In the quiet corridors of Sweden’s Central Bank (Sveriges riksbank) , a transformation is unfolding that mirrors a wider movement across the world’s public institutions. It is not about replacing economists with algorithms or chasing the latest model release. It is about building a sovereign foundation for intelligence, a deliberate rethinking of what buy or build means in the age of artificial intelligence.
When Hugi Aegisberg joined the Sweden’s Central Bank as Innovation Manager and Product Owner for AI, he stepped into one of Europe’s most tightly governed environments. The bank is both the guardian of financial stability and the steward of a nation’s data. The constraints are formidable, with strict privacy rules, geopolitical considerations, and a cautious public-sector culture. Yet it is within these constraints that Aegisberg and his team are redefining how to create resilient, secure, and sovereign AI infrastructure.
The mantra guiding this transformation is simple yet profound. Buy to Build.
The following is a summary of AI After Work (AIAW) Podcast episode 168 where Hugi was a guest.
From Pilot Hell to Platform Thinking
In most large organizations, public or private, AI begins with pilots. A proof of concept is launched with great enthusiasm. A model is trained. A dashboard is built. Then progress slows. Integration stalls. Security reviews begin. The model never makes it to production. The graveyard of prototypes grows deeper.
Aegisberg calls it pilot hell. “Everything just works with this one pipeline that cannot really be repeated any meaningful number of times,” he explained during the conversation. The Sweden’s Central Bank was determined to avoid that fate. The answer lay not in more experimentation but in infrastructure maturity.
Long before its AI program, the bank had already transformed how it worked with data. Teams had moved away from centralized IT ownership toward a federated model where analysts and business experts took direct responsibility for data quality and orchestration. That cultural and architectural groundwork created the perfect foundation for AI.
By the time the AI roadmap arrived, the Sweden’s Central Bank already had the plumbing, pipelines, and governance needed to move beyond pilots. “We are actually very well positioned,” Aegisberg noted. “Our data maturity is in the top tier for public sector organizations. That lays good foundations for what we’re about to do.”
Building a Secure AI Infrastructure
In 2025, the Sweden’s Central Bank is building an on-premise AI infrastructure, a deliberate decision in an era dominated by cloud platforms. For most organizations, the cloud is synonymous with speed and scale. For central banks, it raises questions of sovereignty and security. Sensitive financial data cannot simply be sent to a third-party server overseas.
Yet on-premise no longer means outdated. The ecosystem has matured. “Companies providing these platforms have become better at building infrastructure that just works,” Aegisberg explained. “You can use NVIDIA’s pre-built containers, or blueprints that make deployment much easier than before.”
This shift matters because it changes what small and medium organizations can realistically achieve. Five years ago, setting up GPU clusters and orchestration frameworks was a major engineering challenge. Today, containerization, virtualization, and improved GPU partitioning make it feasible for a 500-person institution to host serious AI workloads internally.
The key principle is flexibility. Instead of rigid hardware allocation, the bank is using time-shared GPUs that allow multiple workloads to coexist dynamically. “It means you can oversubscribe your GPUs eight times compared to static allocation,” Aegisberg said. “It’s about giving people the creative sandbox they need to explore.”
That sandbox is not just a technical construct. It is a governance model. Analysts and economists can experiment safely within defined boundaries. Security teams can vet containers once and trust them across the system. Procurement teams can contract for hardened components rather than building everything from scratch. It is a platform for safe innovation.
Buy to Build The New Logic of AI
For decades, IT strategy revolved around the make-or-buy decision. In the age of enterprise AI, that logic breaks down. “People kept asking me whether we should buy or build,” Aegisberg said. “And I got tired of the dichotomy. Buy to build is what we really do.”
Buying to build means purchasing the right layer of hardened, well-maintained components and assembling them into systems tailored to your needs. It is about flexibility and sovereignty, not wholesale outsourcing.
In traditional software procurement, buying meant adopting a complete product such as ERP, CRM, or data warehouse. In AI, no such finished products exist. Every deployment involves integration with data sources, security processes, and business logic. Even a pre-trained model requires fine-tuning, monitoring, and adaptation.
That reality challenges the procurement mindset in both public and private sectors. “Procurement is still trying to push make-or-buy into the data and AI space incorrectly,” said Henrik Göthberg. “They need to understand that whatever you do with data and AI will require some build.”
Aegisberg agreed but added nuance. The role of procurement is evolving, not disappearing. “Our procurement team gets it,” he said. “They might not know what Kubernetes is, but when I explain it, they understand. And thanks to organizations like the Cloud Native Foundation and OWASP, we can point to established standards rather than reinventing them.”
OWASP, the Open Worldwide Application Security Project, is a quiet hero of the AI era. It compiles lists of known vulnerabilities and best practices for securing new technologies, including large language models. For risk teams demanding assurance, OWASP offers ready-made checklists. “You already have the answer when someone asks, have you thought about the risks of LLMs?” Aegisberg said. “Yes, here’s the list. We’ve gone through it.”
The Sovereignty Imperative
Behind all of this is a question that is reshaping AI strategy across Europe. Who controls the intelligence layer of society? For public institutions, sovereignty is not a buzzword. It is an operational necessity. Sensitive economic models, payment systems, and forecasting algorithms cannot be dependent on opaque foreign systems.
The Swedish Central Bank’s hybrid strategy reflects this balance. Cloud solutions will be used selectively, but core analytical capabilities will remain in-house. The goal is not isolation but resilience.
This mirrors a broader European movement. Projects like SVEA, coordinated by AI Sweden , are building public-sector digital assistants trained on Swedish data. The Central Bank of Sweden is part of this initiative, not because it wants a chatbot, but because it sees the importance of national language infrastructure. As Aegisberg put it, “They are annotating Swedish public-sector language. That could lead to strong mid-sized language agents in the future, and I’m all for that.”
Sovereignty also extends to data architecture. The Swedish Central Bank’s data lake is managed internally, with clear accountability for data integrity resting in the business units, not in IT. This model of business-owned data with centralized enablement is becoming a template for other public organizations. It supports innovation while maintaining control.
From De-Centralization to Enablement
One of the most striking parts of the Swedish Central Bank´s AI approach is its organizational model. Aegisberg’s AI team is not a central service provider building models for everyone else. It is a platform and an enablement team. “It’s a hub-and-spoke model,” he explained. “Our job is to build the platform and inspire use cases. But the competence must grow in the organization.”
The goal is not to centralize expertise but to distribute it responsibly. Economists and analysts who once worked in Excel have already moved to Python notebooks. The next step is for them to experiment with AI tools in a secure, supported environment.
Training and outreach are essential. The AI team educates departments on the kinds of talent they need, from data engineers to DevOps specialists, to bring AI ideas to life. Governance is implemented through automated guardrails rather than manual gatekeeping. “You build it, you run it,” Aegisberg said. “We provide the hardened scaffolding you need.”
This approach reflects a broader trend across leading enterprises, moving from command-and-control to empowerment. It requires trust in people, robust infrastructure, and a clear security framework, the very pillars of the buy-to-build philosophy.
The First Use Cases. Practical AI for Public Good
For all the talk of agents and architecture, the first AI deployments at the Swedish Central Bank are pragmatic. One of the earliest priorities is transcription. In a public institution where meetings and interviews may contain sensitive information, cloud-based transcription is off-limits. An on-prem solution saves hundreds of staff hours per year without compromising confidentiality.
Other early cases include internal document summarization, data-lake exploration, and support for recurring report generation. “We have tons of repetitive work,” Aegisberg said. “Having a template informed by previous reports is a huge help.”
These are modest beginnings, but they lay the groundwork for more advanced workflows. Over time, AI assistants will help staff locate information across vast datasets, check consistency, and even recommend analytical methods. Each step builds confidence, competency, and cultural readiness.
The Emergence of Safe Autonomy
The conversation naturally turned to agents, one of the most hyped and misunderstood concepts in modern AI. Aegisberg was careful to draw a line. “When I say agent, I mean a system that takes autonomous steps beyond simply selecting between options,” he said. “Otherwise, it’s just software.”
That distinction matters. Within regulated institutions, the goal is not full autonomy but safe delegation. “We might start with humans in the loop,” Aegisberg explained. “Agents could support internal service tasks, triaging issues or doing pre-analysis, but not making decisions.”
The future of agency in finance may arrive in small increments. Researchers are already exploring how autonomous systems could manage liquidity, monitor payments, or flag anomalies. The benefits are clear: greater speed, consistency, and auditability. Yet in environments where data leakage or false positives carry real economic risk, caution is a virtue.
As Göthberg observed, it is an emergent stance. You automate safely, step by step. In that sense, the evolution of AI agents may mirror how DevOps transformed IT, a gradual embedding of intelligence into operational workflows rather than a leap to full autonomy.
The Public Sector as an AI Laboratory
The irony of AI in government is that while the private sector races ahead with commercial models, it is often the public sector that grapples most deeply with the long-term implications. Central banks, tax authorities, and national data agencies sit at the intersection of technology, law, and public trust. They cannot afford to move fast and break things.
That constraint has a hidden advantage. It forces rigor in governance, architecture, and ethics. The result is a model that enterprises can learn from, hybrid infrastructure, transparent standards, distributed capability, and human accountability.
Across Europe, similar approaches are taking shape. Finland’s Silo AI is partnering with AMD on sovereign infrastructure. France and Germany are investing in language models aligned with local regulations. The EU’s AI Act, despite its complexity, is pushing organizations to understand and document their model risks.
This environment may slow deployment, but it accelerates maturity. By the time public institutions move to production, they have built the scaffolding of trust that the private sector is now struggling to retrofit.
The Hidden Power of Standards
Behind every success story in AI deployment lies an ecosystem of open standards. The conversation at the Swedish Central Bank highlighted several, including OWASP for security, the Cloud Native Foundation for containerization, and emerging protocols such as the Model Context Protocol, or MCP, for connecting language models to data sources and tools.
Aegisberg described MCP as a successor to naive retrieval-augmented generation. Instead of merely searching static documents, MCP agents can interact with systems and reason across contexts. They bring structure to the growing complexity of AI architectures.
The rise of MCP and similar frameworks signals a broader trend. AI is moving from experimentation to engineering. Enterprises are beginning to design their AI ecosystems with the same discipline once reserved for networks and databases. Standards provide the common language that makes collaboration possible across vendors, regulators, and industries.
Lessons for Enterprise Leaders
What can executives in large enterprises learn from the Swedish Central Bank’s journey?
- Start with data maturity. Before training models, build trust in your data, and place ownership where it belongs, with the teams who use it.
- Treat infrastructure as strategy. Decisions about cloud, on-prem, and hybrid architectures define your sovereignty, cost structure, and ability to innovate safely.
- Replace make-or-buy with buy-to-build. Source hardened, well-supported components, but retain the ability to shape them into solutions that fit your business.
- Build for enablement, not control. A central AI team should be an accelerator, not a bottleneck. Provide safe sandboxes, reusable components, and education so that competence grows across the organization.
- Progress responsibly toward autonomy. The road to AI agents and automated decision-making should be incremental, auditable, and aligned with human accountability. Innovation without governance is fragility.
The Broader Implications of AI as a General-Purpose Capability
Perhaps the most important insight from the Swedish Central Bank’s approach is conceptual. AI is not a project or a product. It is a general-purpose capability, akin to electricity or the internet. As Aegisberg put it, “At some point, you learned to use Word and Excel. Welcome to the future. Now you need to learn to use this.”
That mindset reframes AI from a specialized domain to a basic skill. It democratizes innovation while anchoring it in infrastructure and governance. It also underscores why sovereignty and security are not optional extras. In a world where every organization becomes an AI organization, the ability to build safely and independently becomes a source of strategic resilience.
The story of Sweden’s Central Bank is, in essence, the story of every enterprise trying to modernize responsibly. It shows that progress is not about racing to deploy the latest model but about constructing the durable scaffolding that makes intelligence sustainable.
Conclusion
The age of artificial intelligence is testing the foundations of how organizations procure, govern, and trust technology. The public sector, long dismissed as slow-moving, may be offering the blueprint for doing it right.
Buy to build. Secure the core. Enable the edge. Educate your people. Evolve carefully toward autonomy. These are the new principles of enterprise AI.
As Aegisberg’s experience at Sweden’s Central Bank demonstrates, sovereignty is not isolation, and compliance is not obstruction. They are the architecture of trust. And trust is the only infrastructure on which intelligent systems can truly scale.
This article was enhanced with the help of AI tools, drawing on the podcast transcript and complementary online research. To go deeper into the source material, I encourage you to listen to the full episode and make your own learnings.
Listen, or see the full episode here: https://aiawpodcast.com/
About the podcast
The AI At Work podcast is a leading forum for meaningful discussions on how artificial intelligence transforms organizations, industries, and public institutions. Produced by Hyperight AB and Goran Cvetanovski, hosted by Anders Arpteg and Henrik Göthberg , it connects research, technology, and strategy through conversations with global innovators who are shaping the responsible future of AI. Each episode offers grounded insights for leaders navigating the shift from experimentation to enterprise-scale adoption.
