Workflows for work that runs the business

Mistral AI released Workflows in public preview today. The company describes it as an orchestration layer for enterprise AI, intended to bring durability, observability and fault tolerance to AI-powered processes moving from proof of concept to production.

Mistral said organizations including ASML, ABANCA, CMA-CGM, France Travail, La Banque Postale and Moeve are already running Workflows to automate critical processes. Workflows is part of Studio, Mistral's platform. Developers write workflows in Python, and each workflow can be published to Le Chat so anyone in an organization can trigger it. Every step is tracked and auditable in Studio.

Mistral cited three customer examples. In cargo release automation, a workflow validates shipping documents against customs rules, flags items needing human sign-off, waits for approval using a single line of code, wait_for_input(), then releases the cargo; Studio records the full execution history. In document compliance checking, KYC reviews are completed in minutes, with Studio surfacing each step as a structured timeline down to specific traces with native OpenTelemetry support. In customer support triage, incoming tickets are analyzed, categorized by intent and urgency, and routed automatically, with routing decisions visible and traceable in Studio.

Key features Mistral lists include durable execution that tracks state at every step and resumes where it left off; observability recording every branch, retry and state change; human-in-the-loop approval that pauses a workflow; native integration with Studio agents and connectors; workspaces and role-based access control for enterprise readiness; workflows written as code by engineers and run from Le Chat by business teams; and deployment flexibility where workers and data processing run in a customer's cloud, on-prem or hybrid environment.

Under the hood, Workflows is built on Temporal's durable execution engine. Mistral says it extended it for AI workloads with streaming, payload handling, multi-tenancy and observability. The control plane runs on Mistral, hosting the Temporal cluster, Workflows API and Studio, while workers are deployed on the customer's Kubernetes environment via a separate Helm chart. The Mistral SDK handles retry policies, tracing, timeouts, rate limiting and human-in-the-loop through decorators and single-line configuration. The Python SDK v3.0 is publicly available and installable with the command uv add mistralai-workflows.