Roatel
Roatel's micro-hotel chatbot, built on Llama 3.1 405B, handles 80% of customer service requests in 11 languages
Roatel, a German chain of self-service micro-hotels, uses a multilingual guest chatbot built by Brain4Data on Meta's open-weight Llama 3.1 405B. Meta's case study says the bot handles 80% of customer service requests and went live in two months.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
The problem
Roatel runs self-service micro-hotels in Germany and plans to grow across Europe. Guests deal with a remote service team. As the chain grew, that team struggled to keep up with demand, and many guests come from abroad.
What they built
Brain4Data, a German AI company, built a customer-service chatbot for Roatel on its FRED assistant platform. It uses Meta's Llama 3.1 405B, an open-weight model, meaning its files are published so companies can run it themselves. The bot answers in 11 languages at any hour and gives an overview of the Roatel portfolio during booking. A QR code on site opens it on a guest's phone.
How they did it
The bot uses retrieval-augmented generation: it looks up relevant passages in Roatel's documents and has the model answer from them. Meta's case study says the model was used without fine-tuning and ran on Oracle Cloud Infrastructure, with a vector database and an embedding model from Cohere for the lookup. Meta's case study says the bot was fully operational in two months.
Results
Meta's case study says Llama-powered AI handles 80% of Roatel's customer service requests, and that the bot answers in the customer's language. Roatel's chairman said in the September 2024 announcement that the bot lets the company respond to questions around the clock and in different languages, and that more AI projects were planned.
Lessons for your project
Choose a model that handles many languages well. Roatel's guests are often foreign, and the bot answers in 11 languages.
Try retrieval before fine-tuning. This bot draws on Roatel's own documents and uses the base model without fine-tuning.
Give guests an easy entry point. Roatel puts a QR code on site that opens the assistant on a phone.
Define what 'handled' means before you quote a rate, since the 80% figure comes from the builder and Meta.
Caveats
- The 80% and two-month figures come from Meta's Llama case study about Brain4Data, not from Roatel. Meta marks the results as self-reported. The case study is vendor-reported.
- Roatel's own press release gives no figure. It says only that the bot answers guest questions in 11 languages around the clock.
- The Meta page is undated. The deployment date and publication date are those of the Roatel and Brain4Data press release (23 September 2024), which says the chatbot 'will now be' used, so it is an announcement date.
- Brain4Data, the builder, repeats the 80% figure on its own site in a news item about the Meta case study. That page prints no date and adds no new measurement, so it is not independent confirmation.
- The Llama and Oracle details are from Meta's case study; the press release names no model.
Sources
- Brain4Data
Meta Llama case studies · Undated · Vendor case study · dev.meta.ai
- Roatel now utilizes Artificial Intelligence
Roatel Holding AG and Brain4Data · 2024-09-23 · Press release · brain4data.de
- Brain4Data's AI solution is now a Meta case study
Brain4Data · Undated · Company blog · brain4data.de
Summarized in our own words from the sources above. Spotted an error? See corrections.