Challenge
Giving hotel teams an immediate answer to their operational questions.
As part of the group's AI program for its hotels, one finding stood out: hotel teams juggled scattered tools and documents, and lost time on recurring support requests.
This situation had three consequences.
- Incidents that took longer to resolve. Finding the right procedure took time, every time.
- A support team taken up with repetitive questions. The same requests came back from one hotel to the next.
- An impact on the guest experience. Every minute spent searching for information is a minute less for the guest.
The need: an assistant that answers the operational questions of hotel teams from the support knowledge base.
Solution
A RAG agent evaluated, optimized and integrated into the group's platform.
We treated the assistant as a product in production, not as a demo. Three principles structure the project.
- Answers grounded in the knowledge base. The agent relies on support articles and cites its sources.
- Evaluation before optimization. An automated measurement framework guides each choice: model, context, answer format.
- Integration with existing systems. The assistant joins the group's RAG-as-a-service platform, to benefit from its foundation and enrich it.
Set-up
An agent, an evaluation framework and a shared RAG platform.
- RAG agent. A first version of the agent built on LangGraph on top of the ServiceNow knowledge base, then migrated to the OpenAI Agents SDK within an agentic platform.
- Evaluation framework. Designed in Langfuse: retrieval and citation recall, latency, cost, and LLM-as-a-Judge for faithfulness, tone and redirection.
- Optimization. Structured outputs, context engineering and a benchmark of several LLMs.
- Each hotel's context. Hotel-specific information complements the knowledge base to sharpen each answer.
- RAG-as-a-service platform. Integration of the knowledge base and the business glossary into the group's RAG platform, whose extension we then lead.
A team of eight data scientists and one person dedicated to project management. The team worked with the group's data engineers on the infrastructure side and with the business teams, from product managers to business stakeholders. The project started with three to four people, then accelerated with seven to eight people through to go-live.
Results
In production in 8 months, already deployed in more than 1,500 hotels.
The assistant is in production after an 8-month project. It now equips more than 1,500 of the group's hotels, and the rollout continues.
Each hotel benefits from a personalized experience. The assistant answers in its user's language, wherever they are in the world, and draws on the hotel's context to give a more precise answer.
We measure each evolution on the same criteria: retrieval and citation quality, faithfulness and tone of answers, ability to redirect to the right contact, latency and cost.
Expertises mobilisées
Savoir-faire
- RAG
- AI agents
- LLM-as-a-Judge
- Context engineering
- LLM benchmarking
Technologies
- LangGraph
- OpenAI Agents SDK
- ServiceNow
- Langfuse
Signature
Four mechanisms that make the difference.
- Sourced answers. Each answer links to the knowledge base articles it draws on.
- An answer tailored to each hotel. The assistant takes the user's language and the hotel's context into account, anywhere in the world.
- An LLM-as-a-Judge in service of quality. The framework evaluates faithfulness, tone and redirection automatically at each iteration.
- A shared foundation. Integration into the group's RAG platform benefits other use cases.
For you
An AI support assistant built on your knowledge base.
This approach addresses four situations.
- Your field teams lose time searching for procedures in scattered tools.
- Your support team handles the same recurring requests every day.
- You have an underused knowledge base, on ServiceNow or another tool.
- You want a measured and reliable AI assistant, not a demo.
We start from your knowledge base, your support processes and the questions of your teams. The experts who design the agent evaluate it, put it into production and evolve it with you.
