Executive summary

There is a quiet trap catching out even the most forward-thinking teams today: deskilling our leaders through convenience.

As leaders offload more of their thinking and decision-making to AI, the capabilities leadership depends on most are quietly eroding: the ability to navigate ambiguity, hold hard conversations, and sit with complex problems.

The efficiency gains may seem worthwhile, but if AI's quality degrades or AI tools reprice to become unaffordable, it leaves us reliant on a system that doesn't work for us anymore, without the human capability to replace it. Without deliberate design, we trade deep critical thinking for instant execution, hollowing out the talent pipeline from the bottom up.

Human capability isn't a line item to minimise; it is the advantage technology is least able to replace. This paper is not an argument against AI, but a call for intentionality. The future will belong to companies with the clarity and courage to decide which capabilities must never be handed to a machine.

The choice no one is making on purpose

The real risk of agentic AI is quiet: it erodes leadership capabilities that take years to develop and cannot be recovered on demand.

Agentic AI changes what leaders do and what leaders are for. Most organisations are focused on how leaders can execute with AI while ignoring how AI is redefining a leader's purpose.

When the purpose of leaders is ignored, the central risk is quiet deskilling through convenience. As agentic systems absorb the planning and synthesis work that used to sit with leaders, the capabilities of judgement in ambiguity, navigating difficult conversations without an AI coach, and the willingness to stay with a difficult problem long enough to resolve it well quietly atrophy. Leaders lose the capability and confidence to lead without AI.

One of the first real-world signals of AI-related deskilling in clinical practice surfaced in 2025: clinicians who worked with AI-assisted detection for three months saw their unassisted detection rates fall from 28.4 per cent to 22.4 per cent once the tool was removed (Budzyń et al., 2025). This demonstrates the dependence and vulnerability that can emerge when "assistive" AI tools are introduced without deliberate safeguards.

This leaves us with two core challenges that operate in near-perfect tension:

Leadership development that ignores AI loses relevance.

Leadership development that defers to AI erodes the capabilities it exists to build.

The organisations that get this balance right will do three specific things the others will not: make conscious choices about what it means to be a leader, decide what will never be delegated to a machine, and determine how systems of efficiency and scale can be built around this conviction.

The dependency you can't price

Overreliance on AI tools in an unproven and volatile AI economy leaves organisations vulnerable.

If AI tools change in quality, are withdrawn or rise in price, or a market correction weakens the field, organisations that have ceded their judgement to AI will have no fallback.

The economics of the AI tools leaders are being urged to depend on are unresolved. In its 2026 Annual Economic Report, the Bank for International Settlements warned that disappointing AI returns could turn the current investment boom into a protracted bust (BIS, 2026). At the firm level, a 2025 report from MIT's NANDA initiative found that roughly 95 per cent of enterprise generative AI pilots delivered little or no measurable impact on profit and loss (MIT Media Lab, 2025). Its methodology has been contested, but a Goldman Sachs analysis in early 2026 similarly found no meaningful link between AI adoption and productivity across the US economy, with sizeable gains confined to a few specific uses such as customer support and software development (Goldman Sachs via Fortune, 2026).

This is not an argument that AI fails to deliver value. Whether or not it does, dependence on tools an organisation does not control is a risk in its own right. If a provider raises its prices, scales back its service or folds in a market downturn, leadership capability will not return at the speed the crisis demands. Judgement takes years to rebuild, so the gap opens exactly when leadership is needed most.

Human leadership capability should be treated as insurance rather than overhead. The organisations that enable their people to work both with and without AI tools, while pursuing innovation, are buying cover against a market they do not control.

A leadership-AI relationship model

With the right design, AI delegation can be a deliberate and intentional choice, not an accidental drift.

The Leadership-AI Relationship Model (LARM) is an original LIW diagnostic framework that maps AI capability against human ownership. It exposes how organisations can quietly drift from intentional AI delegation into passive reliance, and what deliberate design would look like instead.

The horizontal axis is how much autonomy the AI in use has. Assistive systems respond to prompts and do what they are asked. Agentic systems pursue goals across multiple steps with limited supervision. Systemic AI involves multiple agents coordinating work across the organisation, at a scale no single person can fully oversee.

The vertical axis is how much ownership the leader retains. Retained means the leader holds full accountability for the thinking and the judgement. Shared means ownership is split between leader and AI, which works only when that split is deliberately structured and governed. Ceded means the leader acts on AI output without questioning it, keeping nominal accountability while the system handles the substance.

Ownership
With agentic AI
Retained
Strong with governanceThe leader sets goals, lets agents run and reviews every output. Workable, and dependent on review discipline holding under time pressure.
Shared
Requires designThe same technology becomes shared decision-making, which is where drift most often begins.
Ceded
High riskThe leader is merely ratifying outputs. Accountability is nominally human but practically hollow.
Reading down one column of the LARM. The technology has not changed across these three cells; only the ownership has. The gravity of the model pulls downward, toward the ceded row.

The practical value of the model is that drift can be detected. This matters because it is often invisible to the people drifting. Four markers do this, some drawing on system data and others on simple review practices.

Four observable markers of drift toward the ceded row
Drift markerWhat it tells you
Ratification rateThe share of AI recommendations accepted without modification. A rate climbing towards 100 per cent signals that responsibility may be being ceded.
Time to acceptThe gap between receiving an AI output and acting on it. When this collapses, scrutiny and human judgement have likely stopped.
Documented dissentHow often a leader records disagreement with an AI recommendation. A sustained absence is a warning.
Reconstruction abilityWhether a leader can explain a recent AI-assisted decision without reopening the tool. If the reasoning lives only in the system, so does the judgement.

Research on cognitive offloading shows that people readily transfer cognitive work to external tools, particularly when a task is effortful or they doubt their own abilities, and that offloading can reduce their own memory for what they have handed over (Risko & Gilbert, 2016). Skill-retention research finds that cognitive tasks decay faster than physical ones when left unpractised, suggesting the complex, judgement-based skills leadership relies on may be especially vulnerable (Arthur et al., 1998).

Context matters too. The same cell reads differently in a workflow, in learning and development, and in leadership decisions. In a workflow, Shared and Agentic creates accountability gaps at unreviewed handover points. In learning and development, it sidelines the facilitator as agents run the learning, quietly removing the developmental friction that builds capability. In leadership decision-making, AI quietly narrows the options considered.

What the evidence already shows

AI is not just changing how work gets done. It is eroding the foundational conditions of leadership.

How people are changing

AI can make performance look strongest precisely where it is weakest. In a study of Boston Consulting Group consultants, those using GPT-4 performed better on tasks within the model's capabilities, but on a complex task outside them they performed worse than a control group, because many accepted answers that were plausible but wrong (Dell'Acqua et al., 2023). A randomised study of experienced software developers found early-2025 AI tools made them 19 per cent slower on real tasks, yet afterwards they believed the tools had made them 20 per cent faster (METR, 2025).

Objectivity and judgement may be most exposed at the top. AI assistants tend toward agreement, partly driven by the preference data used to train them (Sharma et al., 2023). OpenAI withdrew a model update within days in April 2025 after it skewed toward flattering responses (OpenAI, 2025). People given sycophantic AI counsel on interpersonal problems became more convinced they were right and less willing to repair the relationship (Cheng et al., 2025). An assistant that defaults to agreement risks automating the echo chamber. A board should treat an unconfigured assistant in a chief executive's hands as a governance gap.

Clarity collapses into finished-looking work no one can vouch for. A survey of 1,150 US desk workers found 40 per cent had received AI-generated "workslop" in the prior month, each instance costing nearly two hours to resolve by respondents' own estimates (Stanford Social Media Lab & BetterUp Labs, 2025).

Collective attention is fragmenting. Public attention to any given topic now rises and falls faster than in previous decades (Lorenz-Spreen et al., 2019). The capacity to think slowly and resist the pull of the quick answer is foundational to leadership judgement.

How organisations are changing

The pipeline is being cut from the bottom. Payroll analysis found a 13 per cent relative employment decline for workers aged 22 to 25 in the most AI-exposed occupations, while experienced workers in the same jobs held steady (Brynjolfsson, Chandar & Chen, 2025). Organisations shrinking the bottom rungs today are deciding, largely without noticing, where their 2035 senior leaders will not come from.

The non-human colleague has already arrived, and management is becoming the differentiator. An AI agent given a small in-office shop to run lost money at first, and improved once researchers added better models and the kind of scaffolding any competent manager would recognise (Anthropic, 2025).

Accountability for AI outputs is clear in principle but hard to exercise in practice. A Canadian tribunal found Air Canada responsible for what its chatbot told a customer (Moffatt v. Air Canada, 2024). In early 2024 a finance employee at Arup paid out HK$200 million (about US$25.6 million) after a video call on which every face and voice was an AI-generated fake (CNN Business, 2024).

How learning and development practice is changing

A ten-month randomised trial found an AI chatbot coach as effective as human coaches at goal attainment (Terblanche et al., 2022). However, a 2026 randomised comparison of 114 coachees found medium-to-large effects for human coaching across nearly every outcome, while the AI arm showed no significant improvement over a control group (de Haan, Terblanche & Nowack, 2026). And when 28 medical educators were asked to identify the authorship of reflective writing, they could not reliably tell student work from AI-generated work (Wraith et al., 2025).

These signals suggest the pressure is landing on the conditions for developing leaders, not on the content of leadership. That is a design problem before it is a curriculum problem.

Signals at the edge

Two developments are too early to assess but worth watching. The first is agent-to-agent dealings across company boundaries, where no human is party to the exchange. The second is institutional memory accumulating inside vendor systems rather than the organisation's own. Both would quietly shift ownership away from leadership, and both are easier to shape now than to unwind later.

Four futures, and the road between them

Without deliberate choices, organisations tend to drift towards unmanaged adoption that erodes core capability.

Two variables shape which future an organisation reaches: how far it integrates AI, and whether it does so deliberately or reactively.

Four futures for leadership development. The greatest concern is the drift from the Augmented Organisation into the Automated Default.

The goal is not a particular level of AI but a conscious choice about how to engage with it. Klarna's arc is an early warning: it shifted much of its customer-service work to AI in 2024, then began bringing human agents back in 2025 after quality fell (Fortune, 2025).

Organisations rarely choose the Automated Default; they arrive there one sensible decision at a time. Watch for the tripwires:

  • A learning and development budget cut with AI coverage as the stated rationale.
  • No one practises working without AI.
  • Human facilitation retreating to an annual set piece.
  • A ratification rate rising quarter on quarter.
  • Little transparency or shared learning about how AI is used.

What to keep, and how to decide

Three questions separate the skills worth ceding to AI from the skills worth keeping.

Humans have always ceded skills to technology. Few leaders can steer by the stars or do long division at speed, and nothing that mattered was lost. The real question is which capabilities are legacy and which are load-bearing, and it needs a test.

The cede or keep test
TestThe questionKeep it if
RecoveryIf the system failed or repriced tomorrow, could we function while rebuilding this?Rebuilding takes years, and the capability needed could emerge mid-crisis.
SupervisionIs this capability required to check the system's own output?Oversight depends on the very skill being ceded.
DevelopmentDoes exercising this build anything beyond the task itself?Growth comes through practice; cede the task and the growth goes with it.

In our work, running the test across a leadership team's ten most consequential capabilities produces a keep list that tends to map onto LIW's 4 Practices: being present, staying curious, making conscious choices, and continuing to experiment. The convergence is not proof of either model; the two ask a related question by different routes and arrive at a similar answer.

Three conditions shape the outcome: productive friction, which can produce stronger long-term learning (Bjork, 1994); reflective capacity, where depth of reflection drives real change (Moon, 2004); and genuine dissent, which improves decisions even when the dissenting view is wrong (Nemeth, 1986; Nemeth, Brown & Rogers, 2001).

Rebuilding the pipeline is a design job. LIW's DUSA learning model (Diagnose, Understand, Solve, Apply) brings focus to deliberate practice through hard situations, under genuine pressure but within a controlled environment, generating the developmental friction that entry-level roles once supplied for free.

The AI side is a design lever too

How an organisation configures and procures AI tools is one of the few direct levers it has to preserve human judgement.

Interfaces built with cognitive forcing functions, which require the user to engage before the AI's answer appears, reduced overreliance on wrong recommendations, though participants liked them less (Buçinca, Malaya & Gajos, 2021). Across 319 knowledge workers, higher confidence in the AI tool was associated with less critical thinking (Lee et al., CHI 2025). The difference between an oracle and a thinking partner is a configuration choice. Six criteria turn that into a procurement standard:

Position first
The tool withholds its recommendation until the leader defines their own position.
A practical off switch
Teams can run without it for defined periods without workflow collapse.
Challenge on demand
It can argue the other side, red-team a plan or surface the strongest objection.
Uncertainty made visible
It distinguishes what it knows from what it is estimating or inferring.
Reasoning transparency
It exposes its reasoning, so the leader can interrogate the argument.
Acceptance tracking
It records how often people accept its outputs unchanged, giving you the ratification rate.

Governing capability risk

Capability risk must be governed in the language of the board and embedded in risk management processes.

  1. Put capability erosion on the risk register. Treat it the way cyber risk moved from a technical concern to a standing board item: a named owner, leading indicators drawn from the drift markers and tripwires, defined thresholds and a review cadence.
  2. Define a clear decision-rights line. For each material class of decision: who decides, what the AI contributes, and who answers when the AI contribution is wrong.
  3. Fund a judgement reserve. A protected share of development spend that cannot be cut on the grounds that AI can replace it, reviewed like a capital buffer.

The organisational equivalent of aviation's manual flying practice is the AI-free decision drill: a live case worked by the senior team without AI tools, on a schedule, observed and debriefed. It is inexpensive, it gives a direct read on reconstruction ability, and it keeps the recovery test answerable.

The first ninety days

Don't start with "How do we use AI?" Start with "What do we want to achieve and why?"

The LARM answers "Where are we?" The Leadership Impact Chain answers "Why does that matter?", working backwards from the outcome so that behaviour comes last:

  1. Business impactWhat the organisation exists to create
  2. Aligned performanceHow the work must be done to deliver it
  3. Conditions for successClarity, climate and competence
  4. Leadership behaviourDerived last, to create the conditions

LIW's 3Ws name the sequence: W1, what human leadership is for and why; W2, an honest read of where you are, which is where the LARM applies; and only then W3, what to do.

  • Define what human leadership is for, with the executive team, before assessing the role of AI.
  • Choose at least one protected human practice and defend it. It can be as simple as a routine meeting kept device-free.
  • Map the portfolio against the LARM. Most teams find they are designing for fully human leadership while operating in a shared human-AI reality.
  • Run the cede/keep test on your ten most consequential leadership capabilities, and circulate the keep list as a shared standard.
  • Identify an owner for capability risk, tasked with monitoring the five tripwires.

Co-evolution as the goal

The organisations that protect the conditions in which good leadership develops will hold a real and durable advantage by 2030.

As the routine dimensions of leadership are systemised, what remains is the hard part: judgement in complexity, holding others through uncertainty, working through genuine ethical dilemmas, leading with presence. The organisations that proactively navigate and design for the transition will be the ones worth working for in 2030.

References