The tool changed.Now change the training.

“Your skills will be obsolete” is an excellent way to sell anxiety. It is a poor way to read the evidence.

The World Economic Forum expects 39% of core skills to change by 2030. It identifies AI and big data, networks and cybersecurity, and technological literacy as the fastest-growing skill areas. That is an employer outlook, not a verdict on an individual career. [WEF]

A changing skill often combines an old responsibility with a new method. The accountant still has to understand a discrepancy. The recruiter still has to recognise a credible candidate. The manager still has to decide whether the available evidence is sufficient. AI changes the route to the decision; it does not magically inherit its consequences.

For a professional, the useful question is therefore more specific: which parts of my work are becoming easier to produce, and which parts are becoming more important to judge?

WEF illustrates its 2030 training outlook through a hypothetical hundred workers: 41 would not need significant training; 29 could upskill in their current role; 19 could reskill and be redeployed; 11 would need training but be unlikely to receive it. [WEF]

This is a global forecast illustration, not a headcount plan for a Malta employer. Its useful lesson is structural: training demand and training access are not the same thing.

People with the least spare time can be the hardest to train. A team handling live queues cannot simply disappear into a workshop. Design practice into paid working time, make examples relevant to the actual role and give people somewhere to take uncertain outputs without feeling they have failed a test.

Domain knowledge: knowing the customer, the process, the applicable rules and the meaning of the data. Without this, a fluent error can look remarkably competent.

Evidence judgement: checking sources, noticing missing context, distinguishing a current requirement from an old blog post and separating a calculated result from an unsupported inference.

Workflow literacy: understanding where information enters, which tools touch it, who approves an action and what happens when the automated step fails.

Communication: explaining a decision in terms another person can examine. “The model said so” is not an explanation. It is a request to stop asking questions.

These layers reinforce one another. A technically skilled operator without context can automate the wrong thing beautifully. An experienced specialist without workflow literacy may see the problem but struggle to turn their insight into a repeatable system.

In Microsoft’s 2026 survey, 50% identify quality control of AI output and 46% critical thinking among human skills becoming more important. The sample consists of knowledge workers already using AI, not the entire workforce. [Microsoft]

Make review a defined activity rather than a polite expectation. Specify which facts require a source, which calculations require reproduction and which decisions need a second person. Give reviewers examples of acceptable uncertainty.

For a Malta professional serving international clients, language fluency is not enough. A beautifully written answer can still apply the wrong jurisdiction, assume the wrong customer status or ignore a contractual condition. Training should expose those mismatches rather than reward whoever produces the neatest paragraph.

A short practical assessment can be more revealing than another quiz: give someone a plausible but flawed output, ask them to identify the risk, then ask them to repair the workflow that produced it.

Junior work often contains repetition. It also contains learning. Reviewing documents, drafting a response and reconciling records can teach a person how the organisation actually operates.

If AI takes those tasks, the development path needs to be redesigned rather than quietly abandoned. Let junior staff compare drafts, diagnose errors, explain edits and join exception reviews. Pair speed with an obligation to show understanding.

Otherwise, the organisation can become very efficient at producing work while slowly losing the people capable of supervising it. That is our interpretation of a management risk, not a measured forecast of Malta’s graduate employment.

For individuals, keep a portfolio of reasoning, not just outputs. Show the starting problem, the constraints, what the model got wrong and how the final decision improved. A screenshot of a prompt is less persuasive than evidence of competent responsibility.

The European Commission’s AI-literacy guidance puts attention on the people and context in which AI is used. A certificate is not a universal assurance that an organisation’s particular deployment is appropriate. Check current official guidance for the applicable obligations. [EU]

Our proposed training loop has three parts: demonstrate a real task, practise a deliberately difficult version, then review performance on live work under supervision. Record the mistakes that recur and turn them into the next lesson.

Managers need their own version. They should be able to question a vendor’s claims, interpret an evaluation, approve access boundaries and recognise when a cost saving has merely shifted work to another team.

The objective is not to make everybody a machine-learning specialist. It is to make the people responsible for the work capable of using, checking and governing the tools that now participate in it. That is a much more useful qualification.