SAP adopts Mistral AI models to power S/4HANA migration assistant At Swiss Federal Railways

SAP is deploying Mistral AI models to power a multilingual retrieval‑augmented chatbot used in Swiss Federal Railways’ (SBB) migration to S/4HANA, a move designed to speed employee support during a complex ERP upgrade while keeping data on SAP infrastructure in Europe. The assistant addresses day‑to‑day questions from roughly 30,000 SBB employees, surfacing answers from internal documentation and routing issues it cannot resolve to human experts.

At SBB the chatbot recognizes the technical abbreviations staff use and pulls responses from the company’s internal corpus rather than external sources. When the tool escalates a question, experts reply and those human responses are automatically ingested back into the system, creating a closed feedback loop that continually updates the assistant’s knowledge base. SAP presents the deployment as a support aid for migration tasks, not a replacement for subject‑matter experts.

Behind the scenes, SAP says it runs Mistral AI models as part of its migration toolkit, with workloads hosted on SAP infrastructure in Europe so that no data crosses regional borders. The vendor and Mistral began collaborating in June 2024 and expanded that work in October. The initiative specifically targets moves to S/4HANA, SAP’s current core enterprise suite for accounting, procurement and logistics — a common modernization path for customers overhauling legacy ERP landscapes.

Wider customer adoption remains cautious. Many SAP customers are reportedly still “kicking the tires” on AI, and practical barriers such as limited cloud adoption among certain customer bases and complex licensing arrangements could slow broader rollouts of comparable AI‑supported migration tools. For builders and system integrators the SBB example highlights practical design choices: pair retrieval‑augmented generation pipelines with internal corpora; build robust multilingual retrievers that handle abbreviations and domain terms; include human‑in‑the‑loop escalation with automated ingestion of expert replies; and plan deployments that respect in‑region data residency and existing licensing models.