Matthew Jeffery: To take the path of modernisation not austerity is the real test of nerve (part 2)

Matthew Jeffery is one of Britain’s most experienced global talent and recruitment leaders, with more than 25 years advising boards and C-suite executives on workforce strategy, skills, and productivity.

This is part 2 of Matthew’s short series on AI and the AI generation. You can read part 1 here.

The education system represents a major but under-recognised opportunity. AI can significantly reduce teacher workload through automated planning, marking, and administrative support, while enabling personalised learning that reduces long-term remediation costs. Department for Education studies show that workload is one of the primary drivers of teacher attrition, and reducing this pressure improves retention while lowering recruitment and training costs.¹³ Early identification of special educational needs enables earlier intervention, avoiding significantly higher costs later in the system.

Across regulation and compliance, AI enables a shift from periodic inspection to continuous monitoring. Systems can automatically assess compliance in real time across financial services, environmental standards, and health and safety regimes. This reduces enforcement costs while improving outcomes and also lowers the compliance burden on business, supporting economic growth while reducing the cost of the regulatory state.

In defence and national security, AI-driven logistics, predictive maintenance, and cyber capability deliver meaningful efficiency gains without reducing operational effectiveness. Even modest improvements across defence spending translate into billions in savings while enhancing capability, making this both a fiscal and strategic advantage.

A State That Learns

What emerges is not a collection of tools, but a fundamentally different model of the state. A state that learns. A system where every interaction improves the next, where intelligence is shared across departments, and where errors are not repeated because the system continuously adapts.

In practical terms, this means no repeated identity verification, no duplication of effort across departments, and no reliance on static processes that fail to evolve. Services become proactive rather than reactive, decisions are made in real time, and the system becomes more effective over time. This is not incremental reform. It is structural transformation.

From Tools to Agents: The Next Phase of the State

Much of the current conversation around artificial intelligence in government remains focused on tools. Systems that assist, accelerate, or augment existing work. That is where much of the early value has been realised, and it is reflected across the examples outlined throughout this piece.

However, the frontier is already moving beyond assistance and into execution.

The next phase is the emergence of agentic systems. Systems that do not simply support tasks, but plan, coordinate, and complete them. Systems that can manage multi-step processes, operate across organisational boundaries, and act with defined levels of autonomy within clear guardrails.

In practical terms, this changes how government is experienced.

A benefits claim is no longer a sequence of forms, checks, and delays across multiple departments. It becomes a single interaction, handled end-to-end by a system that gathers information, verifies eligibility, flags risk, and delivers an outcome. Appeals are reduced because decisions are more accurate at the point of entry. Administrative overhead falls because the process itself has been removed.

A single agentic system handling a benefits claim could already integrate data from DWP, HMRC, and local authorities, verify identity once, assess eligibility, flag anomalies, and deliver a decision within minutes rather than weeks. The same principle extends across systems. In healthcare, agentic coordination between NHS services and social care could identify high-risk individuals, deploy preventative interventions, and reduce admissions before they occur, not as isolated predictions, but as coordinated system action.

This is where the economic impact shifts from additive to multiplicative. Each layer of coordination removes entire categories of demand, rather than simply processing them more efficiently.

In healthcare, the same principle applies. Systems do not simply identify risk, they coordinate response. They align social care, primary care, and hospital capacity in real time, reducing admissions not through isolated prediction, but through coordinated intervention.

In justice, disputes can be triaged, mediated, and resolved before they ever reach a courtroom. In transport, systems optimise not just traffic flow, but the deployment of assets across entire networks. In procurement, systems continuously manage supplier performance and pricing rather than reviewing them periodically.

This is where the concept of a learning state becomes materially different. It is not simply a state that gathers insight. It is a state that acts on it continuously.

Each interaction improves the next. Each decision refines the system. Errors are not repeated because the system adapts in real time.

The implication is significant. The largest gains do not come from doing the same work faster. They come from removing the need for that work altogether.

This is not a future scenario. Elements of this are already emerging. The question is whether they are deployed systematically, or remain isolated within pilots that never scale.

The Missed Layer of Savings

The picture outlined so far captures the most visible and immediate opportunities, but it may still understate the full scale of what is possible. Beneath the headline areas sits a second layer of savings that is less visible, more systemic, and potentially just as significant.

Social care and adult social care integration represent one of the clearest gaps. Long fragmented from the NHS, social care operates as a parallel system rather than an integrated one. Artificial intelligence can enable predictive risk modelling, optimise workforce rostering, and support remote monitoring that allows individuals to remain independent for longer. The impact is not confined to local authority budgets. It reduces hospital admissions, delays or avoids residential care, and lowers demand across the wider health system. These are savings that compound across services rather than appearing in a single line of expenditure.

Tax administration also offers further opportunity beyond fraud recovery. HMRC already makes extensive use of data analytics, but artificial intelligence enables a step change in automated compliance, dispute resolution, and customer service. Routine queries can be handled instantly, audits can be prioritised using real-time risk signals, and processing times can be significantly reduced. This lowers administrative cost while improving the experience for individuals and businesses alike.

The same applies across benefits and welfare administration. While fraud and error have been addressed, a larger opportunity lies in end-to-end automation of claims processing. AI-driven assessment, reduced appeals through better initial decision-making, and more personalised support for claimants can materially reduce overhead while improving outcomes. This is not simply efficiency. It is a redesign of how entitlement systems function.

Cybersecurity and IT estate rationalisation present another under-recognised opportunity. Government systems remain fragmented, costly to maintain, and increasingly exposed to risk. Artificial intelligence enables continuous threat detection, automated patching, and the consolidation of legacy systems. This reduces both cost and vulnerability, addressing one of the fastest-growing areas of public expenditure.

Pensions administration represents a further area of scale. State and public sector pension systems involve high volumes of processing, verification, and support. AI can reduce error rates, prevent fraud, and streamline administration, delivering savings that accumulate over time across large populations.

More fundamentally, cross-cutting data platforms enable what is often described as a “once-only” principle. Citizens and businesses provide information a single time, and it is reused securely across government. This removes duplication, eliminates repeated identity checks, and reduces friction across services. It is also the foundation of a truly connected system, where intelligence is shared rather than recreated.

There is also a growing opportunity in environmental compliance and net-zero delivery. Artificial intelligence can optimise regulation, grant allocation, and public building retrofits, reducing long-term costs associated with climate obligations while improving outcomes.

Taken together, these areas do not sit outside the core opportunity. They deepen it. They reinforce the central argument that the prize is not confined to isolated interventions, but embedded across the entire operating model of the state.

If pursued with the same discipline and intent, they could plausibly add a further £5–15 billion over time, pushing the upper bound of savings materially higher. As with the core estimates, this depends entirely on whether productivity gains are captured as cashable savings rather than absorbed into system expansion.

The Scale of the Prize

Taken together, the opportunity becomes clearer when separated into two layers: the base case available from today’s tools and augmentation models, and the more ambitious upside available from a fully realised learning state built around agentic systems, shared data, and prevention at source.

Category	Base Case: Tools and Augmentation	Agentic / Learning State Upside
Fraud and error (welfare and tax)	£8 to £12 billion	£10 to £15 billion through real-time prevention and automated compliance
Procurement optimisation	£9 to £15 billion	£12 to £20 billion through continuous AI-led contract and pricing optimisation
NHS productivity and administration	£3 to £6 billion	£10 to £18 billion through prevention, coordinated care, and demand reduction
Energy and estate optimisation	£1 to £3 billion	£3 to £6 billion through fully automated infrastructure management
Local government transformation	£4 to £8 billion	£8 to £15 billion through integrated service delivery and automation
Courts and justice system	£1 to £4 billion	£5 to £10 billion through pre-court resolution and automated case handling
Workforce productivity	£5 to £12 billion	£10 to £20 billion through large-scale removal of administrative work
Transport and autonomous systems	£3 to £7 billion	£8 to £15 billion through accident reduction and system-wide optimisation
Regulation and compliance	£3 to £8 billion	£6 to £12 billion through continuous monitoring and automated enforcement
Education system	£2 to £5 billion	£5 to £10 billion through personalised learning and reduced remediation
Defence and national security	£2 to £4 billion	£5 to £8 billion through predictive systems and autonomous logistics
Cross-system effects	—	£10 to £20 billion through once-only data, end-to-end automation, and elimination of duplication

Savings potential varies significantly depending on ambition. The base case relies on today’s assistive tools and partial adoption. The agentic/learning state case assumes systematic redesign, end-to-end automation, and prevention at source.  The leap from base case to agentic upside comes primarily from shifting from efficiency (doing the same work faster) to prevention and elimination (removing the need for the work altogether)

These figures assume deliberate policy decisions to capture gains as cashable savings rather than absorbing them into expanded services, consistent with observed 12 to 24 month payback periods in pilots.

The base case produces a realistic near-term saving of £30 to £50 billion, with an upside of £35 to £60 billion over time as systems mature and adoption scales.

The agentic and learning state case reflects a fundamentally different operating model. By combining end-to-end automation, prevention, and the removal of entire process layers, total savings could realistically reach £70 to £100 billion annually by 2030.

This is not a continuation of current trends. It is the result of a different model of the state.

To put this in context, £30 to £50 billion is equivalent to cutting around 5p to 9p off the basic rate of income tax without increasing borrowing. At the upper end, £70 to £100 billion would represent a transformation in both the cost and capability of the state without raising taxes or cutting core services.

This is not marginal. This is structural.

A Delivery Problem, Not a Technology Problem

Britain does not lack ideas, and it does not lack pilots. It lacks delivery at scale. The technology already exists, the use cases are proven, and the savings are visible. What is missing is the mechanism to translate productivity into cashable savings and to do so consistently across the system.

A serious government would treat this as a fiscal strategy, not a digital initiative. It would establish an AI Delivery Authority, set binding targets, and ensure independent verification of savings. It would align investment with outcomes and ensure that gains reduce costs rather than expand them.

Making Delivery Unavoidable

If this is to function as a fiscal strategy rather than a collection of disconnected initiatives, the delivery mechanism must be materially stronger than anything currently in place.

An AI Delivery Authority would need to operate with clear, enforceable powers. Not simply to coordinate activity, but to ensure that gains are captured and translated into measurable fiscal outcomes.

This requires binding targets for productivity and cost reduction across departments, aligned to spending settlements and reviewed annually. Where gains are identified but not realised, budgets should be adjusted accordingly to prevent savings being absorbed into new spending.

It also requires independent verification. Claimed savings must be audited, published transparently, and subject to scrutiny. Without this, the risk is that productivity gains are overstated while costs remain unchanged.

Data infrastructure must be treated as national infrastructure. A secure, shared data layer, aligned to a “once-only” principle, should not be optional. Departments should be required to participate, with clear legal and technical frameworks enabling data sharing while maintaining security and privacy.

The objective is not centralisation for its own sake. It is coherence. A system that operates as a system, rather than a collection of disconnected parts.

Without this level of discipline, the pattern of the past will repeat. Pilots will succeed. Case studies will accumulate. And the overall cost of the state will remain unchanged.

The 2030 State

By 2030, a state that successfully applies these tools does not look like the one we have today. It is faster, more responsive, and materially more efficient, not because it has been cut back, but because it has been redesigned. Administrative overhead is lower because time is no longer consumed by processes that add no value. Doctors spend more time with patients. Teachers spend more time teaching. Public servants operate at a higher level of impact.

The system feels different because it operates differently. Decisions are made faster, risks are identified earlier, and demand is managed before it escalates into crisis. Services become more personalised, more proactive, and more effective.

This is not a smaller state. It is a more capable one.

The Trade-Offs That Cannot Be Avoided

Any transformation of this scale carries trade-offs, and avoiding them weakens the argument rather than strengthening it.

In the short term, there will be disruption. Roles will change, some functions will reduce, and parts of the workforce will need to be retrained or redeployed. This is not a consequence of failure, but of success. A system that removes unnecessary work will inevitably change the structure of employment within it.

This requires a deliberate transition strategy. Investment in skills, clear pathways for redeployment, and alignment between productivity gains and workforce planning. However, it also requires honesty. Not every role will remain unchanged, and not every function will persist in its current form.

There are also risks. Model error, bias, cybersecurity exposure, and dependency on a small number of providers are all real considerations. These require robust governance, including rigorous testing, clear accountability, and human oversight in high-stakes decisions.

There are institutional challenges. Resistance within systems that are designed to preserve stability, political caution in the face of perceived risk, and public concern about automation in decision-making.

These are not reasons to avoid change. They are constraints that must be managed.

The greater risk lies in maintaining a system that is already under strain, already inefficient, and increasingly unable to meet rising demand. The status quo is not neutral. It is a path to higher cost and lower performance over time.

A Test of Nerve

This is ultimately a choice about what kind of state Britain intends to build by 2030. One that continues to layer technology onto existing systems, or one that uses it to fundamentally redesign how government operates.

Artificial intelligence offers one of the few credible paths available to reshape the economics of the state. The technology works. The savings are real. The opportunity is immediate.

What remains uncertain is whether the state is willing to act.

Without productivity gains of this scale, the fiscal arithmetic does not hold. The alternative is higher taxes, lower growth, or continued decline in public service performance. Those are the real choices.

Artificial intelligence can transform government. The only question left is whether government has the courage to act.