Your team finished the AI course. Can they challenge the machine?

A dashboard can indicate how many of a company's employees have completed an AI course but it cannot gauge whether those individuals possess the ability to recognize when an AI-generated answer is incorrect. This gap in understanding has become increasingly pressing. The World Economic Forum's Future of Jobs Report 2025 revealed that 77 percent of surveyed employers plan to reskill or upskill their workforce to collaborate more effectively with AI.

However, half of the executives questioned separately pointed to a lack of skills as the primary obstacle to AI adoption. The International Labour Organization estimates that one in four jobs globally could be affected by generative AI, with a higher likelihood of transformation rather than complete replacement. While the quantity of training is important, merely measuring enrolment, hours, and certificates may give the impression of readiness without fostering the critical judgment required to utilize AI effectively. Completion of a course signifies exposure, not competence.

A certificate indicates that someone has finished a course but does not demonstrate their capability to select suitable tasks for AI, safeguard sensitive information, verify an output against reliable sources, or explain why a human should still make the final decision. This becomes crucial even in smaller organizations. An impeccably crafted yet inaccurate response can quickly turn into a customer reply, a hiring recommendation, a financial assumption, or an operating instruction before anyone has a chance to question its validity.

Fluency can introduce a new risk: the more natural the tool appears, the easier it is to conflate confidence with correctness. Basic prompt-writing and feature demonstrations can serve as useful starting points, but they are insufficient.

Singapore's goal of developing AI-competent bilingual workers offers a more insightful perspective on capability. People require both AI fluency and knowledge of their own domain. The true test is whether they can bridge the two. A practical assessment should evaluate three key aspects: explanation, decision, and transfer. The worker should be able to explain the tool in their own words, making clear distinctions between reliable and unreliable information.

They must also demonstrate the ability to make sound decisions under uncertainty by considering factors such as when to involve another person, when to halt the process, and what would warrant such a decision. Finally, the worker should be able to transfer their knowledge into a real workflow, identifying an appropriate task, using an approved tool, verifying the results, and maintaining clear human accountability.

These evaluations do not necessitate expensive certifications but can be conducted through a 20-minute exercise. The learner is given a routine fictional or properly sanitized task, instructed to use AI, identify two uncertainties in the response, verify the critical claims using approved sources, and explain their next steps. A follow-up question asks what would make them decide not to use the output, with the exercise being repeated after a month to assess retention and judgment.

Measuring performance rather than participation is crucial. While course completion can remain an administrative figure, it should not be the primary proof of success. Useful operating metrics could include the proportion of AI-assisted work that passes human review, avoidable rework, appropriate escalation, time saved, and the number of unsupported claims detected before use.

The goal is not to penalize individuals for identifying errors but to encourage a culture where catching a plausible mistake is evidence of critical thinking. Teams should also engage in open discussions about near misses without fear of embarrassment. Employees must feel empowered to say, "I cannot verify this," even when the answer seems convincing.

Experienced workers are essential in this process. They possess a deep understanding of the workflow, including the awkward exceptions, weak signals, and informal checks that may be absent from a neat process map. Their caution should not be dismissed as resistance; instead, it signifies where an AI-supported workflow may be brittle.

Founders should involve experienced employees when designing scenarios, determining what requires review, and defining a good outcome. The domain knowledge of experienced colleagues is not an obstacle to adoption; it is what makes adoption beneficial. Newer employees may bring confidence in the technology, while experienced colleagues contribute context.

AI readiness flourishes when both groups learn from each other. This approach aligns with the ASEAN Responsible AI Roadmap, which emphasizes employment outcomes, industry adoption, industry engagement, and inclusion beyond training participation. For companies in Southeast Asia, the key takeaway is to focus on what individuals can accomplish after training, not merely who attended the training.

A better question for leaders is how to assess practical judgment and ensure that employees are not only proficient in using AI but also capable of identifying its limitations and acting responsibly.