European Commission x deepset: Building Europe's Sovereign AI Platform

At a glance (as of March 2026):

• 68 builders onboarded and activated 
• 29 DGs currently running pipelines 
• 316 pipelines in staging or production 
• 1 unified sovereign platform

The Commission's early AI pilots worked. That created a new problem.

The European Commission's first generative AI project, GPT@EC, was a conversational AI assistant trained and run on the Commission's own data and systems. It proved that AI could add value across the institution. But success exposed a gap: there was no underlying platform to build, govern, and scale AI beyond a single application.

‍

Departments across the Commission began building their own AI tools independently, without a way to reuse components or govern what was being built. For an institution of 32,000 civil servants, strict data classification requirements, and binding EU regulation, that fragmentation was unsustainable.

‍

Three specific constraints made the problem harder than it would be for most organisations.

1. Sensitive Non-Classified (SNC) data cannot leave EU-controlled infrastructure or be exposed to non-EU jurisdictions. For SNC data, public cloud was off the table.
   
   
2. Regulatory compliance at institutional scale requires audit trails, explainability, access controls, and human-in-the-loop validation across every deployment. EUDPR (Regulation 2018/1725), the EU AI Act, and the Commission's internal AI rules all apply.
   
   
3. The Commission required a platform accessible to both technical and non-technical staff. The solution needed to empower AI developers and engineers while enabling domain experts to build AI applications, all without sacrificing governance or security.

The solution: a sovereign, model-agnostic platform built with deepset

To move from experimentation to a stable, governed AI ecosystem, the Commission partnered with deepset to jointly develop the self-hosted AI@EC Platform, based on the Haystack Enterprise Platform.

‍

The platform was co-developed with DG DIGIT, the Commission's Directorate-General for Digital Services. This was a deliberate choice. AI@EC was not procured as an off-the-shelf solution. deepset and EC teams worked side by side on architecture, integration with DIGIT data centres, API Manager integration, security, and guardrails.

‍

The platform runs entirely on Commission-managed infrastructure. It provides a centralized, self-service environment where departments can build, test, deploy, and govern generative AI agents and applications without coding. Teams can build AI applications using templates for use cases like agents, knowledge management, document summarization, chatbots, and multilingual policy analysis. The platform enforces compliance with EC operating standards automatically. Ready-to-use building blocks, including models, guardrails, and prompt templates, mean teams can go from idea to working prototype in days.

‍

The technical backbone is deepset's Haystack Enterprise Platform, built on Haystack, Europe's leading open-source AI framework. The open-source foundation gives the Commission full visibility into every component (e.g., retrievers, models, guardrails, agents) as well as the ability to easily develop their own custom components. They can inspect, modify, extend, or replace anything. If something needs to change, they can do it without replatforming or pulling applications out of production. The modular architecture means departments adopt the corporate building blocks they need, at their own pace, with flexibility to reuse AI architectures and evolve them over time.

‍

"With the AI@EC Platform, we can manage Generative AI projects in a more structured and coherent way, monitor key performance indicators, identify issues early, and reuse components across a wide range of use cases."

Emmanouil Maragkoudakis, Team Leader, Data & AI Professional Services (DIGIT.B.1)

‍

Governance and compliance are structural, not bolted on

The Commission handles sensitive policy data under some of Europe's strictest digital regulations. Governance is embedded at every layer of the platform.

‍

An AI Governance Board of legal, IT, and policy experts reviews all AI projects for regulatory compliance. The Haystack Enterprise Platform supports this with built-in testing and validation tools that let subject matter experts review outputs from deployed pipelines, the gold standard for AI quality assessment.

‍

Automated guardrails monitor both inputs and generated outputs in real time, flagging issues like bias, misinformation, or data leaks. All interactions are logged for audit trails. The guardrails are powered by AISS, a service originally developed by the Joint Research Centre and now operated by DIGIT's B1 unit, integrated into the platform as an API call within each pipeline.

‍

The platform gives users and auditors transparency into how AI applications reach their results, so decisions can be traced, challenged, and corrected.

‍

All SNC data processing stays within EC-managed infrastructure, with EU Login authentication, role-based permissions, and automated redaction of sensitive information. The result is a platform the Commission can trust and control.

‍

"The AI@EC Platform is leading by example, demonstrating that it is possible to deploy reliable and scalable AI solutions in a transparent and ethical way, in full compliance with the Commission's rules on AI, data protection and security."

Roberto Sgnaolin, Deputy Head of Unit, Data, AI & Web (DIGIT.B.1)

‍

How the partnership works

Partnerships are force multipliers. deepset brings open-source AI expertise and startup agility. The Commission brings the rigour and operational complexity of one of Europe's largest institutions. Together they built the platform iteratively across four phases, with each step strengthening all building blocks based on feedback.

‍

The collaboration emphasized reproducible DevOps/GitOps pipelines and extensive documentation so the Commission can operate and evolve the platform independently. Early adopter teams across departments are actively co-designing the platform, providing feedback on usability, templates, and documentation that shapes each release. That growing internal community gives the Commission the foundation to expand AI@EC into use cases across policy, legislation, and citizen services.

‍

“Working inside the Commission's infrastructure taught us a lot about what sovereign AI actually looks like in practice. Every pipeline had to be reproducible, every component auditable, every deployment documented. That rigor made the platform better.”
‍
Oliver Guggenbühl, Solutions Engineer, deepset

‍

What comes next

The AI@EC Platform is live and onboarding new users across the Commission. The first major milestone is the migration of GPT@EC, the Commission's conversational AI assistant, onto the platform. The original version was built as a standalone project. Moving it onto AI@EC means it can be governed centrally, improved continuously, and scaled across the institution rather than maintained as a fixed build. GPT@EC is one of dozens of use cases planned for the platform.

‍

Beyond that, the roadmap extends the platform's building experience for both technical and business users, with new features designed to lower the barrier further. A broad set of new data sources will also be connected to GPT@EC, expanding what staff can access through the assistant and making it more useful across departments.

‍

The modular architecture means the Commission can integrate new capabilities as needs evolve, whether that's new AI models, updated guardrails in response to regulatory changes, or entirely new use case categories.

‍

“The European Commission needed a platform that could meet the most demanding sovereignty and compliance requirements in Europe. Building AI@EC with the Commission showed what becomes possible when you pair open-source foundations with the governance rigour of a major public institution. This is how AI will be adopted at scale in Europe.”
‍
Milos Rusic, CEO, deepset