Mistral Unveils Workflows

Workflows has been released in public preview. Mistral AI is positioning it as an orchestration layer for enterprise AI, stating that it delivers durability, observability, and fault tolerance to run AI processes reliably in production.

Developers write workflows in Python and deploy them to Le Chat, allowing users within an organization to run them immediately. In Studio, every step is tracked and audited, with execution history, branches, retries, and state changes preserved in a timeline.

Representative use cases include the following.

• Cargo release automation: Cargo release after customs, hazardous materials, and safety checks followed by human approval
• Document compliance checking: Extracting and verifying KYC documents and cross-checking them against sanctions/PEP data
• Customer support triage: Classifying inquiries and routing them to the appropriate follow-up process

Core features include durable execution, human-in-the-loop, native Studio integration, RBAC-based multi-tenancy, and cloud / on-prem / hybrid deployment. Internally, it is built on Temporal, and Mistral explained that it extended streaming, payload handling, multi-tenancy, and observability to fit AI workloads. The control plane is operated by Mistral, while execution workers are deployed within the customer's Kubernetes environment, keeping data and business logic within the customer's own boundary.

Additionally, the Mistral SDK handles retry policies, tracing, timeouts, rate limiting, and human-in-the-loop through decorators and a single configuration, and Python SDK v3.0 was also released.