The skills we need in the AI age

Many conversations about AI skills end in the same place: everyone should learn to prompt better. That is useful advice. It is also nowhere near enough.

I have been part of a separate set of ongoing conversations about what Australia needs to succeed in the age of AI. We keep returning to the question of skills, and it is surprisingly easy to talk past one another. “AI skills” has become a catch-all term for several quite different capabilities.

Over the past few years, I have spoken with executives, public servants, educators and people trying to make sense of AI in their everyday work. Most are not trying to become AI engineers. They are asking practical questions. What should I be able to do with these tools? What do I need to understand before I trust an output? And how do we ensure Australia retains the capability to build and govern the technology we rely on, rather than simply renting it from someone else?

Those are different skills problems. Treating them as one will give us shallow AI training, poorly governed experimentation and staff who can produce a plausible answer without knowing whether it is any good.

Four levels of AI capability

Most discussion of AI skills still collapses several quite different capabilities into one vague category. People are told to become “AI literate”, then left to work out whether that means using a chatbot, designing an agentic workflow, managing an AI vendor or building a model.

Those are different jobs. They require different skills, different levels of accountability and different kinds of education. I see four levels of AI capability that organisations and governments need to think about seriously:

1. AI use and judgement: Most people will need this level. It means using AI tools competently in everyday work, giving clear instructions, checking outputs, protecting sensitive information and recognising when an AI tool is unsuitable for the task. It also means retaining responsibility for the work, rather than treating a plausible response as an answer.
2. AI orchestration and workflow design: This is for people redesigning work around AI. They may build no-code automations, create knowledge bases, connect AI tools to approved systems, design agentic workflows and determine where a human needs to review or approve an action. They may never train a model, but they will be creating operational processes that affect customers, staff, services and decisions. That is a consequential capability, and it needs more than enthusiasm for a new tool.
3. AI governance, assurance and leadership: This is the work of leaders, boards, risk and legal teams, procurement, cybersecurity, data professionals and policy people. It includes deciding where AI should and should not be used, setting risk appetite, assessing suppliers, establishing controls, monitoring performance and impacts, and ensuring accountability remains visible. Governance cannot be delegated to a vendor, buried in a policy or left to whoever built the workflow.
4. AI development and technical infrastructure: This is the deeper technical capability needed to build, adapt, secure, evaluate and operate AI systems. It includes LLMs, LRMs, generative AI, machine learning, statistics, software and data engineering, cloud infrastructure, cybersecurity, model evaluation, AI safety and research. Australia needs more of this capability, especially in areas where we cannot afford to depend entirely on technology designed, hosted and controlled elsewhere.

These levels overlap. Someone designing an agentic process needs enough AI literacy to understand its limits. An AI engineer needs to understand governance and the real-world setting in which a system will operate.

But each level carries a different kind of responsibility. Using a chatbot competently does not equip someone to design an AI-enabled business process. Designing that process does not make someone capable of governing its risks. And governing a system is different again from developing the technical infrastructure beneath it.

AI use and judgement

For most people, the immediate task is learning to use AI well in the work they already do. They do not need to build a model or design an agentic workflow. They need to understand enough to use a tool productively without creating problems for themselves, their organisation or the people they serve.

That begins with framing the task properly. AI tools produce better results when someone can explain the context, intended audience, constraints, relevant material and what a useful outcome looks like. People call this prompting. I think it is better understood as briefing well.

The practical skills include:

• Breaking work into sensible steps before handing any part of it to an AI tool
• Giving the tool enough context, examples and boundaries to produce something useful
• Asking for alternatives, checking assumptions and challenging the first answer
• Verifying facts, sources, calculations and claims before using an output
• Editing material so it is accurate, useful and appropriate for the organisation
• Knowing what information must never be entered into a public AI tool
• Keeping a record of AI use where the work is consequential, regulated or may later need to be explained

These are ordinary professional skills, made more important by AI. A poor brief produces poor work whether it goes to a colleague, a consultant or a large language model.

For a small business, this may mean using AI to draft a customer communication, find themes in survey responses, produce a first-pass marketing plan or prepare a meeting agenda. The value does not come from pressing a button. It comes from the person who understands the business, the customers and the consequences of getting the work wrong. Someone still has to decide whether the output is accurate, useful and safe to use. That responsibility does not disappear because an AI tool produced the first draft.

Skills for understanding AI

Using AI effectively is not enough. People also need enough AI literacy to recognise when a system is likely to be wrong, biased, unsafe or simply unsuitable for the task.

This means understanding some basic truths about how current AI works:

• Generative AI predicts plausible patterns in data. It does not inherently know what is true.
• A polished answer can still be incomplete, biased or entirely fabricated.
• AI systems reflect choices about data, objectives, design and deployment.
• Risk depends on context. A weak draft of a social-media post is not the same as an AI-supported decision about welfare, recruitment, health, credit or policing.
• Human oversight only matters if the human has the authority, time, expertise and confidence to challenge the system.

This is where critical thinking becomes more important, not less. The World Economic Forum’s 2025 research identifies analytical thinking as employers’ leading core skill, alongside resilience, flexibility, leadership, creative thinking and technological literacy. It also identifies AI and big data, networks and cybersecurity, and technological literacy among the fastest-growing skill areas.

AI literacy should include privacy, confidentiality, intellectual property, security, bias, accessibility, environmental impact and accountability. It should also include the simple but vital question: should we use AI for this at all?

That question is especially important in government and other high-trust settings. We should not mistake a technically possible use case for a legitimate, proportionate or socially acceptable one.

Skills for building AI

Then there is the building AI layer: the capability to develop, deploy, evaluate and govern AI systems. This is where higher education, vocational education, research institutions and industry all have serious work to do.

Australia needs people who can build systems, not just buy subscriptions from overseas vendors. This is a question of economic capability, but also of digital sovereignty, national resilience and the ability to make informed choices about the technologies embedded in our critical services.

The required skills are broader than data science and machine learning. They include:

• Mathematics, statistics and computational thinking
• Software engineering, data engineering and cloud infrastructure
• Machine learning, natural language processing and model evaluation
• Cybersecurity, including adversarial testing and secure AI system design
• Data governance, data quality, privacy engineering and information management
• Human-computer interaction, accessibility and service design
• AI assurance, audit, monitoring and incident response
• Ethics, law, public policy and domain expertise
• Communication and multidisciplinary collaboration

The most useful AI builders will not be people who treat technical capability as separate from social consequences. They will be able to work with subject-matter experts, understand the operating environment and recognise where a technically impressive system may create unacceptable risk.

Higher education has a larger job

Universities should not respond to AI simply by trying to detect whether students have used it. That is a narrow and ultimately losing strategy.

The larger task is to help students develop the capabilities that remain valuable when AI is readily available: expert judgement, disciplinary knowledge, critical inquiry, practical problem-solving, ethical reasoning and the ability to explain and defend a decision.

Australia’s framework for AI in higher education puts human-centred education, equity, ethical innovation, professional learning and research integrity at the centre of AI adoption. That is the right direction, although the real test will be whether institutions invest in staff capability, redesign assessment thoughtfully and give students meaningful opportunities to learn with AI rather than merely around it.

Students also need to understand that AI fluency is not a substitute for learning. If you cannot assess whether an AI-generated answer is sound, you are outsourcing judgement at precisely the point you need to develop it.

What organisations should do now

Organisations do not need a heroic AI transformation program before they start building skills. They do need a deliberate approach.

A sensible starting point is to:

1. Define which AI tools are approved, which uses are prohibited and where people should seek advice.
2. Provide baseline AI literacy for all staff, including privacy, security, verification and appropriate use.
3. Develop role-specific capability for leaders, procurement teams, risk and legal functions, technical teams and frontline staff.
4. Create safe, bounded opportunities to experiment with low-risk use cases.
5. Measure whether AI is improving quality, safety, productivity or service outcomes, rather than simply increasing activity.
6. Build governance and assurance into projects from the beginning, not after a system has been deployed.

The OECD has warned that the supply of training may not be sufficient to meet growing demand for both general AI literacy and specialised AI expertise.

Human judgement is still the work

AI will change the shape of work, but it does not eliminate the need for human capability. It makes human capability more consequential.

We will need people who can formulate a problem properly, spot weak evidence, understand context, challenge a confident but wrong answer and take responsibility for a decision. We will need technical people who can build systems securely and responsibly. And we will need leaders who understand that buying an AI product is not the same as building organisational capability.

Australia should be aiming for more than a workforce that can use other people’s AI. We need people who can assess it, govern it, improve it and, where it matters, build it ourselves.