allow sufficient fine-tuning on Impel's proprietary automotive data.

How is this customer support AI workflow structured?

Customer inquiry via email or text → Summarize past engagements → Personalize response → Generate follow-up messages → Lead to showroom or sales team.

Related customer support cases