Three variables, and only one you control

Jobs and Skills Australia's framework for understanding generative AI's labour market effects has three terms: exposure, how widely the technology can be applied; adoption, how deeply it is taken up; and adaptation, how workplaces change over time. Its finding is that impact "depends on exposure, adoption, and adaptation" together.

An individual organisation does not set the exposure of its occupations, and only partly controls the pace of adoption in its market. Adaptation is the variable it genuinely owns. It is also the one least likely to have a budget line, an owner, or a measure attached to it.

This is what separates organisational capability from tool deployment. Deployment is a procurement event. Capability is the organisation's standing ability to absorb a change in how work is done, and then to absorb the next one.

Why event-shaped programmes decay

The usual pattern is a licence purchase, a training session, an acceptable-use policy, and a note in the board pack. Twelve months later, usage is uneven, quality is unmeasured, and the training is out of date.

The research explains the decay rather than just describing it. JSA found that generative AI is "intensifying the pace of occupational skill evolution", and separately that this creates a need for "flexible and timely updates to training". If the skill content of roles is turning over faster than it used to, then any fixed curriculum is depreciating from the day it is delivered. The problem is structural, not a failure of the particular course.

The corollary is that the useful unit of investment is not a programme but a mechanism: something that notices when practice has moved and updates what people are taught. That is a durable capability. A completed training day is not.

Different populations need different things

JSA's skills analysis identifies seven distinct workforce personas in the transition, spanning leaders, professionals, educators and affected workers, and concludes that each requires "tailored skillsets and support".

Most organisational programmes ignore this and deliver one thing to everyone, which reliably lands wrong in both directions. Executives receive operational detail they will not use. Frontline staff receive strategy framing that does not tell them what to do on Monday. The people whose roles are actually changing receive the same generic material as everyone whose roles are not.

A workable structure separates at least three populations: everyone, who needs to know what is approved, what must never be entered into a tool, and that output requires review; the teams whose work is changing, who need evaluation skills specific to their domain; and those accountable for decisions made with AI support, who need to understand what they are signing off. Our note on capability uplift in Australian organisations sets out a tiered model along these lines.

Capability and governance are the same problem

It is tempting to treat capability as an HR initiative and governance as a compliance one. In practice they fail together.

An organisation that deploys tools faster than its people can evaluate output has not created a productivity gain, it has created unreviewed work product moving through the business at speed. The control that catches an AI error is a person who knows what a wrong answer looks like in their domain. If that person does not exist, no policy document substitutes for them. This is why JSA's finding that demand is rising for critical thinking alongside digital literacy matters operationally rather than rhetorically.

The same logic runs the other way. Governance without capability produces policies nobody can follow, because staff cannot tell which uses fall inside them. Capability without governance produces confident, capable, unaccountable use. Organisations building either one should be building both, and the reporting line to the board should reflect that they are a single question. For the structural side of this, see our material on AI governance.

Adaptation takes longer than the business case assumes

JSA's observation on complex use cases is worth quoting because it contradicts most vendor timelines: "where use cases are complex, technology often requires development, training" and sustained work to embed. It notes that some firms are only early in large-scale change management programmes, which is why employment effects have not yet emerged in the data.

The study draws an instructive historical parallel. When electricity was introduced, the productivity gains did not arrive with the technology; they arrived after factories were reorganised around what the technology made possible, which took decades. JSA expects a similar lag with generative AI, with adaptation and reorganisation following adoption rather than accompanying it.

For planning, that argues against both extremes. Organisations that expect returns in a quarter will be disappointed and may abandon the work early. Organisations that conclude nothing is happening because nothing has happened yet will find the reorganisation harder when it becomes unavoidable.

What sustained capability looks like in practice

Drawing the evidence together, a capability programme that is likely to survive contact with a moving technology has these properties:

  • An owner and a measure. Adaptation is the controllable variable, so it needs the same accountability as any other managed risk. Usage counts are not a measure of capability. Whether people catch errors is.
  • An update mechanism. Given intensifying skill velocity, build the loop that refreshes content, not just the content.
  • Differentiation by population. At minimum, separate baseline literacy from role-specific evaluation skill from decision accountability.
  • Evaluation taught as a first-class skill. The capacity to recognise a plausible wrong answer in your own domain is the control that actually operates.
  • A deliberate route for developing judgment. Where routine work that used to build expertise is automated, replace the development pathway on purpose rather than discovering its absence later.
  • Equity as a design input. Exposure is concentrated, so untargeted investment predictably under-serves the most affected cohorts. JSA's finding that skills development "must be inclusive" is a design constraint, not a sentiment.
  • A realistic horizon. Plan for adaptation on the timescale the evidence suggests, and resist business cases that assume the reorganisation is free.

None of this requires certainty about how the technology develops. It requires the organisation to be able to change how it works more than once, which is a more defensible thing to invest in than a bet on any particular tool.

This article discusses organisational strategy and workforce capability. It is general information only and does not constitute legal, employment, financial or professional advice. Obligations relating to consultation, restructuring and the use of AI in employment decisions vary by jurisdiction and should be checked with qualified advisers.

Related reading

Sources: Jobs and Skills Australia, Australia's AI Transition: Jobs, Skills and the Future of Work · Our Gen AI Transition: Implications for Work and Skills (14 August 2025) · ILO Working Paper 140 (20 May 2025)