Public Sector AI Faces a Choice Between Innovation and Cuts
By Professor Ashley Braganza
Public sector AI has a productivity problem before it has a technology problem. Greater productivity can come from creating more value with existing resources or consuming fewer resources to maintain existing outputs. Those routes can produce very different public services. AI makes both possible. The risk is that the second is easier to measure. A reduction in expenditure appears in accounts relatively quickly. The value of redesigned services, greater capacity and better outcomes can take longer to establish. That difference could determine the direction of public sector AI adoption.
My guest on The AI Adoption Podcast is Ash Thankey, Managing Director and General Manager for Amazon Web Services Public Sector across the UK, Germany and International Organisations. Ash's perspective is optimistic. He sees AI augmenting skilled work, supporting public sector productivity and lowering barriers to entrepreneurship. His examples also raise a harder question about the form that productivity improvement ultimately takes.
Public sector AI creates capacity before it creates value
Ash describes applications that can productivity benefits that are not confined to reducing the cost of processing any task. In particular, he suggests that human capacity can be redirected. That distinction is central to the productivity debate.
Removing repetitive work creates capacity, but capacity is not itself value. Leaders still have to decide where the released time, attention and resources go. They can use them to improve a service, increase its reach, respond more quickly or redesign the underlying process. They can also remove the capacity from the organisation.
Both may be described as productivity. Ash distinguishes direct pound savings from productivity savings. He argues that freeing civil servants who deal with citizens to concentrate on the tasks that matter can produce benefits with a longer-term effect.
The distinction matters because accounting systems make some benefits much easier to see than others. A lower cost can be identified. The economic and social value of a civil servant spending more time solving a citizen's problem is harder to express in the same way.
Leaders therefore need to decide what they mean by productivity before they deploy AI to pursue it.
The measurable AI benefit may not be the most valuable one
There is a legitimate argument for cost reduction. Public resources are finite and inefficient administrative activity consumes money that could be used elsewhere. Removing unnecessary expenditure is not inherently undesirable.
The problem arises when measurability starts determining strategy. Growth in a national economy requires new activity and new forms of value creation. Within public services, the parallel is similar. Innovation means finding different ways to deliver outcomes rather than simply making the existing organisation cheaper.
That is more difficult. Ash acknowledges that transformation has difficult stages. He describes the journey as containing the good, the bad and the ugly. Legacy applications may need to be moved. Technology has to change. Organisations also need sufficient commitment to make that transition.
Innovation consequently brings uncertainty that a straightforward savings programme may avoid. This creates an asymmetry for leaders. The cost case may arrive with numbers that can be placed into a budget. The innovation case can require an organisation to invest before the eventual form and scale of the benefit are completely known.
That makes conventional cost reduction tempting and there is a strong temptation to define the public sector AI agenda in terms of savings alone. If the principal test of AI adoption becomes the amount of expenditure removed, organisations may optimise for a result that is visible rather than for the result that creates the greatest public value.
The more demanding question is whether AI allows a public service to achieve something it could not reasonably achieve before. Ash's discussion of entrepreneurship makes the same point from another direction. He argues that falling barriers to entry allow people without extensive coding experience to create applications and businesses. He argues that the economic opportunity comes from new activity rather than simply making existing businesses cheaper. The public sector equivalent deserves equal attention.
Model choice will not remove the need for organisational choice
Ash also describes a technical change that could make this landscape more flexible. His argument is that organisations are moving towards a multimodel environment. Amazon Bedrock is designed to give customers access to different models and allow the model supporting an application to be changed without a large engineering change.
The most capable model is not necessarily required for every task. Ash uses the analogy of not needing the Ferrari of models for a basic workflow. A less expensive model may be sufficient. Model choice can therefore become an economic as well as a technical decision.
The possibility becomes more interesting when agents enter the picture. Ash suggests that an agent could potentially determine which model to use, although he also says that organisations may prescribe the workflow and that customers generally have specific views about the models they want.
Whether AI technologies can reliably make these choices across complex public sector processes at the required level of accountability remains an open capability question.
Even if the technology can do so, model choice does not remove organisational choice. A public body still has to determine the outcomes it seeks, the acceptable cost, the risks it will tolerate and the decisions that remain subject to human judgement.
This is where Ash's observation about fragmented adoption becomes important. He argues that organisations without a centralised will to move can end up with different groups wanting different things and a fragmented approach. That is not primarily a model problem. It is an organisational one.
Leaders must decide where the AI dividend goes
Ash's advice to senior decision makers starts with the outcome. He describes AWS working backwards from the customer by asking what the organisation is trying to achieve, its end goal and the speed at which it wants to get there. Existing technology and legacy infrastructure then affect the route.
I think public sector leaders need to take that logic one stage further. The question is not merely which use case to select. Leaders need to determine where the benefit created by AI should go.
Suppose AI reduces the administrative burden associated with a public service. One option is to remove cost. Another is to redeploy the released capacity towards citizens. A third is to redesign the service around capabilities that were previously unavailable.
Those choices lead to different organisations. They also create different measures of success. A cost reduction strategy can reasonably measure expenditure removed. A capacity strategy needs to establish whether released human effort has actually been redirected. An innovation strategy requires measures connected to service outcomes and the additional value created.
Without that understanding, the word productivity conceals rather than clarifies the objective. Ash's point about universities reinforces the scale of the organisational challenge. He argues that higher education has a dual role, preparing students for an AI enabled workforce while also adopting AI as an employer. He expects people to return to higher education to update their skills as technology changes.
Skills, jobs, organisational structures and public services therefore move together. Treating AI productivity as a technology metric misses those connections.
AI productivity is ultimately a choice about public value
The public sector has an opportunity to use AI to remove work that consumes human attention without creating equivalent value. Ash provides examples that suggest this is already beginning.
The unresolved issue is what happens to the benefits. Cost reduction will sometimes be appropriate. But if savings become the default definition of productivity because they are easier to count, AI may make existing public services cheaper without creating the innovation its capabilities make possible.
The leadership challenge is therefore not simply to adopt AI successfully. It is to decide, before the benefits appear, which benefits the organisation is trying to create and for whom.
Listen and follow
LinkedIn Showcase Page:[Showcase Page URL to be supplied]
Podcast website:https://theaiadoptionpodcast.com
Comments