If governments want to act effectively in the age of AI, they must first build the capacity to learn; continuously, reliably, and at scale, explains Johanna Ballesteros, Research Associate, Oxford Internet Institute.
Across every previous technological wave, the public sector has paid for lateness. During the early e-government era, most administrations focused on digitising paper forms instead of redesigning public services. In the agentic era, however, the bill arrives faster. As outlined in The Agentic State, authored by Luukas Ilves, Manuel Kilian, Simone Maria Parazzoli, Tiago C. Peixoto and Ott Velsberg, the cost of delay now exceeds the cost of early, thoughtful experimentation, leaving governments with a simple choice whether they actively shape the transformation or are shaped by it. Progress isn’t inevitable.
As Carl Benedikt Frey argues in his new book How Progress Ends, decentralisation drives discovery, but bureaucracy scales it. When institutions fail to adapt, progress stalls. That’s where my argument begins: to adapt, they must first have the capacity to learn.
Most bureaucratic systems were designed to ensure accountability and predictability in relatively stable environments, not for ever-changing conditions where speed and continuous adaptation become essential. Without institutional learning, they risk turning stability into stagnation.
Max Weber’s model of bureaucracy rested on a powerful idea: public service built on competence rather than patronage. But the kind of competence Weber described, based on technical skills acquired once through education and training, belonged to a slower age. In a world where the skills and capabilities of public servants shift with every advance in AI, that model becomes self-limiting. The modern equivalent of Weber’s meritocracy is not static expertise acquired once but a system that keeps learning alive. In that sense, learning as infrastructure is the next iteration of bureaucratic rationality, one that treats adaptability itself as a core competence.
From sporadic training to learning as infrastructure
This is about reprogramming the operating system of government for an age of continuous change, with learning as the function that keeps it running, improving, and connected. Learning as infrastructure means building the mechanisms through which knowledge flows as predictably as data or finance. It is embedded, budgeted, scheduled and measured like any other utility. It is interoperable, allowing credentials and content to travel across departments. And it is open by design, enabling shared standards and collective intelligence rather than bespoke silos.
We don’t need sporadic training programmes; we need a system that continuously feeds, connects and applies knowledge. Without that learning infrastructure, AI initiatives will remain confined to pilot projects and small showcases rather than reshaping how government works. The real value lies in transforming the core – tax, benefits, casework, procurement – where capability gaps translate directly into slower, costlier, less trusted services.
Building bridges of expertise
To stay at the frontier, governments must draw continuously on the best available expertise from the technology community, academia and applied research. That means treating knowledge exchange itself as part of the learning infrastructure, building mechanisms that connect those who advance technology with those who apply it in practice.
Singapore’s and the UAE’s large-scale AI training programmes illustrate what commitment looks like. Singapore’s mandatory AI literacy course for all public servants and the UAE’s national Chief AI Officers’ training programme signal whole of government capability-building. But the deeper lesson is structural. States need enduring interfaces with the communities driving technological progress, partnerships with universities, research labs and innovation ecosystems, so that foresight and practice evolve in step. As AI and other emerging technologies develop, governance models must evolve with them, through sustained collaboration between science, technology and public administration.
What public servants should be learning
Two broad capabilities matter. Technical literacy is now basic competence: understanding how digital systems work, how data and algorithms interact, and how technology shapes decisions and services. Not everyone needs to be a developer or data scientist, but everyone should grasp that they operate in a technology-mediated environment.
Institutional literacy is equally critical: how to buy, govern and steward technology responsibly. That includes contracting for AI and digital systems, managing risks and audits, and aligning incentives to public outcomes. Procurement, finance and oversight must evolve from static compliance to continuous assurance, shaping markets rather than simply checking boxes. Because influence is unevenly distributed, learning must extend across all levels of decision-making and delivery, equipping senior leaders to act strategically and enabling practitioners to translate those choices into adaptive practice, through formats tailored to their distinct contexts and responsibilities.
Why traditional training programmes won’t be enough
Classic retraining is episodic and backward-looking. As argued in AI and the Retraining Challenge, providers need to look ahead to the skills that will matter in the future, not merely to those required at present. The lesson for governments should be similar when it comes to continuous learning. That means combining foresight and experimentation with hybrid formats, micro-credentials and highly curated content, supported by evidence of what works, for whom and at what pace.
From my experience in public-sector learning practice, hybrid formats, curated pathways and strong administrative context consistently make the difference between short-lived training and lasting capability. To scale them, governments need learning infrastructure that makes such formats repeatable and connected. Ukraine’s Mriia ecosystem, a personalised AI-driven educational platform, shows how AI can become part of the infrastructure, not a separate tool.
A paradigm for a learning state
Learning as infrastructure is less about new platforms than about a new premise: adaptation must be a public-sector core function. Whether through shared systems, portable credentials or tighter feedback loops between research and administration, the goal is to embed learning into the architecture of government. Technology alone cannot make institutions adaptive. Developing learning capabilities is a question of state capacity and institutional resilience in an AI-driven world.
As Carl Benedikt Frey reminds us, progress is never automatic; it depends on institutions able to evolve with the technologies they adopt. The task ahead is to design a state that not only scales innovation but keeps learning from it, a state that learns faster, together.
About the author
Johanna Ballesteros is a Research Associate at the Oxford Internet Institute. She also leads the learning and upskilling unit at GovTech Deutschland.
