Building AI on solid ground: why data foundations matter for government transformation

The UK public sector is under intense pressure to modernise, deliver better services, and harness the potential of AI, all against a backdrop of tight budgets and political urgency. But while attention often focuses on new tools and front-end services, the biggest obstacle to progress remains much less visible.

“The silos that existed still exist,” says James Underhill, Senior Government Account Director at Cloudera. “The move to the cloud, in some cases, is creating new silos as well.” While cloud adoption has brought flexibility and a wave of new capabilities, it has also multiplied the complexity of managing and integrating ever-growing volumes of data.

Underhill draws on a simple but telling analogy: building a house. “The foundations are probably the most boring part, but they’re one of the most important. Without them, the house is likely to collapse.” For him, preparing data to be AI-ready is just that: critical, often unseen work. It’s what determines whether an organisation can make reliable, trustworthy decisions in an AI-driven world.

And trust, he argues, is the lynchpin. “If you don’t know where that data is coming from, what happened to it, and why a decision was made - and if you can’t bring the public along with you, then why are you doing it?”

When cloud creates new silos

Despite a concerted effort over the past decade to break down departmental silos, fragmentation persists, and in some cases, has been reinforced by the shift to cloud. While delivery teams are keen to adopt AI to improve services, data teams are often left managing the underlying integration, governance, and security issues.

This tension can easily lead to what Underhill describes as ‘shadow IT’ where teams bypass central IT to get work done faster, but at the expense of governance and oversight. “If you stop people doing something, they may just choose to do it themselves. That leads to even more silos, no governance, no security.”

The answer, he says, is not to clamp down, but to enable innovation within clear guardrails. “The IT team should be an enabler rather than a blocker, but there has to be some control, whether that’s from a legal point of view or spend. Someone is ultimately going to have to be in control of that.”

For Underhill, the challenge is to give delivery teams the tools they need, without creating an unmanageable patchwork of systems and datasets. This means embedding robust governance and security from the outset not as an afterthought.

“Having robust security, governance, lineage across everything that’s being done is very important,” he says. “If someone creates something really valuable, you want to trace that back to see how it was created.” That visibility not only protects against risk, but also ensures that innovation can be scaled and replicated across teams.

In a world where civil servants increasingly expect the same seamless experience they get from their bank or online retailer, the pressure to modernise is only growing. “People in government know that if they interact with their bank, it will be seamless. So why can’t it be in government? Why can’t they use the latest tools, why can’t they use AI to do what they need to do?”

One source of truth

Cloudera’s answer to the fragmentation problem is to maintain a single, authoritative dataset, and bring tools and models to that data, rather than moving or duplicating it.

“Our view is: have one copy of the data, one single source of truth, and bring tools, bring models to that data,” says Underhill. “Don’t copy it, don’t ship it off to somebody else to do something on, you’ll incur a cost of doing so, and who knows what’s been done to that copy?”

This approach, he argues, offers multiple advantages: lower costs, stronger security, and better collaboration. “Even within an organisation, it’s difficult. You might have a data science or an AI team wanting to do something but not knowing how to get access to the data they need, they may not even know it exists.”

As data volumes grow and the pace of change accelerates, Underhill believes that adopting modern architectures will be essential. “The problem is only going to get worse unless modern architectures are adopted.”

While this approach addresses fragmentation within a department, larger cross-government initiatives aim to solve the same challenge on a national scale.

Central initiatives and use-case thinking

Large-scale projects such as the National Data Library aim to give the government a more connected, accessible data environment. Underhill sees clear potential in such initiatives but stresses that the value comes from clarity of purpose.

“Conceptually, it’s a brilliant idea. We just need to be clear about what the end goal is. That’s where I always start in the conversations I have with government: they may come to Cloudera and say, ‘We want to do AI,’ and our question back is, ‘Why? What’s the use case?’”

He points to examples like the Tell Us Once service, where citizens report a death a single time and the information is shared across departments, as a model of simple, high-impact design. “It’s not complicated to do. It’s not AI-backed. It’s very simple in how it’s been architected, but it has a real impact on people.”

For Underhill, starting with well-defined use cases ensures that investment delivers tangible benefits. “If we’re going to invest public money to do something, there should be a value to doing it otherwise, either remain with the status quo or don’t do it.”

Towards a single citizen data profile

The idea of a single citizen data profile, Underhill explains, is well established in the private sector as the “single customer view” or “golden record.” The principle is simple: hold a joined-up, accurate record of all the interactions an individual has with an organisation.

“In government, a lot of the time you repeat the same information. That’s not efficient from a customer point of view, and it means data is being replicated and duplicated,” he says. Without data-sharing agreements in place, information is rarely joined up between departments, leaving valuable insights untapped.

“If government were a bank, that would not be the case. Everything would be joined up so there would be a single record of me and all the interactions I have with government.” Such a profile, he argues, could be the foundation for more personalised services, with AI harnessing the information to anticipate needs and tailor delivery.

But the sensitivity of the data makes security and trust paramount. “One of the blockers we see is government departments being quite rightly concerned about exposing that information to AI tools. Our view is you bring the tools to the data, not the other way around. Having that security and that governance around it breeds trust.”

Targeting AI investment where it counts

Modernising legacy systems while delivering visible improvements is never easy especially when budgets are tight. Underhill is clear that government cannot compete with the scale of private sector investment in AI, but believes targeted action can still deliver meaningful results.

“I don’t envy them,” he says. “The investments being made in things like AI are staggering, and the government doesn't have that available headroom to compete at that level. But for me, it is worthy of an investment, where that investment needs to be targeted.”

He advises looking for areas where modest investment can both improve services and generate future savings. “Rather than boiling the ocean and potentially ending up with nothing, you can start small, and the learnings that government makes from the project can be built on.”

When asked what could prove the value of a unified citizen data approach, Underhill points to two opportunities: fraud detection and citizen engagement.

Fraud, he argues, is an ideal candidate for AI-driven improvement. “Fraud prevention using AI is endemic in private sector as it has a direct impact on profit and loss - The same principle within Government could protect public funds to be reinvested where it is needed” He stresses this is not limited to benefits or tax fraud, but could include a wide range of activities that, if addressed through better data sharing and AI, could protect public funds.

The other is citizen engagement, where tools such as AI-powered chatbots could free up capacity for complex cases. “When I deal with a bank now, I often wonder; is this a human, or is this a chatbot? Because they’ve got it spot-on. You could potentially filter 95% of all the cases out and free up people to focus on the 5% that actually warrant a human interaction.”

This is not, he cautions, a one-size-fits-all solution. “Some people may be very happy not speaking to civil servants when they’re trying to do something; some may want that. But we free up people to really focus the time and engage with those who need more help.”

Foundations for the future

Underhill’s message is clear: AI will only ever be as good as the data it relies on. Without strong foundations, secure, governed, interoperable, and trusted, technology investments risk under-delivering on their promise.

“Bad data in, bad AI out,” he says simply. For him, that’s the principle that should underpin every AI project in government. By starting with clear use cases, targeting investment where it can have the biggest impact, and protecting sensitive information through secure, modern architectures, departments can begin to break down silos and build towards more connected, citizen-centred services.

The prize, he suggests, is not just efficiency, but public trust. And in a world where AI’s role in public life is only set to grow, trust may be the most valuable foundation of all.