Is AI really the answer to government efficiency targets?

As part of its drive to improve government productivity and value for money, the UK Government has published tough efficiency expectations, including a target of at least 5% efficiencies and savings by 2028–29, alongside a reduction in departmental administration budgets of at least 16% in real terms by 2029–30. Departmental efficiency plans published with the Spending Review identify almost £14bn of annual efficiency gains by 2028–29.

The justification for setting such ambitious targets is that AI will be used to achieve them. However, this is a big leap. AI is a powerful enabler, yes, but it is not the answer. Whilst the promise of AI isn't in doubt, government departments don’t (yet) have a clear path to realised value.

This distinction matters because the pressure on departments is real. They are being asked to cut costs, drive productivity, modernise ageing digital estates, maintain or improve service quality and do it all in ways the public can trust.

In this environment, AI is an obvious temptation. It promises automation, speed and scale. To unlock real value, however, departments must approach AI implementation with clear baselines, defined use cases and measurable outcomes from the start.

Too often, the conversation starts with: “How do we adopt AI?” Instead, we should be asking: “Where does AI genuinely improve productivity, efficiency and outcomes, and how can we unlock that value more quickly?” In other words, the hype is about tools, whereas the route to success is about selecting the right use cases.

Selecting use cases where AI genuinely adds value

The most mature organisations understand that AI is more than a cost lever; they see it as a way to improve efficiency, raise productivity and transform the customer experience. In government, that means increasing capacity to meet demand, improving workforce productivity, rethinking how services are designed and delivered, and creating innovative ways to deliver better outcomes for citizens.

These are the goals that should shape how organisations approach AI. The right framing is not 'AI transformation'. It is productivity transformation, with AI tested pragmatically, applied selectively and scaled where it can truly reduce cost-to-serve, release capacity and improve outcomes. The most powerful use cases are those rooted in real service pain.

To identify credible opportunities, departments need a clear view of the end-to-end user journey, the operational flow and the workforce. Where does demand enter the system? Where does work stall? Where is effort duplicated? Where do staff spend time on low-value activity? Where is quality control done too late, or not consistently enough? Where are we experiencing failure demand, where users have to make repeat contact because their issue was not resolved properly the first time?

Answering these questions allows leaders to identify where AI can help, sequence use cases sensibly and judge success against real operational and service metrics, rather than abstract ambition – and choose delivery approaches accordingly.

Discipline matters. If a team cannot articulate the current baseline, the intended improvement and the safeguards needed before building anything, the use case is probably not ready. It’s vital to consider operational and service-led metrics, including first contact resolution, complaint volumes, self-service resolution, backlog volume, and adherence to service-level agreements (SLAs). These should be considered alongside measures of quality, accuracy, consistency and user trust. Faster is not better if the wrong things are being done more quickly, or if failure is simply pushed further downstream.

Once value is defined in these terms, it becomes easier to see which use cases have genuine potential.

The real opportunities for AI in the public sector

The highest-value opportunities are often unglamorous. They tend to lie in augmenting the capability, productivity and experience of employees working across the public sector. Automation of customer-facing interaction often proves less valuable (at least for now).

These findings are supported by Baringa's recent AI research, which analysed survey responses from 1,000 consumers and more than 37,000 online reviews of AI-enabled customer service. Half of consumers believe AI should only be used to enhance and augment employees' work, and 82% say it's essential or important to speak to a human to deal with a customer service issue (full report here). For government, these findings reinforce that the productivity prize sits in augmenting staff - reducing administrative burden, improving case handling and removing avoidable rework, rather than replacing the human interactions citizens value most.

In practice, this means the most valuable opportunities for introducing AI include:

• Improving intake and triage by categorising requests, extracting key information, identifying missing evidence and routing work to the right team faster
• Automating document processing, case summarisation, and information extraction to reduce administrative effort
• Reducing complaints and failure demand through root-cause detection and response drafting, so organisations can fix recurring problems faster
• Strengthening quality management by increasing assurance coverage and consistency while reducing cost
• Connecting front‑end contact management with back‑end caseworking via a shared AI layer that captures intent and enriches context then routes, pre-populates and updates cases end-to-end. This reduces handoffs, duplication, and rework.

By contrast, the most overhyped use cases tend to optimise fragments while leaving the underlying operating model unchanged. Standalone chatbots and isolated call summarisation pilots generate noise rather than savings, because they don’t change the economics of service delivery. Similarly, using AI to make sensitive eligibility decisions is often presented as an efficiency breakthrough when, in live operational environments, it often increases friction, risk and assurance burden.

A useful rule of thumb is this: if a capability does not improve the end-to-end journey or operating model, it is probably more performative than productive. If it reduces failure demand, backlog or manual casework, it is far more likely to generate real value.

Once it is clear which use cases will drive meaningful productivity gains, organisations must then decide how best to deliver them within their broader operating model. In many cases, value is unlocked faster through proven software-as-a-service (SaaS) solutions rather than bespoke builds. Leading platforms, particularly in areas such as contact management, already optimise core functions such as routing, case handling and assisted workflows. Choosing whether to buy, configure or build is therefore a critical operating model decision. Getting this wrong often leads to unnecessary complexity and delays in realising value.

What it takes to scale AI responsibly

Even when organisations choose the right use cases, AI programmes often falter when teams attempt to scale them.

The pattern is familiar. A department pilots an AI capability - say, automated case summarisation in a casework team. It shows encouraging results. But when the team tries to roll it out across the operation, they hit barriers the pilot never surfaced: inconsistent data quality across regions, no agreed process for human review of AI outputs, and no clarity on who owns accuracy when the tool gets it wrong. The pilot succeeded because a small, motivated team compensated for all of this. At scale, those workarounds collapse.

Scaling AI is an operating model challenge, not purely a technology one. It plays out in how work is designed and governed day to day: workflows, decision-making, roles, incentives and assurance. If organisations try to scale use cases without addressing these conditions, technical debt builds through point solutions that are hard to maintain, operational efficiency declines as workarounds multiply, and trust deteriorates as front-line teams disengage.

Most importantly, time is lost. Efficiency targets do not wait while organisations experiment.

To overcome these challenges, departments must think beyond pilots and tools and consider the operating model needed to turn promising use cases into lasting service improvement. The goal is not to design the perfect operating model upfront, but to establish one that can support experimentation and evolve as value becomes clear.

Ownership must be clear from day one - not only for the performance of the tool, but for service outcomes, adoption, oversight and continuous improvement. Roles and responsibilities often need to change, and in some cases entirely new roles will emerge. Governance, risk and assurance must be aligned to the reality of AI-enabled services, not bolted on afterwards. Workflows may need redesigning, organisations will need to plan how AI fits into the wider technology estate, and training and change management needs to be active rather than assumed. And AI is not a one-off investment; funding must reflect the need for continuous improvement, with value realised over time in business-as-usual.

Start with productivity, then apply AI with precision

AI can absolutely help government meet ambitious efficiency targets, but efficiency is only part of the prize. The wider goal is productivity: delivering more or better public outcomes with the same or fewer resources. In practice, that means improving outputs, boosting quality and making better use of people, systems, data and budget.

This calls for a broader lens than cost reduction alone. The question is no longer where AI can save money, but where it can help services handle more demand, improve the consistency and quality of decisions, and make better use of scarce capacity.

The most effective route to success is:

• Start with the service outcomes that matter and understand productivity constraints
• Use insight from the customer journey, the operation and the workforce to identify where value really sits
• Prioritise a small portfolio of use cases with clear baselines
• Decide pragmatically which capabilities to buy, configure or build
• Put ownership, governance and safeguards in place early
• Track benefits rigorously over time
• Plan and allocate longer-term funding.

That is the difference between AI creating hype and delivering tangible results.

Productive beats performative. The departments that grasp that distinction early, and build accordingly, are the ones that will deliver results.

If you’d like to learn more about how Baringa can help you drive value from AI, please get in touch with Hannah Bolton and James Ainsley.