Why Public-Sector AI Depends on Better Data Integration

UK Government An app for AI opportunitiespublished in early 2025, committed to embedding AI in all public services at an accelerated pace. NHS England has accelerated its figures digital transformation agenda. Many local authorities are exploring generative AI for casework, procurement and citizen services. Internationally, the pattern is the same. But there is one problem with this acceleration. The data infrastructure required to run these systems it is not designed for it.

Infrastructure problem

Ask most public sector leaders if they have data on their most vulnerable service users, and they’ll say yes. Ask them if they can see it all together—with one unified image, before the situation escalates—and the answer is almost universally different.

This is not a new problem. But it’s one that’s even more important now as governments try to put predictive and generative AI into the same segregated space. Probate programs were developed by different departments, in different decades, for different purposes. They don’t talk to each other. And the result is now an underground space where social services powered by AI are being built.

In the past five years, the stakes of the division have been high: slow decisions, missed transfers and closed services. Today, they have a structure. When an AI system trained on incomplete data is used to assess social cases, predict risk to protect or allocate NHS services, gaps in that data simply do not produce less accurate answers. They produce biases in a systematic way.

What creates this new urgency in 2026 is not just the pace of AI adoption, but the convergence of pressures driving it. Austerity-era cuts shut down the human power that previously compensated for bad data—the social worker who knew the family, the doctor who saw a pattern. Generative AI is now being called upon to fill that gap. At the same time, lead times are shortening and political pressure to demonstrate AI progress is increasing. The result is a race to invest in unprepared foundations.

Spending on AI, investing less in integration

Global public sector AI spending is expected exceed 25 billion annually by 2027and the UK is among the most ambitious investors relative to GDP. Yet the Government Digital Service and its successors have consistently struggled to secure ongoing funding for the less glamorous work of data integration—the establishment of identifiers across departments, governance structures for interagency data sharing, the creation of interoperability pipelines.

Other countries make different choices. In Estonia X-Road infrastructurea secure data exchange layer that connects government systems, has been in operation since 2001 and is now the core of its public AI applications. Finland and Denmark have invested heavily in creating integrated databases that allow for cross-agency understanding without aggregating sensitive records. In contrast, the UK public sector lags behind.

What organizations that receive this privilege have in common

Organizations making progress are not, in most cases, those with large AI budgets. They are the ones who did the groundwork first.

Take the National Association for People Abused in Childhood (NAPAC). Their ability to increase support for survivors, influence police policy and work with the Department of Justice did not depend on advanced machine learning. It depended on a central dashboard: a single, consistent view of their data that gave them visibility they didn’t have before. That appearance, not the algorithm, was the change. It is a model that many larger and better resourced civil society organizations have yet to implement.

Or consider Health Education England (now part of the NHS), which has combined multiple data points with a machine learning model to predict which junior doctors are likely to leave their training programmes, and provide support before they walk out the door. The platform achieved a prediction accuracy of over 60 percent. But its most important lesson is programmatic: understanding was only possible because basic data was combined with proper management. Apart from that integration work, there was no sign to be found. AI was the last layer, not the foundation.

Both cases point to the same basic truth: the data that can most help governments identify vulnerable populations—signs of financial instability, mental health struggles, domestic difficulties—already exist. It is not lacking. It’s divided, divided across the departments, so it doesn’t look like the whole picture. In that gap, between what is held and what cannot be seen, people fall.

Structural shifts will determine what comes next

The question is not whether governments should pursue AI-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 legal, privacy-based concerns that hold them back.

Three building shifts are required, and none of them are technical.

The first is to reframe data integration as infrastructure, not overhead. In the same way that no serious government would attempt to run a national rail network without agreed standards, no government would deploy AI across public services without agreed data standards. The cross-departmental frameworks for data sharing that are currently being developed are a start, but they remain advisory rather than mandatory and often have relatively little money compared to the AI ​​procurement budget above them.

The second is to change the way risk is understood. Currently, the dominant risk framework around public sector AI focuses on the harm of misuse—a discriminatory algorithm, a system that violates privacy. These are real dangers and deserve strong governance. But the harms of inaction—an early warning sign that went unnoticed because the data wasn’t compiled, a family that went unsupported because no system captured the whole picture—are just as real. They are not easily readable by administrators and are not visible to the public.

The third is to recognize that integration only works when the people whose data it affects are part of designing it. The communities most likely to fall into disparate systems—those dealing with poverty, homelessness, domestic abuse and mental health issues—are also the most likely to trust that aggregated data will be used to their advantage rather than against them. That trust gap is rooted in history. 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 be creating something that will outlast the AI ​​toolset. They will have built the institutional capacity to see their citizens clearly—and act on what they see. Those who skip this step, rushing to install AI on different platforms, will have created their own blind spots. The difference will not immediately appear on procurement dashboards or ministerial announcements. It will come from people who are not reached in time.