There's a version of the AI conversation in public services that I find increasingly difficult to sit through.
It usually starts with a compelling demonstration; a large language model summarising casework, a predictive tool flagging risk earlier, an automation that handles a process in seconds that used to take hours. The demonstration works. The use case is real. The potential is genuine. Then someone asks the question that tends to bring the energy out of the room: what would it take to deploy this at scale, in a live service, on the systems we actually have?
The honest answer in most cases is considerably more than the pilot suggested.
This isn't a criticism of the technology. The tools are genuinely powerful. It's an observation about the gap between where AI performs well, in controlled environments with clean data and against well-defined problems, and where public services actually operate. That gap is largely defined by legacy. By data that was collected across decades in incompatible formats, sitting in siloed systems that were never designed to share it. By workflows and undocumented workarounds that were built around the constraints of old technology and have since become embedded in how organisations actually function. By integration challenges that don't show up in a pilot because the pilot was carefully scoped to avoid them.
The pattern I see most often is what I think can be called the pilot trap. An AI initiative is stood up with enough resourcing and careful scoping to demonstrate genuine value. The demonstration succeeds. The business case is made, the work of scaling it into a live service environment begins and encounters the estate that the pilot was designed not to engage with. The result, more often than it should be, is an initiative that works in theory and stalls in practice.
This isn't inevitable but avoiding it requires treating the data and systems question as part of the AI strategy, not as a separate technical workstream that can be addressed later. But later, in this context, tends not to arrive.
The data quality problem is the most underappreciated challenge in public sector AI adoption. Legacy systems are frequently the custodians of decades of institutional knowledge, but that knowledge is stored in ways that reflect the technology and practices of the time it was captured. Fields that mean different things in different systems from different times. Records that were never reconciled when organisations merged or restructured. Data that exists but can't be trusted without significant work to validate and contextualise it. Teaching a model to work with data in this condition produces outputs in the same condition, if not worse. The quality of what comes out is bound by the quality of what goes in.
There's a governance dimension that sits alongside this and deserves equal attention. AI in public services isn't a back-office efficiency question, it involves decisions that affect people's lives, often people who are already in difficult circumstances. Getting the human oversight right, building the audit trails, ensuring that the people responsible for outcomes understand what the system is doing and why – these aren't constraints on AI adoption, they're the conditions under which it can be trusted.
What this adds up to is a framing that I think is more useful than the one that dominates most AI conversations, specifically: productivity and innovation in public services are systems problems, not technology problems. The three blockers I see most consistently, weak digital foundations, innovation that can't get beyond the pilot stage and delivery that's fragmented across teams and programmes, aren't solved by a better model or a more ambitious strategy. They're solved by the patient, precise work of understanding what you have, reducing the risk it carries and building the integration layer that allows new capabilities to operate in the real world.
None of this is a reason for pessimism. The organisations that have done that foundational work, that have taken an honest look at the estate, sequenced their modernisation intelligently and invested in the data infrastructure that AI actually needs, are the ones where the pilots are starting to become programmes. The potential that gets demonstrated in those early use cases is real. The question is whether the foundations exist to catch it when it scales.
Ready to unlock productivity in your organisation?
Discover how leading organisations can modernise complex legacy estates without disrupting critical services. Join myself and Fiz Yazdi as we share practical lessons on moving beyond pilots, managing risk in-flight, and building the data, governance and integration capabilities needed to unlock productivity and scale AI with confidence.
Watch the webinar here: