Skills fade in the age of AI:
a resilience question for boards

There has been considerable rhetoric from technology leaders suggesting white collar work will not exist in eighteen months. At the same time, software engineers are reporting AI fatigue and a loss of confidence in their own skills from constant AI oversight. These are not separate concerns. They are early warning signals of a resilience gap. McKinsey's State of AI report for 2025 shows 88 percent of organisations now use AI in at least one function, but only a third have begun scaling it across the enterprise.

I do not buy into the short-term horizon being described, and I find Westpac's recent analysis of the AI investment boom makes an important distinction worth holding onto: the intellectual property of a software firm is less that they have people who know how to code, and more that they have people who know what makes a good payroll system, or drawing software, or whatever the product may be. That design knowledge is harder to replicate through AI than the code itself.

This piece examines a specific concern sitting at the intersection of AI adoption, organisational resilience, and professional capability. As we progressively do less foundational work and move into AI oversight, what happens to the underlying human skill, and what does that mean for organisational resilience over time.

Why professional work will still exist

Westpac's analysis draws an important line. Organisational intellectual property sits less in execution capability and more in contextual judgement. AI is strong at process. Professional work requires understanding organisational history, navigating political dynamics, and making decisions when rules conflict with each other.

The complication is that contextual judgement is built on foundational skill. A lawyer who understands legal strategy but has never personally researched case law cannot validate whether AI-generated research is sound. If professional roles evolve into purely supervisory functions, and the foundational capability beneath that judgement is allowed to erode, how does anyone validate AI outputs or operate independently when the systems fail?

The infrastructure dependency problem

AI is infrastructure-heavy. It requires power, data centres, network connectivity, and cloud access. In an increasingly contested global environment with escalating threat profiles, it is reasonable to plan for scenarios where AI systems are unavailable, not for an hour, but for a sustained period.

Cyber attacks that take down cloud providers, severe weather events that cut power for days or weeks, and infrastructure failures that make data centres inaccessible are real risks rather than hypothetical ones. The Allianz Risk Barometer for 2026 identifies AI as the second-highest global business risk, noting that adoption is outpacing both governance and workforce readiness. ASIC has named operational resilience, AI governance maturity, and cyber risk as key supervisory issues for 2026.

The skills fade pattern

As AI use increases, foundational skills may degrade simply because they are no longer being exercised. Software engineers are already describing both the fatigue of working alongside AI constantly and the loss of their own confidence and capability as a result.

The consequence becomes visible during sustained infrastructure disruption. When outages caused by cyber incidents, severe weather, or other infrastructure failures take hold, teams cannot deliver critical work effectively without AI. The system they have come to depend on is unavailable, and the capability to work without it has quietly eroded in the meantime.

This concern intensifies when the timeline is extended. As trust in AI increases, automation scales and human roles become increasingly supervisory. If AI becomes embedded in mission-critical workflows, what happens when systems are unavailable for days or weeks, not today, when senior professionals still remember how to work manually because they built their expertise before AI was widespread, but in 2035 and beyond, when much of the workforce may have trained entirely with AI throughout their careers. When the lawyers who learned legal research through AI tools become partners. When the clinicians who used diagnostic AI from residency become department heads. When the analysts who never built a model manually become team leaders.

Skills assumptions and organisational resilience

Business continuity conversations are starting to emerge for the AI era. The frameworks recognise that humans may need to perform AI's role during a system failure, and that redundancy of skill matters. But they appear to assume those skills will still exist when needed. If resilience is defined as the ability for an organisation to sustain operational output under adverse conditions, that assumption may not hold unless the capability is deliberately preserved.

The professionals who can validate AI outputs today learned their craft before AI was ubiquitous. They built legal arguments, diagnostic reasoning, financial models, and policy analyses manually for years before AI tools became available, which is precisely what gives them the foundational capability to recognise when AI outputs are sound and when they are breaking down.

What is missing is a deliberate pathway for the next generation to build that same foundational capability, and a willingness to keep using those skills actively in the short term rather than letting AI absorb them by default.

What resilience could look like

Emergency response services, aviation, and the military routinely train for infrequent scenarios specifically to prevent skill degradation. Aviation offers a useful reference point. Despite autopilot systems handling the vast majority of flight operations, airlines deliberately preserve manual flying capability through regular simulator rehearsal of scenarios a pilot may statistically never encounter across an entire career.

That capability is preserved precisely because it may be used infrequently while still being critical during system failures, extreme weather, or other scenarios where the automated system cannot be trusted. Aviation recognised decades ago that autopilot outputs cannot be validated, and manual operation cannot occur during an emergency, if pilots have lost the foundational skill required to fly the aircraft.

Over time, this preserves a working level of foundational capability required to validate AI outputs and operate when systems fail. It treats the capacity to work without AI as a strategic asset that requires deliberate cultivation rather than something that simply persists on its own.

Knowledge-based organisations could adopt a similar principle: periodic resilience exercises where teams practice first-principles work without AI, simulating extended cyber disruption or infrastructure failure as a discipline rather than a contingency footnote.

In the context of professional capability, organisations are effectively making a bet: that AI systems will never experience sustained failures requiring manual operation, or that professional capability can be rapidly rebuilt when needed, or that the cost of being unable to operate during disruption is an acceptable one.

A question for leadership

As organisations accelerate AI adoption in an environment where cyber threats, weather events, and infrastructure vulnerabilities are genuine risk factors, will there be investment in AI capability alongside protection of the professional capability required to operate without it?

The answer becomes visible over time, in how organisations respond when AI systems are unavailable for an extended period, and in whether the professionals responsible for validation still hold the foundational skill required to do it.

A reasonable question worth raising for boards and executive leadership is whether capability preservation is being addressed as a design consideration in AI transformation, or treated as an afterthought.

Organisations approaching AI as enterprise redesign have an opportunity to build resilience into the transformation from the start. Those treating it as a technology upgrade may find the capability assumptions sitting inside their business continuity plans do not hold when they are actually tested.