Banking AI Gains Without Hollowing Workforce Capabilities

Efficiency Gain, or Capability Loss? The tasks generative AI automates most easily are often the ones through which professionals build their judgement. In the public sector, where legitimacy rests on human reasoning, lasting gains depend on careful delegation.

A reflection by MPA Capstone students at the London School of Economics and Political Science, School of Public Policy.

Every few decades a technology arrives that changes not just what organisations produce, but how the people inside them think and work. Generative artificial intelligence (AI) is one of those technologies but not just another general purpose technology. What makes it unusual is its reach: unlike past tools, which landed on one function at a time, it touches almost every knowledge worker at once, and the limits of what it can do well remain genuinely uncertain. For public sector organisations, where legitimacy rests on the quality of human reasoning rather than the sheer volume of output, the stakes of getting adoption right are extremely high.

The productivity case is not in doubt. Used well, these tools draft faster, clean data faster, and write code faster, freeing scarce human capacity for the higher-value analytical and judgement-heavy work that public institutions exist to do. The harder question, and the one that determines whether those gains last, is what happens to the capabilities that the automated work was quietly building earlier. This piece sets out why workforce skills belong at the centre of any public sector AI adoption strategy, and what the wider evidence suggests an institution can do about it. Some of these ideas were explored with a small group of staff at the Bank of England as part of a broader reflective exercise on future skills, an engagement that helped sharpen the questions below, though the argument here rests on the wider literature.

The Productivity Paradox at the Heart of Adoption

There is a paradox buried in the efficiency story. The tasks AI handles most easily, the first draft, the preliminary code, the data formatting, are often the very tasks through which professionals have traditionally built their foundational expertise. When a junior analyst no longer writes a first pass from scratch, or stops wrestling with messy data by hand, something beyond a few saved hours may be changing. The development pathway itself is altered, with consequences for how an organisation grows its future talent.

The research literature gives this concern a name. When people delegate cognitively demanding work to a tool, their engagement with the underlying reasoning can fade through a process researchers call cognitive offloading. Over time this can accumulate into what some describe as cognitive debt: a widening gap between what someone appears able to produce and what they can actually do unaided. The risk is not the tool. It is the quiet substitution of practice by output, where the work still gets done but the capability behind it slowly thins.

Efficiency gains that depend on a capability the workforce is no longer maintaining are gains built on borrowed time. The most productive use of AI is therefore not the one that delegates the most, but the one that delegates deliberately, preserving the human capabilities that make the outputs trustworthy and the gains durable.

A Jagged Technology Frontier and an Uneven Workforce

Two features of the evidence make blanket adoption strategies risky. The first is that AI capability is uneven in ways that are hard to predict. Researchers describe a jagged technological frontier, where these systems excel at some tasks and fail at others of seemingly similar difficulty, so competence in one area cannot be assumed to carry across to the next. In practice this means experienced professionals often disagree about which tasks can safely be handed over and which must stay human-led. This is not due to confusion, but because the boundary is contested and shifts as the technology moves rapidly. The boundary between what to preserve and delegate should be a living diagnostic an organisation revisits, not a fixed list.

The second feature is that the impact falls unevenly across the workforce. Early-career staff have the most to gain from hands-on practice and, for that reason, the most to lose from skipping it. Evidence suggests AI delivers its largest immediate productivity boost precisely to less experienced workers, which is exactly why it can tempt them past the foundational work that builds judgement. Senior staff face a different problem: keeping the oversight instincts sharp when much of the routine work that once honed them has been automated.

There is also a subtler risk that self-assessment alone cannot catch. People do not always notice when a capability is fading, because confidence can remain intact while the underlying skill quietly declines. This divergence between how able someone feels and what they can actually do unaided functions as an early warning sign, one that surfaces only at the moment the skill is needed and found wanting. It is part of why collective, team-level reflection tends to reveal blind spots that individual introspection misses. However, with Gen AI providing solutions with less frictions than having to walk over or dial a colleague, the reducing sociality could worsen this risk.

Disappearing First Rungs of the Career Ladder

Recent work on how AI reshapes jobs, rather than whole occupations, sharpens the point. The useful unit of analysis is not the job but the bundle of tasks inside it. AI automates tasks, but organisations still hire and value bundles, and the work that resists clean automation tends to be the messy, relational, judgement-laden kind. It requires a professional to coordinate under ambiguity, resolve conflicts, know how the workflow actually happens rather than how the manual says it should. These capabilities become more valuable as routine cognition becomes abundant.

The uncomfortable implication is what one might call the broken ladder. The same technology that makes today's experts more productive can quietly remove the rungs by which novices used to climb toward that expertise. Routine work was the training ground. If it is automated away wholesale, the question arises - where would the next generation of expert practitioners come from. This is a threat that compounds silently if left unaddressed as AI becomes more advanced.

A related trap is worth naming. AI will often get a piece of cognitive work to something like ninety percent, and that ninety percent is good enough to feel finished. Yet the climb from ninety to a hundred is frequently where judgement is actually trained. If organisations let AI stop the climb early, they will save time today but forgo the very practice that produces tomorrow's expertise. Designing in the push past ninety percent, deliberately, becomes a discipline in its own right.

A Governance Challenge, Not Just a Training One

The clearest conclusion from the wider evidence is that skill preservation cannot be left to learning-and-development as an afterthought. It is a governance issue. When an institution's accountability depends on its people being able to explain, defend, and independently verify their work, and to carry institutional memory forward, the erosion of those abilities is not merely an HR concern. In a body with oversight responsibilities over markets or the economy, it can shade into a systemic risk.

This is also where comparison across peer institutions is revealing. Across central banks, regulators and government bodies, a great deal of effort has gone into helping staff use AI tools and use them safely, through tiered training, competency frameworks, dedicated learning platforms and programme evaluation. Less attention has gone to whether that adoption might be eroding the capabilities the institution depends on. Public sector bodies are aware of the risk but, while they have competency frameworks which define the skills staff should have, few appear to actively monitor whether existing skills are quietly decaying at present. That gap is an open frontier, and an opportunity for any institution willing to lead on it.

Ideas Worth Considering

Several directions follow from the evidence. Each of the ideas is meant to complement the AI governance, learning and adoption infrastructure that capable institutions already have in place.

Treat higher-order, meta-cognitive capabilities as the durable human capital investment. Critical thinking, contextual judgement, and the ability to spot what AI gets wrong are what keep human involvement meaningful. These respond well to deliberate practice, which suggests embedding them into workflow design and performance expectations rather than treating them only as a course to be completed. A practical lever could be sequencing: asking staff to engage with a problem before they see the AI's output, so the tool sharpens human reasoning rather than quietly replacing it.

Move beyond individual self-assessment toward collective, team-level skill audits, since group deliberation surfaces the blind spots that private reflection misses. Differentiate by career stage so that junior staff still build foundational expertise through direct practice, while mid-career and senior professionals focus on maintaining independent analytical capacity and oversight rigour respectively. A blanket policy serves neither.

Make skill vulnerability visible and dynamic. A living tool that tracks which tasks are being delegated, which skills are consequently exposed, and how that exposure varies across roles gives an institution a diagnostic rather than a static snapshot, one that can be revisited as the technology and the work both move. Embedding incentives so that learning is not an invisible cost of efficiency: practical formats such as verification exercises, where staff compare their own work against AI-assisted versions, and cross-functional communities of practice help make skill-building a shared, recognised activity rather than a private burden that competes with delivery.

The Bigger Picture

Public institutions, and central banks in particular, make consequential decisions under uncertainty. Their authority rests on the quality of the reasoning behind those decisions and the confidence of the many stakeholders who rely on them. AI can strengthen that reasoning, but only if adoption is designed so the technology supports and extends rather than slowly displacing the capabilities workforce and organisations depend on.

The goal is not to move slowly with AI adoption. It is to move deliberately towards efficiency gains without eroding the human capital that makes gains trustworthy. As these systems advance, this is a conversation every knowledge-intensive institution in the public sector will need to have, and to have it sooner than later.

Authors:

Abhishek Sudke, Kurnia Kusuma Wijayanti, Leyla Aghayeva, Simon Anquetil, and Najiba Taghiyeva are MPA candidates at the School of Public Policy of the London School of Economics and Political Science.