Anthony Finkelstein argues that the UK needs stronger tech foundations to reap the benefits of artificial intelligence
This article offers a deliberately sober form of techno-optimism. It argues that, whilst artificial intelligence (AI) holds real promise for public services, the decisive factor will be whether government is prepared to confront its accumulated, and accumulating, ‘enterprise tech debt’ and rebuild the digital foundations on which meaningful AI deployment depends. It sets out a practical path for managing tech debt and realising the strategic benefits of AI.
I believe firmly in the potential of AI to provide significant improvements in the delivery of public services and in the development of policy. Whilst I am not blind to its limitations – limitations that are for the most part inherent to the technology – I judge them to be substantially over-played. There are important near-term opportunities, some of which have been usefully highlighted in previous Heywood Quarterly articles (most notably those by Laura Gilbert and Michael Padfield).
Our ability to gain the greatest advantage from AI, however, is critically inhibited by the state of digital systems across government and the public sector. We are weighed down with legacy technology and, more importantly, by the legacy of our approaches to building and sustaining these systems.
Securing value
The key benefits from AI can be secured: first, by leveraging the extraordinary data resources we possess as a basis for learning; second, by using AI to orchestrate and supplement adaptive and goal-based workflows (technically known as ‘Agentic AI’); and third, through dynamic AI-mediated interaction with services. Each of these require significant technical and organisational integration but deliver, by far, the greatest payoff.
None of the benefits, though, can be achieved without plotting a path to confront and address our collective ‘enterprise tech debt’. Enterprise tech debt is best defined as the accumulated liability created through sustained underinvestment in software and systems. The debt manifests itself in, amongst other things, outdated systems and frameworks, accumulated piecemeal changes, inadequate testing, quick hacks (or shortcuts) and hasty compromises.
Confronting systems realities
The starting point for digital change is therefore our current systems landscape, as it presents us with a set of structural constraints that materially limit what can be achieved.
Our data is weakly managed and of highly variable quality, locked in silos. Our workflows are hard-coded and often stitched together across systems thanks to temporary or imperfect solutions. Our systems are too often closed, without the technical interfaces required to access their functionality. Worst of all, we frequently do not possess the information necessary to understand what we already have. Systems were built by contractors now long gone, and under commercial relationships that have since ceased to operate.
This is, of course, not a universal picture. There are islands of good practice where we have some opportunities but, as a generality, we are locked out of the advantages we should enjoy.
In what follows, I briefly examine how this came about, what we can do – recognising that simply ripping things up and starting again is not an option – and finally, how we can accelerate the path to more substantial AI benefits.
How this came about
Failures in budgeting, repeated bouts of costly organisational transformation, and shifting priorities can all lead to periods of underinvestment. Organisations like government departments that operate with annual budgeting cycles and have limited capacity for strategic technology planning are, in my experience, particularly prone to this.
These periods of underinvestment do not need to be prolonged to seed serious problems. Missing even a single technology cycle can be enough. Once that happens, the accumulation of tech debt becomes rapidly harder to stop. As operating requirements continue to evolve, the response is to make immediate, tactical fixes simply to sustain operations and keep the show on the road. New requirements, additional data and ever more complex systems integrations are layered onto an already fragile foundation. Each change increases overall system complexity and the cost of change rises accordingly.
This creates a vicious cycle. The growing cost of maintenance and adaptation consumes any residual budget that might otherwise have been used to pay down tech debt. Investment is progressively skewed away from renewal and innovation and towards survival. The problem is therefore not merely running to stand still but running in order to fall further behind.
In the best-case scenario, sufficient system knowledge and capability will have been retained to make changes. In too many cases, though, failures to retain system knowledge emerge cumulatively through staff turnover, outsourcing, organisational churn and crisis-driven change. Leaders leave and take design intent with them; delivery is pushed to suppliers that lack incentive to preserve architectural understanding, let alone render it accessible; and repeated restructurings break continuity of ownership. As systems become fragile, the effort shifts from learning to simply keeping things afloat, fixes are made under pressure, and documentation decays or is not produced at all. Over time, skills atrophy, technologies fall out of favour and the capacity to reason confidently about one’s own systems is lost. At that point tech debt ceases to be merely technical – rather, it morphs into a wholesale loss of operational capacity to act.
| Case Study 1 A UK Government department with a nationally-focused mission relies on a complex set of legacy digital systems inherited from predecessor organisations. Over time, these systems have required continual maintenance and modification to reflect changing operational demands. The cumulative effect has been increasing complexity, fragility and cost. The department does not possess a complete or reliable picture of its current digital architecture. As a result, leaders cannot readily assess the impact of proposed changes. Digital delivery is skewed towards short-term stabilisation. Locally developed capabilities cannot easily be absorbed into the core platform. Investment decisions are made without a clear view of the total cost of ownership, or the long-term architectural consequences. The capacity to deploy more advanced digital capabilities, including AI-enabled approaches, is limited. |
What we can do
Enterprise tech debt is a consequence of management choices, even though it manifests as a technical problem. It should, however, be treated as a strategic liability and governed alongside finance, risk and operational performance. That means making it visible and governable. At a minimum, it requires a shared map of both the ‘system-as-is’ and as it is intended to become – the ‘system-to-be’. This provides a common basis for decision making.
A ring-fenced allocation of budget and time for simplification, renewal, and ‘tech debt’ reduction is essential, particularly in fiscally-constrained environments. Short budgeting cycles, unrealistic programme horizons and success measures that focus solely on delivery will bias decisions towards what may appear to be expedience. The emphasis should be on reducing complexity and decreasing the cost of change, even when these deliver no immediate functional benefit.
Ownership must be restored and sustained. That requires clear accountability for systems over their full life, continuity of responsibility in the face of change and deliberate steps to develop and retain architectural capability. This is ultimately a stewardship challenge.
Technical interventions matter a great deal – reducing unnecessary complexity by decommissioning, removing duplication, standardising platforms and simplifying integrations. A key part of this is to reduce systematically the cost of change, principally through the design of interfaces so as to deliver incremental improvements, plug in new systems and build services quickly and safely. The process of simplification can be used as a basis for deliberately building system knowledge. The new architecture should be documented, kept current and treated as operationally critical. Knowledge should be embedded through shared ownership and design review processes. Tech debt reduction must be made part of delivery, not deferred to the next speculative stage of an extended digital transformation programme.
There is some good news. AI makes much of this technical work materially easier.
Case Study 2 The UK government department in the previous example began by establishing a clear view of its digital platform. The existing architecture is being documented, including systems, interfaces and data flows, with targeted reverse engineering where knowledge had been lost. This information is treated as operationally critical. Decision-making has been simplified. Roles and accountabilities for digital governance are mapped, including a strengthening of independent challenge. Investment decisions are now aligned to a coherent digital strategy and an explicit target architecture, rather than to disconnected projects or funding lines. A ring-fenced allocation of capacity has been created to support simplification, stabilisation and the transition to a supported operation. These steps are creating the conditions for sustainable change by lowering the cost and risk of future decisions. |
Path to AI benefits
Once tech debt is under control, the opportunity to secure the larger benefits of AI opens up. If we are serious about this, it requires a deliberate shift.
First, the basics. Scalable computing generally delivered through cloud or hybrid services is a prerequisite. Security and identity must be built in from the start. AI should be developed and deployed as an ‘enterprise capability’. Though experiments and early pathfinders have their place in building understanding and capacity, AI stacks (layered assemblages of technologies) should not proliferate. A unified AI platform, providing access to models, orchestration, lifecycle management and monitoring will support consistency and control.
Data is a strategic asset, to be managed for long-term value, rather than simply for immediate operational needs. Leveraging data resources already held across the public sector means more than aggregation. It requires curation, semantic coherence and, of course, trust. Data must be discoverable and governed in ways that permit reuse across boundaries.
AI creates value when it is embedded directly into workflows. Standalone tools rarely yield the larger benefits we are seeking. AI that operates inside case management systems, operational platforms or professional tools, supporting decisions at the point they are made, is how transformation occurs.
If AI is to act as an intelligent intermediary between users and services, then those services must be exposed cleanly and consistently. Legacy systems tightly bound to interfaces, hard-coded policy logic and informal exception handling are structural blockers. Only modular, callable and observable services can be orchestrated dynamically across systems in real-time.
Agentic systems function most effectively where objectives are made explicit. That demands the surfacing of what is often implicit: decision criteria, tolerances, success conditions and escalation points. In practice, this means redesigning processes so that AI systems pursue defined goals within constraints, while handing off to human judgement where authority or legitimacy requires it. Clear escalation paths, explainable outputs, override mechanisms and auditable decisions must therefore be part of the technical design. They cannot be a policy or governance bolt-on. The same is true of model approval, monitoring, compliance checks and the application of ethical constraints.
To date much of the focus in government has been on AI skills and familiarity; these are undoubtedly important but insufficient on their own. Architectural and systems thinking are foundational. We also require greater expertise in process redesign, service and interface design and an understanding of decision-making. Where these skills are absent, progress will be limited.
Case Study 3 Prior to the initiation of the programme described above, our exemplar government department launched a pilot to use AI for an important operational task. The model itself performed well in testing but the failure came elsewhere. Case data was spread across multiple systems with no authoritative source of truth. Key decision steps were embedded in legacy workflows, some hard-coded in software, others executed through informal temporary workarounds. There was no reliable way to expose the end-to-end process as a callable service. To deploy the AI tool in live operations would have required changes across several interdependent systems. No one could state with confidence what those changes would break, how long they would take or who ultimately owned the risk. The pilot was therefore left running in parallel, producing insights that could be admired but not acted upon. The problem was not trust in AI outputs. It was the inability of the digital estate to absorb the required change. AI revealed the limits of the organisation’s systems far more clearly than it delivered immediate benefits. |
Strategic payoff
If we get all of this right, the prize is substantial. We move from brittle systems that constrain policy and delivery to adaptive public services that learn, improve and respond intelligently to changing needs. Decisions become better informed, faster and more consistent, without displacing judgement or accountability. Services become easier to access, easier to adapt and cheaper to change.
AI gives us an opportunity to rebuild government digital capability on more sustainable foundations. Doing this requires us to make choices that privilege long-term capacity over short-term expedience. That is a familiar challenge necessitating clarity of purpose and a stewardship mindset, something that would certainly have been recognised by the late Jeremy Heywood, after whom this publication is named.
