In Episode 166 of the AIAW Podcast, Magnus Enzell and Peter M Nordström from the Regeringskansliet, Finansdepartementet spoke about one of most ambitious public sector case studies I have seen this year: the introduction of more than thirty AI assistants across ministries. These agents now draft, summarize, search, and support decision processes inside Regeringskansliet which is the heart of Sweden’s government machinery.
It’s a quiet revolution. Instead of grand announcements, Regeringskansliet is building digital civil servants that operate behind the scenes, transforming how policy work, documentation, and administration unfold. What makes this initiative remarkable is not just its technical scope but its cultural realism. As Magnus put it, “We didn’t start with technology; we started with trust. Without that, no AI system can succeed in government.”
This article explores the deeper significance of Regeringskansliets approach on how agentic AI is changing the very structure of administrative work, what lessons this offers for other governments, and why the Swedish model may become a template for responsible AI adoption across the public sector.
A Government Laboratory for the Agentic Era
While many public organisations are still debating pilot projects, Regeringskansliet has already moved from prototypes to deployment. They now operate dozens of AI assistants supporting tasks such as document classification, policy analysis, and correspondence handling.
Peter Nordström described it succinctly: “Our idea was not to build a single super assistant, but many small ones that solve real problems.” That philosophy, a pragmatic modularity sets Regeringskansliet apart. Instead of chasing a single general-purpose platform, they created a constellation of narrow, purpose-built agents, each trained to understand the context of its host department.
This distributed model mirrors Regeringskansliets administrative culture, where agencies retain operational independence under shared governance frameworks. It also enables faster experimentation and safer iteration. “Every agent we deploy,” Peter noted, “has a human twin or someone accountable for its actions and outputs. That’s what makes it acceptable inside the organization.”
At a time when many governments still talk about AI in abstract terms,Regeringskansliet’s system already processes real documents, supports real decisions, and helps real people. Yet it does so within strict boundaries that maintain legitimacy and oversight.
Beyond Automation: Toward Augmented Governance
Magnus Enzell framed the effort not as automation but as augmentation. “We wanted to free human time, not replace human judgment,” he said. This distinction which is subtle but profound it runs through the entire project.
The AI assistants handle the repetitive, high-volume tasks that previously consumed hours: reading long reports, extracting summaries, cross-referencing regulations, and drafting internal memos. Civil servants then focus on the interpretive and deliberative parts of their roles – the work that demands reasoning, empathy, and political sensitivity.
Henrik Göthberg, supported the digital innovation described it as “a shift from clerical to cognitive public service.” The aim is not efficiency for its own sake, but the creation of institutional capacity and the ability to respond quickly to new issues without expanding bureaucracy.
Anders Arpteg, added a critical nuance in the episode: “The hardest part isn’t data or code. It’s helping people see AI as a colleague, not a competitor.” That human transition from psychological and cultural is as central as any technical milestone.
The Regeringskansliets Way: Governance First, Technology Second
What is most novel about Regeringskansliet’s approach is the order of operations. Governance precedes technology. Before an AI assistant is built, the team defines its legal footing, its boundaries, and its human supervision model. Every deployment is accompanied by documentation that explains data access, audit trails, and escalation processes.
This approach reflects the country’s long administrative tradition of transparency and accountability. Sweden was the first nation in the world to introduce a Freedom of Information law, in 1766. The same ethos now shapes its AI policy.
Peter explained it clearly: “Trust in government is our most valuable asset. Every new system has to earn it.” That perspective informs how Regeringskansliet manages data, logs every AI action, and builds explainability into interfaces. Each assistant must be able to justify its reasoning, show the sources it used, and indicate uncertainty when confidence is low.
In a world where most governments fear that AI will erode accountability, Regeringskansliet is demonstrating how digital systems can deepen it by making processes more transparent and auditable than before.
The Architecture of Coexistence
A striking theme throughout the conversation was coexistence. The Regeringskansliet AI agents are not hidden behind walls of code; they operate alongside people. Each department has designated stewards who supervise their agents, test them, and report anomalies.
This design choice is both cultural and technical. It builds ownership and demystifies the technology. “We treat the agents as part of our teams,” Magnus said. “When they fail, we talk about it. When they improve, we share what worked.”
That language, “part of our teams”, reveals how deeply integration has already gone. By normalizing the presence of AI assistants, Regeringskansliet is cultivating a hybrid workforce where human and machine collaboration becomes everyday practice.
Observers often underestimate how radical this shift is. For centuries, public administration has relied on hierarchical workflows and written procedures. Now, a machine can draft a document, analyze a policy, or summarize a meeting faster than a person but only if the system is allowed to act within clearly defined boundaries. The challenge is not capability, but choreography.
Building Trust through Visibility
Transparency emerged as the most powerful design principle. Every action an AI assistant takes inside the Swedish Government Offices is logged and reviewable. Managers can trace outputs to inputs, identify which version of a model was used, and see who approved each result.
This kind of traceability might sound bureaucratic, but it is the foundation of legitimacy. It transforms AI from a black box into a glass one. “If you can explain what happened, people stop being afraid,” Peter observed.
Regeringskansliet’s transparency framework is already influencing other European administrations exploring generative AI. Its design anticipates the EU AI Act, which will soon require public institutions to document and disclose how AI systems operate. In this sense, Regeringskansliet is not only adapting to the new legal order rather it is helping define it.
From Pilots to Platforms
What began as small departmental experiments is now evolving into a shared infrastructure. The government is building reusable components or connectors, audit services, and model governance layers that allow new agents to be developed safely and quickly.
Magnus described this shift as “moving from pilots to production.” Once an assistant proves reliable in one ministry, the framework allows others to adopt and adapt it without starting from scratch. This accelerates innovation while maintaining control.
The result is a new kind of digital platform. It is not centralized in a single system, but federated across ministries, bound by common standards. It mirrors Regeringskansliets model of decentralized yet coordinated governance, which may turn out to be exactly what the agentic era requires.
Change Management as the Real Innovation
If there is a single thread running through this episode, it is that technology was never the hardest part. Change management was.
Civil servants had to learn how to collaborate with digital colleagues, how to question AI outputs without fear, and how to see the value of augmentation rather than replacement. The leadership team invested heavily in communication, workshops, and co-creation.
Magnus reflected on this cultural shift: “We asked people what tasks they wished they didn’t have to do. The answers were remarkably consistent. That’s where we started building.” The result was an adoption curve driven by genuine need, not top-down mandates.
By aligning technology with everyday pain points, the Government Offices created pull, not push. Staff began requesting new assistants rather than resisting them. That reversal of psychology from fear to demand may be the most replicable lesson of all.
A Playbook for Deploying Agentic AI in Government
Step 1: Frame a high-impact, principled pilot domain
Start with a well-scoped domain where value is tangible and risks are contained. This could be document summarization, memo drafting, or internal case triage. Sweden began in departmental bottlenecks before expanding. Select areas where errors are manageable and stakeholder buy-in is achievable.
Step 2: Co-design governance around the agent
From the outset, define how oversight, escalation, and compliance will function. Decide who can override the agent, how actions are logged, and how legal obligations are met. Research shows that agentic AI requires continuous supervision rather than periodic review. Embedding governance into daily operations, rather than treating it as an external function, is key.
Step 3: Conduct rigorous risk assessment
Before launch, perform a comprehensive risk audit that examines technical, legal, data, and ethical dimensions. Address issues such as model drift, privacy, bias, and accountability. The unique risk with agents is that they act autonomously, so safeguards must be stronger than those for conventional AI systems.
Step 4: Design human-agent collaboration carefully
Define when and how control switches between AI and human operators. Agents should propose and humans should decide. When uncertainty arises, escalation must be automatic. Magnus and Peter emphasized that trust depends on clarity: civil servants must always know where responsibility lies and how to intervene.
Step 5: Integrate gradually with existing systems
Rather than rebuilding IT from scratch, add agents as layers on top of current infrastructure. This lowers costs and accelerates learning. Most government IT environments are legacy-heavy, so incremental integration is often the only practical path. Sweden’s success rested on this pragmatic approach.
Step 6: Prioritize interpretability and auditability
Every action an agent takes should be explainable. Systems must record data sources, reasoning steps, and human overrides. These logs serve accountability, help trace errors, and reinforce public trust. Without transparent audit trails, agentic AI in government will stall.
Step 7: Measure continuously and build feedback loops
Implement metrics to track accuracy, speed, human overrides, satisfaction, and system drift. Use this data for improvement. Sweden’s team built feedback into daily use, allowing them to detect issues early and refine performance over time.
Step 8: Scale through modularization
Once pilots prove their value, expand through modular agent functions that can be orchestrated together. This approach manages complexity and enables reuse across departments. However, as scale grows, governance must evolve to handle new interdependencies.
Step 9: Build institutional capacity and culture
AI agents change the nature of work. Staff transition from executors to supervisors and designers. Invest in training, cross-functional teams, and a culture that values experimentation and learning. Governments must accept that small failures are part of progress.
Step 10: Maintain transparency and citizen trust
Legitimacy is the cornerstone of public sector innovation. Citizens must understand how AI systems are used, what decisions they influence, and how those decisions can be challenged. Communicate openly through public documentation and clear explanations. Transparency is not a risk—it is protection.
Lessons for Governments Everywhere
Regeringskansliet’s case study offers a set of insights that transcend context. Start small, but start with governance. Make transparency visible and continuous. Keep humans in control, not just in the loop. Invest in trust as deliberately as you invest in technology.
Perhaps the most novel insight is that agentic AI can strengthen, not weaken, the social contract if it is deployed under conditions of clarity and accountability. In Regeringskansliets case, each AI assistant operates within a clear mandate, supervised by a named individual, and reviewed through auditable logs. It turns automation into a new layer of public service rather than a threat to it.
Peter concluded with a quiet optimism: “The value isn’t just in what the agents do. It’s in what they make possible. A government that learns faster, reacts faster, and works more humanely.”
The Broader Horizon
Agentic AI in government is still a new frontier, but Regeringskansliet’s work shows what a credible path looks like. It is an incremental, accountable, transparent modernization of governance. It demonstrates that AI can coexist with democratic values and perhaps even reinforce them.
As other nations watch this experiment, one truth becomes clear: the agentic era will not wait for perfect conditions. Institutions must build the capabilities, governance, and trust structures now. Regeringskansliet’s journey is proof that it can be done without breaking public trust or bureaucratic stability.
In the words of Magnus Enzell, “We don’t need to automate democracy. We just need to make it work better.”
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, listen to the full episode and make your own learnings.
