Microsoft's Copilot Studio gained a powerful new AI model on May 28, 2026, with the addition of Mistral Medium 3.5. The model rolls out worldwide in early release environments, bringing advanced agent-building capabilities and enterprise-grade governance controls—including data residency for European Union customers.

The move deepens Microsoft’s multi-model strategy in the low-code AI agent market. It pairs Paris-based Mistral’s privacy-first architecture with Copilot Studio’s existing orchestration layer, allowing organizations to build agents that keep data within the EU while tapping into a top-tier reasoning engine.

What Mistral Medium 3.5 Brings to Copilot Studio

Mistral Medium 3.5 is the latest mid-range model from the French AI startup known for steering clear of US hyperscaler training pipelines. Benchmarks place it close to GPT-4o in reasoning and multilingual tasks, but with a smaller compute footprint and native support for French, German, and other European languages.

Inside Copilot Studio, the model functions as a drop-in option for topic creation, generative answers, and custom GPT-like actions. Admins select it from the model picker in the authoring canvas—no extra licensing required if the tenant holds a Copilot Studio capacity pack or Power Platform premium license.

Key differentiators:
- Data sovereignty: Inference runs in Azure’s EU data centers (Amsterdam, Dublin, Frankfurt, Paris) with customer data encrypted in transit and at rest using customer-managed keys.
- Transparency: Responses include citations drawn from configured knowledge sources, with optional hallucination scores visible in test mode.
- Multilingual fluency: Outperforms US-centric models on French and German legal, medical, and manufacturing terminology—critical for EU-regulated industries.

Agent Governance Gets a Boost

Alongside the model launch, Microsoft updated the Copilot Studio admin center with new governance tools that apply to Mistral-powered agents and existing GPT-4o agents alike. These controls address the top pain point for IT managers: shadow AI agents spinning up without oversight.

Environment-scoped model policies let admins restrict which models can run in each environment. For example, a bank might allow Mistral Medium 3.5 only in the production environment that resides in an EU data center, while limiting test environments to less sensitive models. Policies roll out in early release starting May 28 and graduate to general availability within 60 days.

Agent lifecycle management now includes mandatory approval flows for publishing. Agents that generate external-facing content or access SQL connectors must pass a two-stage review: a technical audit (endpoint checks, PII masking) and a business owner sign-off. Audit logs stream to Microsoft Purview, where compliance teams can trace every interaction.

Usage dashboards in Power Platform Admin Center break down token consumption, response latency, and error rates by model. A new “cost analyzer” projects monthly spend based on current usage patterns, factoring in the Mistral model’s lower per-token cost (reportedly 30–40% cheaper than GPT-4o for text generation).

Data Residency for the EU: How It Works

EU customers have clamored for Copilot Studio to keep their prompts and agent outputs inside the bloc. Mistral Medium 3.5 delivers that through Azure’s “EU Data Boundary” pledge, which went fully operational in 2025.

When an admin deploys the model in a West Europe, North Europe, or France Central region, all processing—from prompt intake to response generation—stays within EU-located compute and storage. Microsoft’s own support engineers cannot access the data without explicit customer authorization, a process now streamlined through the Customer Lockbox for AI preview.

For organizations subject to GDPR, the setup removes legal friction. Data protection impact assessments (DPIAs) become simpler because the model never touches US infrastructure. Microsoft also updated its Data Protection Addendum to list Mistral as a sub-processor under the same contractual guarantees as Azure OpenAI Service.

How Admins Enable Mistral in Copilot Studio

Turning on the model requires a few clicks:
1. Navigate to Power Platform admin center > Environments > [select environment] > Settings > AI Builder > Models.
2. Enable the Mistral Medium 3.5 entry. This surfaces the model in Copilot Studio’s canvas for all authorized makers in that environment.
3. Optionally, set a default model per environment—if Mistral is default, new agents auto-select it; otherwise makers can pick it manually.
4. Apply data loss prevention (DLP) policies that prevent Mistral agents from connecting to non-EU connectors or exporting data to unsanctioned endpoints.

Microsoft ships a prebuilt DLP template for “EU-regulated agent builders” that blocks over 200 connectors by default, including all US-based social media and productivity services outside the Microsoft 365 boundary.

The service auto-scales. No GPU quota management required; Copilot Studio taps Azure’s reserved Mistral endpoints, so capacity expands with demand. Early tests show cold-start latency under 800ms for most text-generation tasks.

What It Means for EU Orgs and Makers

The combination of Mistral’s model and Copilot Studio’s orchestration fills a gap for EU enterprises that wanted advanced AI agents without handing data to US-based inference servers. Makers inside these companies can now build:
- Internal knowledge bots that parse multilingual SharePoint libraries and respond in the employee’s language without translation layers.
- Customer service agents that handle French, German, or Dutch inquiries with native fluency and route complex cases to human reps based on PII detection.
- Compliance assistants that scan 100-page contracts, flag risky clauses, and generate summaries in the same EU region where the originals are stored.

Large government contractors and automotive suppliers already piloting the integration report a 40% reduction in agent development time because Mistral’s instruction-following requires fewer prompt engineering iterations for non-English languages.

Pricing and Licensing

Mistral Medium 3.5 consumes AI Builder capacity credits, which are included in the standard Copilot Studio subscription. Microsoft confirmed a consumption-based metering model with the following approximate rates (prices in EUR, subject to early release adjustments):

Task Credits per 1,000 tokens
Prompt text 0.08
Completion 0.24

A 500-agent, low-complexity conversational bot that handles 10,000 queries per month would cost roughly €120/month, about 35% less than the same bot built on GPT-4o. Volume discounts kick in at 10 million tokens per month.

Enterprise Agreement customers can negotiate reserved capacity, and Microsoft plans to introduce a monthly cap feature in Q3 2026 to prevent budget overruns.

Competitive Landscape

Microsoft’s embrace of an independent European model puts pressure on Salesforce’s Einstein GPT and ServiceNow’s Now Assist, both of which still route EU data to US infrastructure by default. AWS Bedrock offers Anthropic’s Claude in EU regions, but lacks a native low-code agent builder with the governance depth Copilot Studio now provides.

At the same time, Google Cloud’s Vertex AI Agent Builder added Mistral last quarter, but without the tight Power Platform integration—meaning no unified DLP policy that spans both model selection and connector access. Microsoft’s advantage lies in the admin-center consolidation.

Roadmap and What’s Next

Microsoft’s AI at Scale team outlined upcoming capabilities on the roadmap:
- Fine-tuning via Copilot Studio (public preview Q4 2026): Makers will be able to upload domain-specific documents to adapt Mistral Medium 3.5 within the same EU data boundary, with no need to move data to a separate training service.
- Agent-to-agent collaboration: Mistral agents will summon GPT-4o agents and vice versa within a single conversation, orchestrated by Copilot’s router model. This multi-agent mode will enter private preview in fall 2026.
- EU-specific compliance packs: Pre-configured intent libraries for GDPR, DORA, and EU AI Act readiness will ship later this year, reducing the time to build regulatory agents from weeks to hours.

Bottom Line

Mistral Medium 3.5 in Copilot Studio isn’t just another model dropdown. It’s a statement that Microsoft intends to meet EU sovereignty demands without walling off customers into a separate, less capable platform. The governance layer that ships alongside—model policies, mandatory approvals, cost controls—gives IT leaders the levers they’ve been asking for since generative AI exploded into the enterprise.

For any organization still holding back on AI agents because of data residency fears or governance gaps, the combination dissolves two of the biggest blockers. The early release gives businesses a chance to experiment before general availability locks in the feature set around July 2026. Act now, and you might build the agent that redefines internal operations well before your competitors finish their DPIA.