The U.K. government’s AI Opportunities Action Plan, published in early 2025, committed to embedding A.I. across public services at speed. NHS England has accelerated its digital transformation agenda. Dozens of local authorities are piloting generative A.I. for casework, procurement and citizen services. Internationally, the pattern is the same. But there is one problem with this acceleration. The data infrastructure being asked to power these systems was never designed for it.
The infrastructure crisis
Ask most public sector leaders whether they have data on their most vulnerable service users, and they will say yes. Ask them whether they can see all of it together—in a single coherent picture, in time to act before a situation escalates—and the answer almost universally changes.
This is not a new problem. But it is one that has become dramatically more consequential now that governments are attempting to layer predictive and generative A.I. onto the same fragmented estate. Legacy case management systems were built by separate departments, in separate decades, for separate purposes. They do not talk to each other. And the result is now the substrate on which A.I.-powered public services are being constructed.
Five years ago, the stakes of fragmentation were largely operational: slower decisions, missed referrals and siloed services. Today, they are structural. When an A.I. system trained on incomplete data is used to triage welfare cases, predict safeguarding risk or allocate NHS resources, the gaps in that data do not simply produce less accurate answers. They produce systematically biased ones.
What makes this newly urgent in 2026 is not just the pace of A.I. adoption, but the convergence of pressures driving it. Austerity-era cuts have hollowed out the human capacity that previously compensated for poor data—the social worker who knew a family, the physician who spotted the pattern. Generative A.I. is now being asked to fill that gap. At the same time, procurement timelines are shortening and political pressure to demonstrate A.I. progress is intensifying. The result is a race to deploy on unprepared foundations.
Spending on A.I., underinvesting in integration
Global public-sector A.I. spending is projected to exceed $25 billion annually by 2027, with the U.K. among the most ambitious investors relative to GDP. Yet the Government Digital Service and its successors have consistently struggled to secure sustained funding for the less glamorous work of data integration—the standardization of identifiers across departments, the governance frameworks for cross-agency data sharing, the plumbing of interoperability.
Other countries are making different choices. Estonia’s X-Road infrastructure, a secure data exchange layer connecting government systems, has been operational since 2001 and is now the backbone of its public A.I. applications. Finland and Denmark have invested heavily in federated data architectures that allow cross-agency insight without centralizing sensitive records. In contrast, the U.K.’s public sector remains behind.
What the organizations getting this right have in common
The organizations making progress are not, in most cases, the ones with the largest A.I. budgets. They are the ones who did the foundational work first.
Take the National Association for People Abused in Childhood (NAPAC). Their ability to expand support for survivors, influence policing policy and work alongside the Ministry of Justice did not depend on cutting-edge machine learning. It depended on a centralized dashboard: a single, coherent view of their data that gave them visibility they had not previously had. That visibility, not the algorithm, was the transformation. It is a model that many larger and better-resourced public bodies have yet to replicate.
Or consider Health Education England (now part of the NHS), which combined multiple data points with a machine learning model to predict which junior doctors were likely to leave their training programs, and to offer support before they walked out the door. The platform achieved over 60 percent predictive accuracy. But its more important lesson is systemic: the insight was only possible because the underlying data had been brought together with appropriate governance. Without that integration work, there was no signal to find. The A.I. was the final layer, not the foundation.
Both cases point to the same underlying truth: the data that would most help governments identify people at risk—the signals of financial instability, of mental health struggle, of domestic difficulty—already exists. It is not missing. It is fragmented, isolated across departments, and therefore invisible as a whole picture. In that gap, between what is held and what can be seen, people fall.
The structural shifts that will determine what comes next
The question is not whether governments should pursue A.I.-enabled public services. That decision has been made. The question is whether they can overcome the challenges they face in improving data availability, quality and the legal, privacy-based concerns that stand in their way.
Three structural shifts are required, and none of them is primarily technical.
The first is a reframing of data integration as infrastructure, not overhead. In the same way that no serious government would attempt to run a national rail network without agreed track standards, no government can responsibly deploy A.I. across public services without agreed data standards. The cross-departmental data-sharing frameworks currently being developed are a start, but they remain advisory rather than mandatory and are chronically underfunded relative to the A.I. procurement budgets above them.
The second is a shift in how risk is understood. At the moment, the dominant risk framework around public-sector A.I. focuses on the harms of misuse—the algorithm that discriminates, the system that violates privacy. These are real risks and deserve rigorous governance. But the harms of non-action—the early warning sign that was not visible because the data was not joined up, the family that was not supported because no system held a complete picture—are equally real. They are simply less legible to regulators and less visible to the public.
The third is a recognition that integration only works when the people whose data it concerns are part of designing it. The communities most likely to fall through fragmented systems—those experiencing poverty, homelessness, domestic abuse and mental health issues—are also the least likely to trust that joined-up data will be used for their benefit rather than against them. That trust gap is historically grounded. Closing it requires participatory design, genuine co-production and a demonstrated commitment to using integrated data to support people.
Those who get this right will have built something more durable than a set of A.I. tools. They will have built the institutional capacity to see their citizens clearly—and to act on what they see. Those who skip this step, racing to deploy A.I. on fragmented foundations, will have automated their blind spots. The difference will not show up immediately in procurement dashboards or ministerial announcements. It will show up in the people who were not reached in time.
