AI Sovereignty: What Breaks When You Build AI

In the rapidly evolving AI landscape, the concept of sovereignty has emerged as a critical factor, particularly for organizations operating in regulated sectors or those prioritizing data privacy and control. Bilge Yücel, Sr. DevRel at deepset GmbH, delivered a presentation titled "What Breaks When You Build AI Under Sovereignty Constraints" at AI Engineer Europe, shedding light on the complexities and considerations involved in developing Sovereign AI systems.

Understanding Sovereign AI

Yücel defined Sovereign AI as the ability of an organization to design, deploy, and operate AI systems on its own terms. This entails having explicit control over data flow, model choice, infrastructure, and operations. She broke down the concept into four key pillars: Data Sovereignty, Model Sovereignty, Infrastructure Sovereignty, and Operational Sovereignty.

The Four Pillars of Sovereign AI

Data Sovereignty governs how data is accessed and used in AI systems, emphasizing that data should be stored and processed within trusted jurisdictions to meet compliance requirements, and that access permissions must be respected. Yücel highlighted that for European citizens, data sovereignty often means data must remain within Europe, citing GDPR as a prime example.

Model Sovereignty refers to who controls the running models and the origin of the training data. This involves having the freedom to choose and switch models, ensuring swapability without architectural changes, and understanding the origin of the training data.

Infrastructure Sovereignty addresses where compute happens. Yücel presented a spectrum from maximum control (air-gapped, on-prem, no egress, EU AI Act safe) to maximum convenience (SaaS, full vendor, cloud act risk). Options like private clouds with GDPR compliance and EU-operated infrastructure fall in the middle, offering a balance.

Operational Sovereignty focuses on monitoring, evaluating, and managing AI systems over time. This includes monitoring AI systems in production, incorporating human-in-the-loop processes in high-stake environments, and managing versioning and updates in a controlled, auditable way.

Engineering Challenges and Solutions

Yücel then delved into the engineering challenges that arise when building sovereign AI systems. She illustrated these challenges by explaining what needs to be done and what breaks in existing systems when transitioning from cloud-based solutions to more sovereign, self-hosted models.

For instance, when replacing a frontier API with a self-hosted model, developers need to translate API logic to the new model architecture. They also need to manage multiple databases and instances, which can lead to issues with searchability and discoverability. Furthermore, moving private data into the required jurisdiction for compliance can create complexity.

The presentation showcased how a good orchestration framework, like Haystack, can help solve some of these problems. Haystack offers features such as a consistent interface, explicit data flow, serializability to YAML, and being truly open and extensible. This allows for greater control and flexibility in managing AI systems, particularly when dealing with diverse models and infrastructure.

Bringing it all Together with Haystack

Yücel demonstrated how these concepts can be integrated into a sovereign architecture using Haystack. She presented a pipeline that includes components like a tracer, a LangChain connector for LLM observability, and an agent. The agent is configured with a system prompt that defines it as a sovereign agent with access to multiple tools, including a custom component to connect to a self-hosted LLM. The pipeline also incorporates input and output guardrails for content moderation and safety, ensuring that potentially unsafe requests or outputs are handled appropriately.

The example code illustrated how to define tools, such as a knowledge base search and a PDF report generation tool, and how to integrate them into the agent. The confirmation strategies, such as "BlockingConfirmationStrategy" and "AlwaysAskPolicy," highlight the control mechanisms available for managing agent interactions and ensuring compliance.

Sovereignty Checklist

To help organizations assess their AI sovereignty, Yücel provided a checklist with key questions:

• Can you swap models without changing the application logic?
• Do you have reproducible run logs, stored in a compliant way?
• Can your team respond to an incident without calling a vendor?

These questions aim to prompt a deeper evaluation of an organization's control over its AI systems, from model flexibility and data lineage to operational resilience and vendor independence.

In conclusion, Yücel emphasized that building sovereign AI is a challenging but essential endeavor for many organizations. By understanding the four pillars of sovereign AI and leveraging robust orchestration frameworks like Haystack, companies can navigate these challenges and build AI systems that meet their specific needs for control, compliance, and operational autonomy.