Efficiency now, erosion later: rethinking AI transformation in the public service
Conversations about AI in the public service focus on what we gain: speed, efficiency, lower cost. Far fewer ask what we quietly lose along the way.
Drawing on academic research from 2024 to 2026 and two decades in transformation practice, this piece argues that the real risk of rapid AI adoption is not job replacement but capability erosion. The kind that does not show up on a dashboard for years.
It introduces the "doorman fallacy" (the trap of assuming that what is hard to measure has no value), explains why AI tends to amplify the system it sits inside rather than fix it, and walks through seven deficits the research is now documenting: weakened critical thinking, a broken apprenticeship ladder, homogenised thinking, eroded trust, workplace loneliness, automation bias, and encoded power asymmetry.
The conclusion is not "use less AI." It is that governance, leadership, and institutional design (not the technology itself) determine whether AI strengthens public institutions or hollows them out. The strongest institutions will be those that integrate AI while deliberately protecting the human judgement, trust, and accountability that public systems depend on.
Efficiency now, erosion later. The challenge is designing systems where human and machine capability strengthen each other rather than quietly eroding one another over time.
As a transformation professional with more than twenty years of experience leading complex organisational change across public and private sectors, I can see both the opportunities and the weaknesses emerging through the rapid adoption of artificial intelligence.
The productivity gains are real. AI can accelerate analysis, reduce administrative burden, support synthesis across large volumes of information, and help institutions move faster in environments defined by increasing complexity and constraint.
But after spending much of my career inside large-scale transformation programmes, restructures, operating model redesigns, and institutional reform, I have learned something else: change almost always carries secondary consequences that organisations underestimate at the beginning.
Efficiency gains are usually visible early. Capability erosion often appears much later.
That tension feels increasingly relevant as governments accelerate AI adoption while reducing workforce costs and administrative overhead. Recent New Zealand Government announcements regarding further public sector headcount reductions reflect a broader international pattern. Public institutions are being asked to modernise rapidly, adopt AI-enabled ways of working, and deliver more with fewer people. At first glance, this can appear rational and even inevitable. Yet institutions can quietly weaken long before formal performance indicators reveal distress.
The doorman fallacy
Recently, I came across a concept sometimes referred to as the “doorman fallacy.” A doorman in a building may appear, from a narrow efficiency perspective, to contribute little measurable value. The role can easily be reduced to opening doors, greeting visitors, or performing routine tasks that seem replaceable during cost-cutting exercises.
But over time, organisations often discover the doorman was performing many invisible institutional functions: noticing unusual behaviour, recognising emerging issues early, maintaining social connection, preserving informal accountability, creating trust, and acting as part of the relational fabric that helped the system function coherently. Once removed, those stabilising functions only become visible after problems begin emerging elsewhere. The fallacy lies in assuming that because something is difficult to quantify, it lacks value.
I increasingly wonder whether parts of the current AI and public sector efficiency discussion risk falling into a similar trap. Public service institutions do not function through efficiency alone. They rely on judgement, legitimacy, stewardship, institutional memory, tacit knowledge, and human capability developed slowly over time. Many of these capabilities sit beneath formal structures and reporting lines. They are difficult to measure directly, yet they often hold systems together during periods of uncertainty and pressure.
AI learns the system it sits inside
One of the least understood dynamics in organisational AI adoption is that AI adapts to the context it operates within. AI systems do not arrive neutral in practice. They learn from organisational language, behavioural patterns, data structures, workflows, incentives, governance settings, and historical decisions, and over time they begin reinforcing the norms already embedded within the institution.
If an organisation already suffers from siloed thinking, poor governance, excessive bureaucracy, risk aversion, low trust, or fragmented decision-making, AI may amplify those patterns rather than correct them. Systems thinking teaches us that technology rarely transforms a dysfunctional system by itself. More often, it accelerates existing dynamics. An AI system trained within a compliance-heavy, risk-averse bureaucracy may optimise for procedural safety while suppressing innovation. A system embedded within siloed structures may reinforce fragmentation because it learns from disconnected workflows.
The question is not simply whether AI is intelligent. The question is whether the system it is learning from is healthy.
The deficits emerging in the research
The dominant narrative around AI emphasises productivity, augmentation, and competitive advantage. A quieter but rapidly growing body of academic and industry research from 2024 to 2026 is documenting a different story: the deficits, the cognitive, relational, structural, and cultural costs that accompany rapid AI integration. These findings are not anti-AI. They make visible what is being lost so that institutions can make integration choices with eyes open. Seven distinct deficits stand out.
1. Cognitive offloading and the atrophy of critical thinking
The most empirically robust deficit concerns the erosion of critical thinking through cognitive offloading: delegating mental work to AI and, over time, losing the disposition to do it oneself. Gerlich’s 2025 study of 666 participants found a significant negative correlation between frequent AI use and critical-thinking performance, strongest among younger users. A survey of knowledge workers by Lee, Sarkar, Tankelevitch and colleagues at Microsoft Research and Carnegie Mellon, published at CHI 2025, found that generative AI made tasks feel cognitively easier, with workers self-reporting reduced effort and lower confidence in their own reasoning.
Much of public service capability is built through cognitive discipline: analysing ambiguity, testing assumptions, interrogating evidence, identifying unintended consequences, and reasoning carefully through uncertainty. If AI absorbs that cognitive strain, institutions risk becoming faster on the surface while gradually weakening the depth of judgement underneath. The deeper risk is reduced institutional capacity to challenge assumptions, recognise weak analysis, and respond when conditions change unexpectedly.
2. The broken apprenticeship ladder
Public sector capability has traditionally developed through layered experience. Junior staff learned by drafting briefings, supporting operational work, conducting analysis, observing senior decision-makers, and gradually developing judgement through exposure to complexity. AI increasingly automates many of those developmental tasks. Research summarised in Communications of the ACM and MIT Sloan studies, including work by Li, Raymond and Brynjolfsson, warns that organisations may unintentionally remove the very work through which future expertise traditionally formed, and that the effect on novices’ skill formation is poorly understood and potentially negative.
Five or ten years later, agencies may discover they have fewer experienced policy thinkers, weaker implementation capability, diminished institutional memory, and thinner leadership pipelines, because the developmental tasks that used to forge experts have been outsourced to AI.
3. Homogenisation of thought and output
Where individual users gain creative leverage, collective creativity may shrink. Doshi and Hauser’s Science Advancesstudy (2024) found that AI-assisted writing improved creativity scores for less naturally creative individuals, but the similarity between stories rose measurably, indicating the collective set had become more uniform. A parallel study of Boston Consulting Group consultants by Dell’Acqua and colleagues found AI users generated higher-quality solutions, but the diversity of those solutions decreased. The effect is workflow-dependent: when humans generate initial ideas and AI supports refinement, diversity is preserved; when AI is used in early ideation, outputs converge. For the public sector, this means analysis, advice, and options across agencies, and across jurisdictions using the same tools, will trend toward a shared statistical mean, making genuine policy differentiation harder to produce and harder to recognise.
4. Erosion of trust, psychological safety, and team cohesion
AI integration, when poorly managed, corrodes the relational fabric on which high-performing teams depend. A 2026 Frontiers in Psychology study found that misunderstandings in human and AI teams reduce trust, impairing communication and outcomes. Inconsistent adoption creates “AI fragmentation,” invisible walls between teams, with skilled users feeling superior and others feeling left behind. When rules are unclear, employees hide their AI use for fear of being reprimanded, eroding psychological safety. This matters because trust is not a soft variable: Google’s Project Aristotle established psychological safety as the single strongest predictor of team performance, and large-scale public sector reform depends heavily on relational coordination between policy teams, delivery agencies, regulators, and frontline workers.
5. Loneliness and the hollowing of workplace connection
Studies in the Journal of Applied Psychology and Harvard Business Review document a counter-intuitive finding: working extensively with AI makes employees lonelier, even when productivity rises. De Cremer and Koopman’s research found that the more employees collaborated with AI, the more socially deprived they felt, with downstream effects including isolation, insomnia, and increased alcohol use. The mechanism is relational: many small, friction-laden interactions through which colleagues build rapport, asking for help, walking through a problem, are precisely the interactions AI now absorbs. Set against Gallup’s 2025 finding that only 21% of employees feel strongly connected to their organisation’s culture, the efficiency case for AI assumes connection happens elsewhere. The evidence suggests it does not happen automatically anywhere.
6. Automation bias and the decay of verification
A 2025 systematic review in AI & Society of 35 studies found that automation bias, the tendency to over-rely on automated recommendations, is amplified when AI outputs are presented with high confidence or polished formatting. People become less likely to independently verify information when it appears authoritative. A 2024 Deloitte survey reported that 38% of business executives made incorrect decisions based on hallucinated AI outputs, with consequences including regulatory non-compliance and reputational damage.
Public sector systems rely heavily on verification, peer review, challenge processes, assurance, legal scrutiny, and governance oversight. Historically, friction within systems prompted deeper review. AI reduces friction, and that is part of its appeal. But if institutions remove too much friction without preserving verification disciplines, they may unintentionally weaken their own safeguards. The risk is not simply technical error. It is institutional overconfidence.
7. Algorithmic power asymmetry and encoded inequality
The seventh deficit operates structurally: AI changes who holds power and how it is exercised. The Academy of Management Annals review by Joyce, Smith-Doerr, Alegria and colleagues identifies four facets of AI-related workplace inequality, encoded, evaluative, wage, and relational. Empirical work shows the promised neutrality is often illusory: a study of Deliveroo’s shift-allocation algorithm found it sanctioned workers without allowing them to justify absences for strikes, caring responsibilities, or health issues, and was ruled discriminatory by courts. AI used to allocate work, monitor performance, or screen candidates can reproduce and magnify bias at scale. Without explicit governance, decision rights tend to flow toward the systems’ designers and away from the workers and citizens most affected, entrenching inequalities while presenting them as objective.
Cross-cutting themes
Several patterns recur across the seven deficits.
The “feels easier” trap. The subjective experience of AI assistance is one of reduced effort, and that ease is precisely the signal of the deficit. The cognitive engagement that used to build skill and judgement is the same engagement now bypassed.
Individual gain, collective loss. Productivity and decision speed often improve at the individual level while diversity, pipeline depth, trust, and connection degrade at the collective level. Most measurement systems track the former and miss the latter.
Time-lagged costs. Broken apprenticeship, skill atrophy, homogenisation, and eroded verification produce no immediate symptoms. They appear years downstream, by which point the cause is no longer cleanly attributable.
Governance, not technology, determines outcomes. Almost every study concludes the deficits are not inherent to AI but emerge from how it is deployed and what behaviours it incentivises. Where leadership defines clear norms and preserves deliberate human practice, the deficits attenuate.
The productivity trap
The economic logic behind AI adoption is compelling. Governments are under pressure to reduce expenditure while maintaining service delivery, and treasury systems naturally gravitate toward technologies that promise measurable efficiency gains. However, public value has never been reducible to productivity alone.
Herbert Simon warned decades ago that organisations tend to optimise what they can measure while neglecting what matters most. Thinkers such as Mariana Mazzucato and Mark Moore argue that the role of public institutions is not merely to minimise cost, but to create long-term social value, resilience, trust, and capability. Those qualities rarely appear neatly on performance dashboards, yet they are the foundations on which stable societies depend. The risk is that institutions start confusing velocity with capability.
Public service work is also not purely technical. Policy, regulation, and delivery functions involve balancing competing interests, exercising judgement under ambiguity, interpreting intent, and applying discretion within legislative frameworks. A policy recommendation is not simply a technical output. It carries downstream impacts across communities, services, funding, and trust. AI can assist these decisions. It cannot hold accountability for their consequences. When systems fail, ministers are accountable, chief executives are scrutinised, and communities experience the impact directly. Public trust depends on whether systems are experienced as fair, understandable, proportionate, transparent, and human, qualities that cannot be fully automated.
A Balanced debate matters
Two unhelpful extremes are emerging. The first is technological utopianism: the belief that AI will solve institutional inefficiency simply through automation and scale. The second is technological fear: the belief that AI should be resisted entirely. Neither is useful. AI will almost certainly become embedded across public institutions because the pressures driving adoption are structural: fiscal constraints, workforce shortages, rising complexity, and public expectations for faster service. The constructive debate is not whether AI should exist, but how institutions preserve human capability, accountability, and trust while integrating it.
Designing AI transformation that holds
The frame that emerges from the research is not “use less AI” but “deploy AI in ways that preserve the human capacities AI is most efficient at replacing.” The strongest public institutions will not be those that automate the fastest. They will be those that integrate AI while deliberately protecting the capabilities public systems depend upon most. In practice, that means:
- preserving apprenticeship and mentoring pathways
- maintaining protected spaces for unaided reasoning and practice
- protecting critical thinking capability
- retaining robust verification and assurance disciplines
- strengthening institutional learning and memory
- designing workflows that put humans in early ideation, with AI supporting refinement
- supporting collaboration rather than fragmentation
- governance that distributes decision rights rather than concentrating them
- and ensuring humans remain accountable for judgement and stewardship
This is not simply a technology challenge. It is a systems and institutional integrity challenge. Public institutions are not held together by productivity alone. They are held together by trust, judgement, relationships, legitimacy, stewardship, and the collective capability to navigate complexity responsibly over time.
The cost of getting this wrong is not visible on next quarter’s dashboard. It is visible, over a decade, in the absence of a generation of experts, the homogenisation of an institution’s thinking, the quiet exit of disconnected staff, and the erosion of the institutional judgement on which public systems ultimately depend.
AI can accelerate systems. But acceleration without coherence, capability, and integrity can still produce hollow institutions. The challenge for the public service is not choosing between humans or AI. It is designing systems where technological capability and human capability strengthen one another rather than quietly erode each other over time.
That is the transformation question worth getting right.