As companies train employees to work with more capable AI systems, course completion alone says little about whether people are actually becoming better at their jobs. The harder question is what changes in day-to-day work after training, including how employees judge AI output, decide what to delegate and use time saved by automation.
In this TNGlobal Q&A, Felicia Tan, Director of Tribe Academy, discusses how employers can measure AI training after 30, 60 and 90 days, why domain judgment remains durable as tools change, and how job design needs to evolve as AI becomes more agentic.
The discussion builds on a broader productivity question TNGlobal also recently explored with Workday ASEAN: faster individual tasks do not automatically translate into better organizational outcomes.
At the National Day Rally, PM Lawrence Wong highlighted how AI has moved from systems that answer questions to agents that can carry out tasks. From a workforce-development perspective, what changes when employees need to learn not only how to use AI, but how to supervise systems that can act on their behalf?
Most of what we think of as “using AI well” today assumes a fairly simple interaction: asking a question, receiving an output, reviewing it, and making a decision. Once we move into AI agents that can carry out tasks on our behalf, we are now supervising a chain of decisions and actions, some of which may happen without us seeing every step. This changes the nature of the risks involved.
From a workforce-development perspective, employees need to build three key capabilities.
First, they need to understand what should and should not be delegated. Much of today’s conversation focuses on what AI can do, but not enough attention is paid to what it should never do without human involvement. Employees need to distinguish between work that can be fully handed off because it is repeatable and rules-based, work where AI can assist but still requires human approval, and work where judgement, accountability, or relationships remain central.
Second, employees need to learn how to test AI systems rigorously before implementation. It is not enough to run a workflow once and decide the output looks reasonable. Organizations need to test standard cases, edge cases, and real-world data to ensure the system performs reliably. Just as importantly, workflows should be designed for graceful failure, where issues are surfaced clearly rather than allowed to compound.
Third, organizations need to protect the development of human judgement. Delegating repetitive or mechanical work makes sense, but if employees hand over difficult conversations, prioritisation decisions, or judgement-heavy tasks too early, they risk weakening the instincts they need to assess and challenge AI effectively.
The goal should not be to automate as much as possible, but to automate in a way that removes low-value effort without reducing human capability.
Many training programs can measure enrollment, attendance and course completion. What should an employer look for 30, 60 or 90 days later to determine whether AI training has actually improved how people work or make decisions? Are there particular measures organizations tend to overvalue?
At 30 days, organizations should look at whether employees are actually using AI in their day-to-day work. This is usually when the novelty wears off. People either integrate it into their workflow or quietly return to their previous ways of working. This is best assessed through actual usage patterns: login frequency, feature usage logs, or output volume rather than self-reporting, since people often overestimate adoption when asked directly.
At 60 days, the focus should shift from usage to quality of intervention. Organizations should review samples of AI-assisted work and assess them against the same standards applied before AI: was anything incorrect, and did the employee catch it before the work was shared?
At 90 days, organizations should look at whether saved capacity is translating into higher-value work. If the hours saved have not moved anywhere, either the role has not been redesigned around the gain, or the automation was not addressing a meaningful bottleneck.
One measure organizations should be careful about overvaluing is output volume. So “We produced 100 social media posts in a day” says nothing about whether the quality of those posts has improved. The stronger measure is whether people are completing work faster without sacrificing quality, and whether that saved time is being redirected towards more valuable activities.
AI tools and models are changing quickly. Which capabilities are proving durable enough to transfer across different tools and generations of technology? For example, how important are domain judgment, breaking down a task, evaluating an AI-generated result, asking useful questions and knowing when to escalate to a person?
The better AI models become, the more convincing their outputs appear. However, fluency does not mean judgement. AI can produce responses that appear considered without possessing the lived experience, organizational context, or domain judgement behind the problem, and that gap becomes harder to spot as models improve.
This is where domain judgement becomes critical. For example, an analyst who understands that a client operates on an unusual fiscal year may immediately recognise a flawed revenue projection. Someone without that context may simply accept the answer because nothing about the output appears obviously wrong.
Other skills, such as breaking down tasks, asking useful questions, and evaluating AI-generated results, can be taught and applied across different tools. Domain judgement is different. It is built through experience, what people often describe as “gut feel”, and it remains one of the most important human capabilities in working effectively with AI.
What does superficial AI fluency look like in the workplace? Can an employee become faster at producing work with AI while becoming less reliable at checking outputs or understanding the underlying problem?
Superficial AI fluency is when someone becomes very good at getting polished work out of AI, but not necessarily better at judging whether that work is correct, relevant, or fit for purpose.
Someone producing more output, faster, using all the right tools, can in fact be getting worse at the actual job. Speed and reliability are not the same axis, and AI makes it easy to mistake one for the other.
It is like the difference between someone who has memorised a route using GPS and someone who actually understands the city. Both may arrive at the destination, but when the GPS fails or the route changes unexpectedly, the difference becomes clear.
Managers should look out for warning signs such as work being produced unusually quickly but not being defensible when questioned, employees struggling to explain why a recommendation was made, outputs becoming more generic and less sensitive to context, factual errors slipping through, or increasing dependence on AI even for tasks where independent reasoning should still be developed.
As AI becomes more agentic, some roles may shift from directly performing tasks towards setting objectives, reviewing exceptions and supervising automated workflows. How should organizations rethink job design, accountability and training?
The move towards agentic AI does not mean people should simply be moved into supervisory roles. The people best placed to catch AI mistakes are often those who still understand the work deeply because they have actually done it.
We have already seen examples of this challenge emerging. Ford, for instance, recently hired back 350 veteran engineers after AI-driven design processes struggled with vehicle quality issues, highlighting that AI outputs still require human expertise and judgement.
Hence, the better move is to keep people rotating through real hands-on work alongside supervision, so accountability stays meaningful rather than just a formality. Automated tools also depend on the training and expertise that experienced employees bring.
How much of successful AI reskilling depends on the employee, and how much depends on the organization redesigning the surrounding workflow?
Successful AI reskilling requires both. Employees need curiosity and confidence to experiment, but training only creates value if they have somewhere to apply those skills immediately.
Otherwise, adoption remains highly individual. One employee may save two hours a week using AI in their own workflow, but without shared playbooks, institutional memory, or redesigned processes, the gain lives and dies with that individual.
Ultimately, organizations determine whether AI remains a personal productivity tool or becomes an enduring new way of working. This requires identifying early champions, creating space to test and share what works, rather than leaving them to experiment alone.
Managers also need to understand how AI changes workflows. If leaders continue evaluating people against old processes, they may unintentionally undermine the very behaviors that AI training was meant to build.
Which workers are most at risk of being left behind as AI adoption accelerates?
The workers most at risk are not necessarily those with lower technical skills. Technical ability was never the biggest barrier because people do not need to code to use AI effectively.
The bigger factor is when someone starts. Someone who began experimenting with AI earlier has experienced the technology evolving step by step. They have developed an instinct for where AI is reliable, where it fails, and how to work around its limitations.
However, domain expertise remains a powerful advantage. From our experience conducting AI training for the past two years, someone who understands their function deeply can often get more value from AI than someone who only understands the tool itself. Domain knowledge gives AI the context it needs to become useful.
This is also why fear of being replaced is the wrong reason to wait. Sitting out does not protect a role; it just delays finding out how AI complements or threatens the work, and leaves less time to adapt once that becomes clear. All scenarios considered, there is no better time to start learning about AI than now.
If AI capability continues changing quickly, what should continuous learning look like in practice?
If AI models change every few months, the question is not how to repeatedly train people on new tools, but whether training gives people the skills to keep adapting as those tools change.
Good AI training should focus on the underlying skills that remain relevant: understanding where AI commonly fails, knowing what can safely be delegated, and developing the judgement to evaluate outputs.
Courses still matter because they provide the foundational mental model and shared vocabulary needed before employees begin experimenting independently. But training should be viewed as the starting point, not the finish line.
Everything after that has to happen on the job. Companies should set aside dedicated time each month for experimentation, similar to how technology companies protect time for research and development. Without a protected slot, it does not happen, because it loses to whatever is urgent.
It does not need to be complicated. A protected hour, a shared repository of lessons learned, and regular conversations about what worked and what did not can create lasting capability.
The goal is to build systems around people, not just around tools. Tools will continue changing, but the ability to work effectively with AI will remain valuable.
Editor’s note: This Q&A has been lightly edited for clarity and TNGlobal house style. The responses remain those of the interviewee.
Share your perspective: TNGlobal welcomes contributed insights and expert commentary from across Asia’s technology and innovation ecosystem. Submit a contribution for editorial consideration, or explore more conversations in our interview archive.
Workday’s Jess O’Reilly on why AI productivity gains get lost between disconnected systems [Q&A]