The AI Skills Gap That No Training Programme Is Closing
Created on 2026-03-01 22:26
Published on 2026-03-01 22:28
Two people sat down at 9am on a Tuesday. Same company. Same AI tool. Same job title.
The first typed a request: “Create a summary of our Q3 performance for the leadership meeting.” The AI produced something roughly 70% right. Wrong formatting. Missing the regional breakdown. Didn’t match the company’s slide conventions. She spent 45 minutes cleaning it up. She was satisfied. Without AI, this would have taken two hours. She saved time. This is what most organisations call “good AI usage.”
The second spent 12 minutes before opening the AI tool. She wrote down the purpose of the summary, who would read it, what decisions it needed to support, which data sources to reference, the required format, the tone, what should be emphasised, what should be excluded, and three sentences describing what a successful output looked like. Then she gave all of that to the same AI tool. She went to get coffee. She came back to a completed summary that hit every requirement. She did the same for three other deliverables before lunch.
Same tool. Same Tuesday. One did a morning’s work. The other did a week’s.
I have watched this exact pattern play out dozens of times now across training rooms in Singapore, Kuala Lumpur, Bangkok, and Jakarta. The gap between the best AI users and the average ones is not 20%. It is not 50%. It is closer to 10x. And the difference is not intelligence, not technical ability, not how many hours someone has spent with the tools.
The difference is a skill that almost nobody has named, almost nobody teaches, and almost nobody knows they’re missing.
The 25% Problem
The AI skills training industry has a structural problem. It teaches the 2024 skill in a 2026 world.
Prompt engineering was the right starting point. It was individual, synchronous, session-based. You typed a question, read the answer, refined the question. The skill was about the instruction: how clearly could you tell the AI what to do.
That skill still matters. But it is now table stakes. Teaching prompt engineering and calling it AI skills training is like teaching someone to type with all ten fingers and calling it computer literacy.
AI systems no longer just answer questions. They execute work. They run autonomously for hours, sometimes days. They produce complete deliverables. The human is no longer sitting there watching and correcting in real time. The human sets the work up and walks away.
This changes everything about what the skill actually is.
When you sit in a chat window iterating in real time, you are the safety net. You catch mistakes. You provide missing context. You course-correct. When the AI runs independently, everything you relied on in that conversation, your judgment, your organisational knowledge, your understanding of what “good” looks like, must be encoded before the AI begins. Not during. Before.
That is not a harder version of prompt engineering. It is a different discipline. And in the field work I’ve done across APAC enterprises over the past 18 months, I’ve found that what we call “prompting” is actually hiding four distinct skills. Most employees are practising only the first. The gap between people who operate across all four and people who don’t is the 10x difference I keep seeing.
I call these the 4Cs of Context Engineering.
C1: Communicate. Design the Instruction.
This is the layer everyone knows. How you structure a clear, unambiguous instruction for AI. Specify the task. Provide examples. Set guardrails. Define the output format. Resolve ambiguity before the AI has to guess.
Most AI training stops here. And the training is often poor even at this level. I’ve sat through vendor-led workshops where a full day was spent teaching people to add “think step by step” to their prompts as though those four words were a universal fix. They are not a methodology. They are a sometimes-useful trick.
The real skill at this layer is thinking like a delegator, not a conversationalist. If you gave your instruction to a capable person who had just started at your company that morning, could they produce what you need without asking a single follow-up question? If the answer is no, your instruction is incomplete. AI is always the person who started this morning.
This matters. But it is 25% of the skill.
C2: Contextualise. Build the Information Environment.
This is the layer where most people have never thought to look.
Your typed instruction might be 200 words. The context window it sits inside might be 100,000 words. Your instruction is less than 1% of what the AI processes. The other 99% is context: conversation history, system instructions, uploaded documents, the AI’s training data, whatever defaults your organisation has configured.
Most people manage none of this deliberately. They are asking AI to do work in an information vacuum and then wondering why the output feels generic.
Here’s the analogy I use in workshops: imagine giving the same task to two employees. One has worked at your company for five years. She knows the culture, the clients, the standards, the politics, the terminology, the history. The other started this morning. Both receive the same written instruction. Who produces better work?
The answer is context. The five-year employee has it. The new hire doesn’t. When you use AI without providing context, you are always working with the new hire.
Amazon learned this the expensive way when Alexa launched in Singapore. The voice assistant had been trained and tested extensively in American English. It worked brilliantly in the US market. In Singapore, it couldn’t handle Singlish, the natural creole that blends English, Malay, Mandarin, Hokkien, and Tamil in ways that shift mid-sentence depending on who you’re talking to and what you’re talking about. Alexa couldn’t parse the code-switching. It couldn’t handle the local references. It failed in ways that no testing environment in Seattle would ever catch, because the context of how Singaporeans actually speak was never in the system. Amazon had to invest significantly in localisation after launch, playing catch-up in a market where first impressions matter and where consumers quietly moved on to alternatives that understood them.
The technology was fine. The context was wrong. And the cost of deploying without the right context, what I call the Context Tax, was months of wasted momentum and consumer trust that never should have been lost.The technology was fine. The context was wrong. And the cost of deploying without the right context, what I call the Context Tax, was months of wasted investment and customer frustration that never should have happened.
The people who are 10x more effective with AI are not writing 10x better instructions. They have built 10x better context. Their AI starts every interaction with the right information already loaded. Their instructions can be simple because the context does the heavy lifting.
C3: Commission. Align AI With Intent.
This is the layer the industry is just beginning to discover. And it is where the most expensive failures come from.
Context engineering tells AI what to know. Commission tells AI what to want. The distinction is the difference between telling AI “resolve customer complaints quickly” and telling it “resolve customer complaints in a way that builds long-term trust, even if it takes longer.”
Those instructions produce radically different behaviour. Klarna discovered this when their AI customer service agent handled 2.3 million conversations, cut resolution times from 11 minutes to two, saved $60 million, and destroyed customer trust so badly that their CEO had to publicly admit the whole approach had produced “lower quality.” They started rehiring the human agents they’d let go.
The AI had clear instructions. It had context. It did not have intent. Nobody had encoded the judgment that a long-tenured customer expressing frustration needs generosity, not efficiency. Nobody had defined when relationship preservation matters more than ticket closure speed.
When you commission work from a human contractor, you don’t just describe the task. You describe the goal, the trade-offs you’d prefer, what “good” looks like, what should be escalated versus decided, what the work is ultimately in service of. You describe your intent.
AI needs the same thing. And it needs it explicitly. A human contractor might absorb your intent through conversation and relationship over time. AI absorbs nothing through osmosis.
In practical terms, commissioning means defining four things before AI begins:
What must the output include, no matter what? What must it never include or do? When multiple valid approaches exist, which does your organisation prefer? And under what conditions should the AI stop and ask rather than decide on its own?
Most organisations have never had to make these trade-offs explicit. They didn’t need to. Experienced employees carried them intuitively. But AI is not an experienced employee. And if you skip the commissioning step, you get what Klarna got: a technically brilliant agent optimising for the wrong objective.
C4: Codify. Define the Work So Completely That the Output Needs No Fixing.
This is the layer that separates people who use AI from people who direct it.
Go back to the Tuesday morning story. What Person B spent 12 minutes writing was not a prompt. It was a specification. A document that describes what an output should be with enough completeness, structure, and internal consistency that a capable system, AI or human, can execute it without asking follow-up questions.
The discipline of codification includes self-contained problem statements where all necessary information is present, acceptance criteria that let an independent observer verify the output, constraint architecture embedded from the commissioning step, decomposition of large tasks into independently executable components, and evaluation design so you can test whether the output actually meets the standard.
This sounds like a lot of work. It is more work upfront. And that is exactly the point. The best AI users take longer to set up a task. They finish 10x faster because the output is right the first time.
As AI becomes capable of longer, more complex autonomous work, this skill becomes the single highest-leverage human capability in any organisation. The person who can define an outcome with enough precision that an autonomous system can execute it is the new centre of gravity. Title irrelevant. Function irrelevant. The skill is universal.
Why This Matters Beyond the Individual
MIT’s Project NANDA research found that 95% of organisations are getting zero return from AI investments. The finding that should reframe your AI training budget: “This divide does not seem to be driven by model quality or regulation, but seems to be determined by approach.”
Approach. Not which tool you bought. Not how much you spent. How your people work with the tools you gave them.
The 95% are not failing because they lack access to capable AI. They are failing because their people are operating at C1. Writing instructions and hoping. Getting back 70% outputs that require 40 minutes of cleanup. Concluding that AI “isn’t ready yet” and going back to doing things the old way.
The 5% have people operating across all four layers. Their instructions are clear. Their context is rich. Their intent is explicit. Their specifications are complete. They are not using AI differently. They are using it at a fundamentally higher level of the same skill.
Now multiply that individual gap by an entire organisation. If one person with four skills is 10x more effective than one person with one skill, what happens when 500 people across your company move from one to four? That is not a 10x gain. It is a structural competitive advantage that compounds every day.
This is particularly urgent across APAC, where organisations are making significant AI investments, deploying tools broadly, and measuring adoption through logins and usage dashboards. Those dashboards show activity. They do not show whether the activity is producing value or producing 70% outputs that require manual cleanup.
BCG found that 74% of organisations fail to achieve or scale AI value, and 70% of the challenges are people and process. Not technology. People and process.
The process challenge is well-understood. The people challenge is almost entirely about skills. Not about whether your people can use AI. About whether they can use it at a level where the output actually serves organisational purpose.
What Changes Monday Morning
If any of this resonates, three things you can do this week that cost nothing.
First, test the gap yourself. Take a task you regularly give to AI. Write the instruction exactly as you normally would. Run it. Save the output. Then spend 10 minutes adding context about your role, your organisation, your audience, and your quality standards. Add your intent: what is the real goal, not just the task? Define what “done” looks like in three sentences. Run the revised version. Compare the outputs.
If the difference is significant, you have just experienced the gap between C1 and the full stack. Now imagine that difference multiplied across every knowledge worker in your organisation, every day.
Second, ask what your AI training actually covers. Most vendor-led training teaches button-clicking and basic prompt structure. That is C1, and often poor C1 at that. If your programme doesn’t address how employees build context, align AI with organisational goals, and write complete specifications, you are investing in the 25% and ignoring the 75%.
Third, look at where AI output is being manually corrected. Wherever your people are spending significant time cleaning up, reformatting, or redoing AI outputs, you are looking at a skills gap. Not a technology gap. The AI is capable. The instruction, context, intent, or specification was incomplete.
The technology is ready. The question is whether your people have been given the full skill to make it productive, not just the first quarter of it.
If you want to assess where your organisation stands across all six dimensions of AI readiness, including skills and capability, the AIR APAC Readiness Scorecard takes 15 minutes. It won’t recommend a training vendor. It will tell you whether your current approach is building capability or just building familiarity.