Mistral AI Workflows Public Preview

Mistral AI has launched Workflows in public preview, an orchestration layer designed to move AI-powered business processes from proof-of-concept to production. Workflows addresses the common failure modes of enterprise AI—such as silent failures in production, network timeouts in long-running processes, and the lack of mechanisms for human approval—by providing a durable, observable, and fault-tolerant execution environment.

Core Capabilities of Workflows

Workflows provides a set of enterprise-grade features that allow developers to focus on business logic rather than recovery and infrastructure logic:

Durable Execution: The system tracks state at every step, allowing failed processes to resume exactly where they left off.
Observability: Every state change, retry, and branch is recorded in Studio, providing a full auditable timeline of how decisions were reached.
Human-in-the-Loop: Developers can pause a workflow for human approval using a single line of code (wait_for_input()). The workflow pauses without consuming compute and resumes once a reviewer responds via Le Chat, a webhook, or another connected surface.
Native Studio Integration: Workflows utilize the same agents and connectors as the rest of the Mistral Studio ecosystem, eliminating separate integration work.
Enterprise Governance: Studio Workspaces provide project separation, while role-based access control (RBAC) ensures consistent enforcement of access rules.
Deployment Flexibility: The control plane is hosted by Mistral, while workers and data processing run within the customer's own environment (cloud, on-prem, or hybrid).

Real-World Enterprise Applications

Several organizations, including ASML, ABANCA, CMA-CGM, France Travail, La Banque Postale, and Moeve, are utilizing Workflows to automate critical business processes:

Cargo Release Automation

In global shipping, Workflows is used to validate shipping documents against customs rules and check for anomalies. The system flags items requiring human sign-off and uses the wait_for_input() function to pause execution until approval is granted, preventing cargo delays and compliance breaches.

Document Compliance Checking

For KYC (Know Your Customer) reviews, Workflows automates the extraction of identity documents and verification against sanctions lists and PEP databases. This reduces the process from hours of manual analyst time to minutes, while providing a structured timeline of reasoning and native OpenTelemetry support for regulatory auditability.

Customer Support Triage

Workflows automates the analysis, categorization, and routing of incoming support tickets based on intent and urgency. Because routing decisions are visible and traceable in Studio, teams can correct miscategorizations at the workflow level without needing to retrain the underlying model.

Technical Architecture and Implementation

Workflows is built upon Temporal's durable execution engine, which is used by companies like Netflix, Stripe, and Salesforce. Mistral AI has extended this engine to support AI-specific requirements, including streaming, payload handling, multi-tenancy, and enhanced observability.

Deployment Model

The architecture separates the control plane from the data plane:

Control Plane: Mistral hosts the Temporal cluster, the Workflows API, and Studio.
Data Plane: Users deploy workers in their own Kubernetes environment via a Helm chart. These workers connect back to the central cluster using secure credentials, ensuring that business logic and data remain within the user's perimeter.

Developer Experience

Developers write workflows in Python using the Mistral SDK. The SDK manages complex operational requirements—including retry policies, timeouts, rate limiting, and tracing—through decorators and single-line configurations.

Workflows are now available via the Python SDK v3.0, installable via uv add mistralai-workflows.