What challenges did ELCA face before AI implementation?

The main challenge was to build a chatbot capable of answering employee questions related to internal IT standards and documents while safeguarding sensitive information. Handling hallucinations and inaccuracies frequently encountered in language model outputs for niche and internal company domains was a key issue. Running costly third-party cloud models was not feasible due to privacy, compliance, and cost concerns.

Source

MicrosoftProductionEvidence: Low40/100

ELCA Switzerland: Cost-Efficient AI Chatbot Using LoRA for Internal IT Support

Use case typeConversational assistantsUpdated Jun 13, 2026

ELCA, a Swiss independent IT company with 2,300 experts, developed an accurate, cost-effective, and trustworthy internal chatbot to address employee queries about IT standards, legal documents, and proprietary workflows. The chatbot is designed to keep sensitive data secure while significantly reducing hallucination problems.

Organization: ELCA
Industry: Tech & Comms
Location: Switzerland
Published: July 2025

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: ELCA
Provider: Microsoft
Maturity: Production
Linked source: LinkedIn

The team applied Low-Rank Adaptation (LoRA) techniques for model fine-tuning which drastically cut hardware requirements and operational costs

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 2 of 2

1Conversational AI
2Custom AI Model Fine-Tuning

The main challenge was to build a chatbot capable of answering employee questions related to internal IT standards and documents while safeguarding sensitive information.
Handling hallucinations and inaccuracies frequently encountered in language model outputs for niche and internal company domains was a key issue.
Running costly third-party cloud models was not feasible due to privacy, compliance, and cost concerns.

ELCA implemented a Retrieval-Augmented Generation (RAG) system combined with fine-tuned open-source LLaMA 3 models using Microsoft Azure infrastructure.
The team applied Low-Rank Adaptation (LoRA) techniques for model fine-tuning which drastically cut hardware requirements and operational costs.
They built a domain-specific question-answer dataset by curating thousands of QA pairs, including manual question creation and augmentation using LLMs.
Custom metrics were designed to evaluate the model, prioritizing source citation accuracy to reduce hallucinations.
Fine-tuning was performed using QLoRA (Quantized LoRA) on LLaMA 3.1 8B base model with 4-bit quantization, enabling efficient training on a single Nvidia L40S GPU.

Technologies

Microsoft Azure Retrieval-Augmented Generation (RAG)Low-Rank Adaptation (LoRA)LLaMA 3 QLoRA

Architecture

The architecture uses Retrieval-Augmented Generation (RAG) with fine-tuned open-source LLaMA 3.1 8B models using QLoRA technique for efficient fine-tuning. The chatbot pulls information from a curated internal data source, specifically a Confluence space containing IT tutorials and guidelines. The model was trained and evaluated with a customized source-based metric ensuring high precision and low hallucination. The solution is hosted on Microsoft Azure infrastructure and trained on a single Nvidia L40S GPU.

Sources & evidence1

Evidence: Low40/100Evidence strength