The 10 platforms for AI governance in the mid-market

The closest German AI workspace. EU residency is real. Ownership of the application layer is not.

Langdock sells to the same buyer: a GDPR-compliant enterprise workspace with chat, agents, workflows, RAG and MCP. Frankfurt hosting is residency. On-premise on your Kubernetes exists — behind an Enterprise contract, typically sized for 1,000+ users, still licensed per seat, as a closed binary. For mid-market DACH it is the best closed-source alternative. It is still rent.

• EU hosting in Frankfurt; on-prem exists, but it is Enterprise
• No public source you can audit or fork
• Included models are fair-use capped; fallback to a cheaper model
• BYOK: your keys, their model list

No US parent. That closes the CLOUD Act hook. You still do not own the application layer.

Mistral is EU-headquartered, processes in the EU, ships a DPA under EU law, and offers on-prem for regulated deployments. ISO 27001 is public. The jurisdiction argument is the strongest after an open-source application you run yourself. Pricing sits in genuine enterprise (quotes, often well above mid-market budgets), and what you buy is Mistral’s product — not your fork.

• EU headquarters, no CLOUD Act hook via a US parent
• On-prem available; entry is enterprise pricing, not mid-market
• Application layer stays closed
• Models are also sold via US clouds (Bedrock, Vertex) — the routing path counts

German models, on-prem and EU-sovereign operating models. For mid-market teams the missing piece is the application layer where the work happens.

Aleph Alpha has oriented around sovereignty for government, defence and regulated industry. Pharia offers on-prem and EU-sovereign deployment, a DPA under German law, and a no-training commitment. That is a model layer that survives a CLOUD Act conversation. It is not an open-source workspace with agents, workflows and MCP for marketing, legal and sales. Buy Pharia for inference — you still need something that owns the application.

• On-prem and EU-sovereign, DPA under German law
• Built for sovereign and regulated deployments, not a typical mid-market rollout
• No public workspace you can fork
• Fits as a model behind Navigator, not as a replacement for it

Joule lives in the SAP landscape. Sovereign-cloud options and European models do not change who owns the application.

For SAP mid-market, Joule is the path of least resistance. It is model-agnostic (including Mistral and Aleph Alpha) and SAP markets Sovereign Cloud. The generative AI hub is not classic on-prem with the LLM on your servers. The application layer stays SAP. Governance here means SAP processes plus SAP control — not a workspace you can fork when the contract ends.

• ERP depth no generic workspace replaces
• Sovereign-cloud tiers; not a classic on-prem LLM on your hardware
• Application layer closed, seat-licensed
• Sensible beside a workspace you own — not instead of one

Inventory, risk, evidence. Exactly what a KI-policy file does not do — and still not the inference chain.

watsonx.governance is the most mature dedicated governance suite on this list: model inventory, risk scoring, policy, evidence for the AI Act. Mid-market teams rarely buy it alone. IBM is a US parent. The overlay documents systems that often run on IBM or hyperscaler cloud. A policy in watsonx does not change who can compel your prompts.

• Strong for inventory, risk and AI Act evidence
• US parent; CLOUD Act stays relevant where inference sits on IBM/hyperscalers
• Does not replace a workspace or your own model layer
• Complementary to a stack you own

Claude is often the strongest model in the room. Availability and jurisdiction are set by someone else.

On 12 June a US export-control order told Anthropic to cut off its two newest models for foreign nationals. Anthropic disabled them for everyone, three days after launch. That is not GDPR theory. That is a switch. Enterprise residency and zero-retention options exist; they are promises from a US provider. For mid-market: use frontier where it earns the price, and put the work that has to keep running on models you can host.

• Frontier quality that routine work rarely needs
• US parent; export control can cut access with no warning
• Zero retention is an option and a contract, not a proof
• Fits behind routing you control — not as the only layer

Vertex AI in europe-west is residency. The company holding the bytes is still Google.

Gemini for Workspace and Vertex AI are the second default after Copilot once Google is already in the building. EU regions, a DPA, no-training on the enterprise tier: all residency and contract. The CLOUD Act still reaches the US parent. For AI privacy (processor terms, RoPA) that can be enough. For sovereignty under the AI Act — who can compel access — it is not.

• EU regions and Workspace integration are real
• US parent; CLOUD Act and DPF risk unchanged
• Application layer closed
• No mid-market self-host of the workspace

The name on every AI policy. Residency and no-training on Enterprise are real controls. They are not sovereignty.

ChatGPT Enterprise and the API with zero-data-retention (on the right account) are what most mid-market AI policies describe today. Default API retention remains a window, commonly up to 30 days, on US infrastructure until ZDR is approved and verified. The company is US-controlled. The EU-US Data Privacy Framework many transfers sit on is the third attempt and is in court again. The two before it were struck down.

• Enterprise: no-training, SSO, residency options — documented controls
• Default retention is not zero; ZDR is a request plus verification
• US parent; CLOUD Act and DPF challenge
• You rent the application layer entirely

The most common tick in a mid-market AI policy. Microsoft’s own lawyer said under oath that he cannot guarantee the outcome.

Copilot wins procurement because Office is already there and “EU data boundary” is on the slide. Residency in Azure EU regions is real and useful. Sovereignty follows the US parent. In June 2025 Microsoft’s Director of Public and Legal Affairs for France told the Senate, under oath, that he could not guarantee French citizens’ data would never reach US authorities without the French state’s approval. For a bank under DORA, a hospital, an insurer or a public body, the acceptable level of unresolved jurisdictional risk is zero. Willingness is not capability.

• Deepest Office integration, lowest rollout friction
• EU region = residency; US parent = CLOUD Act
• Own statement under oath: no guarantee against US access
• AI policy and Purview document use — they do not change the architecture