AI Training Didn't Work? A 5-Symptom Diagnosis for Malaysia

The certificates went out. Someone posted the group photo. Six months later the finance team still closes the month exactly the way it always did, and the only trace of the AI training is a line in last year's spend.

If that is your company, you are not an unlucky outlier. The MIT Project NANDA report on generative AI returns found 95% of organisations getting zero return — preliminary findings, drawn from a review of more than 300 publicly disclosed AI initiatives, interviews at 52 organisations and a survey of 153 senior leaders. Malaysia has been buying training at scale into that same pattern.

More than a million Malaysians completed AI Untuk Rakyat in under six months. Employers fund training on top of that at real scale: HRD Corp collected a record RM2.3 billion in levy during 2024. That figure is levy collected across all training, not AI training specifically, but it sizes what Malaysian employers already spend on upskilling. And MIT's Project NANDA, which reviewed over 300 publicly disclosed AI initiatives, interviewed representatives from 52 organisations and surveyed 153 senior leaders, found 95% of organisations getting zero return on generative AI, against USD 30–40 billion of enterprise investment. The report presents itself as preliminary findings, and the phrase it uses for the majority is no measurable impact on profit and loss.

So the awareness worked. The work did not move.

You have already spent the money, so the useful question is not what you should have planned. It is which of five things went wrong, because each one has a different fix and the most expensive mistake now is to treat all five as a training problem and buy another course.

First, the reason a second course rarely helps

The instinct after a failed programme is to blame the content and go looking for a better provider. It is usually the wrong read.

Baldwin and Ford's transfer-of-training model, widely cited since 1988, treats the work environment as one of three co-equal inputs to whether training reaches the job, alongside the training design and the trainee. They are careful to note the empirical evidence on the work-environment leg was thin, and flag it as a gap needing study. We still find it the most useful map of where transfer breaks. People come back to the same templates and the same approval steps they left, routed through the same two people who always sign off. They may know exactly how to use the tool. The official process gives them nowhere to use it.

There is a second reason, and it is arithmetic. The Center for Creative Leadership's 70–20–10 guideline puts roughly 70% of development in challenging experiences and assignments, 20% in developmental relationships and 10% in coursework and training. CCL is clear it is a guideline and not a formula. Even so, a claimable workshop was always the 10. A budget that funds the workshop and nothing else has bought a tenth of a capability.

None of that means the course was bad. It means a course was never going to be sufficient on its own, and buying a better one does not change the ratio.

The diagnosis

Five symptoms. They look similar from the outside and need opposite responses.

Two terms recur below, and they are easier to see in an example than in a definition. Say the process is supplier invoices: four hundred a month, keyed in by two people in accounts.

The operating measure is the number that process is judged on day to day. Here it is invoices cleared per person per day. The guard is the second number that stops the first being gamed — invoices posted to the wrong account. Speed on its own can always be bought by being sloppier, so the guard is what makes the speed mean anything.

Then the dates. A month in, count how many of the four hundred are genuinely going the new way. Read the operating measure at three months, not before. And if adoption is above 80% at three months while that measure has not moved, stop.

Those three timings are ours. They come from our own delivery work rather than from any study, so stretch them if your process runs on a longer cycle. What matters is not our numbers. It is that yours exist, written down, before the programme starts.

1. People use it privately, but nothing official changed

Your team demonstrably can use the tool. Someone drafts in it, pastes the result into the approved template, and files the work through the same route as always. Nothing in the documented process mentions it.

This is capability without redesign. It is one plausible route to the outcome MIT measured, rather than something MIT itself diagnosed: the organisation bought the training and skipped the part where the documented process is rewritten around the new tool. MIT does describe a shadow AI economy of staff using consumer tools without employer approval. Read each instance as a signal about which part of a job feels mechanical to the person doing it.

More training will not touch this one. Pick a single process, write down the new sequence of steps including which steps disappear, name who owns it, and set the date the old path closes.

2. A new process exists and nobody trusts it

The opposite failure. The workflow was redesigned, announced, and within a quarter people are quietly doing it the old way with the new tool bolted on top.

This is redesign without capability. The process assumes a fluency the team does not have yet, so the first awkward result sends everyone back to what they know. This one is a training problem, but a specific one: train the people who work that process, on that process, in the same two weeks the redesign takes effect. Generic AI literacy delivered three months early gives people nothing to practise on in the meantime, which is Baldwin and Ford's point about the work environment restated as a scheduling decision.

3. Adoption stalled below half, and you are told it needs more training

This is the one worth measuring rather than debating, because it separates symptoms 1 and 2 cleanly.

Go back to the four hundred invoices and count. If fewer than two hundred are going the new way a month in, and the two people in accounts can demonstrably use the tool when you watch them do it, then the fault sits in the redesign or in the closing date. More training will not repair either. Buying more is how a stalled programme becomes an expensive one.

Note also that the first month is supposed to look bad. Brynjolfsson, Rock and Syverson describe the productivity J-curve: organisational investment alongside a general purpose technology depresses measured output before lifting it. Judging the operating measure at week four can make a working programme look like a failure.

4. The dashboard improves and the business does not

Someone reports that the AI is now handling most of the incoming WhatsApp enquiries. Renewals are flat. Both are true.

Gartner surveyed 5,728 customers and found that only 14% of customer service issues are fully resolved in self-service. Handled and resolved are different events, and only one of them is what the customer wanted. AI is unusually good at moving the first number while the second stands still.

Motivation is not the lever here. Change what the measure counts: tie it to something the business already tracks, and attach the guard. Quotes go out faster and the pricing errors go up with them, and you have moved the problem rather than solved it.

5. Nobody can say whether it worked

No baseline was taken, so the argument about whether the programme delivered has no evidence on either side and has been running for two quarters.

Kirkpatrick's four levels evaluate training at reaction, learning, behavior and results — the model's own American spelling. Most programmes measure level one, attendance and how people felt, or level two, usually a quiz. Levels three and four ask whether behaviour and business results changed, and both need a defined process and a baseline you can no longer collect retrospectively.

You cannot recover the original numbers. What you can do is stop spending management attention on an unwinnable argument, baseline the process as it stands today, and treat that as the starting line for the next change rather than the verdict on the last one.

Why this keeps happening in Malaysia specifically

Two structural reasons, and neither is anybody's incompetence.

The funding pays for the part that was never the constraint. The levy funds training. Budget 2026 adds a further 50% tax deduction for smaller companies on AI training recognised by MyMahir, once in two years, for applications up to the end of 2027. Nothing in either mechanism asks whether the workflow changed. A company can therefore do everything the system rewards and still see nothing move, which is a design feature of the incentive rather than a failure of the company.

And the delivery market is supply-driven. The Edge has reported that Malaysian upskilling programmes remain supply-driven, designed largely by training providers with minimal input from employers. The provider is paid for attendance. Nobody in that transaction is paid for behaviour change, so the buyer has to supply that discipline.

Jazz Tong develops this at length in How Malaysian Businesses Can Benefit From AI. He splits BCG's "70% people and process" into two lines a business can buy separately, 30% capability and 40% workflow redesign, then argues that no Malaysian levy scheme or tax deduction pays for the second one.

He is explicit that the 30/40 weighting is his judgement and not a measured figure. His full split is 10% model and tool, 20% data and plumbing, 30% capability, 40% redesign. (Worth flagging, since the digits collide: BCG's 10/20/70, this 10-20-30-40, and the Center for Creative Leadership's 70-20-10 further up are three unrelated frameworks that happen to share numbers.)

That essay is the fuller case. This piece is what to do when you are already on the wrong side of it.

What to do next

Match the symptom to the fix, and resist the one purchase that fits none of them.

• Private use, no official change → one process, a written sequence, a named owner, and the date the old path closes. No further training until those exist on paper.
• New process, no trust → train that team, on that process, in the same two weeks the change takes effect.
• Adoption under half at a month → fix the redesign or the closing date, not the training budget.
• Dashboard up, business flat → re-tie the measure to something already counted, and add the guard.
• No baseline → baseline today, and stop litigating the last programme.

Then set the rule that most programmes skip: decide in advance what would make you stop. If adoption is above 80% at three months and the operating measure has not moved, the process was the wrong candidate. The honest response is to pick a different process, not a better model.

What a good first candidate looks like

MIT's finding on where value actually appears is counter-intuitive and worth acting on. Half of generative AI budgets go to sales and marketing, but the report finds back-office automation often yields the better return, and that the wins come from narrow, heavily customised use cases. Front-office tools get the attention; back-office tools deliver the savings. Boards want the chatbot. The money is often in reconciliation and document handling, the work nobody demos.

So look for a process that is:

• High volume — it happens often enough that a week of improvement is visible.
• Low variance — the same shape every time, not a judgement call wearing a template.
• Internally facing, at least at first, so an early mistake costs you an apology and not a customer.
• Tolerant of a review step while trust is being built.
• Already measured by somebody, because a process with no existing number cannot show you a change.

A factory floor scores well on the first four, which we work through in predictive maintenance; the cost side of building and running one is worked out separately.

If the honest answer is that the last programme taught people a tool with nowhere to use it, the sequence to fix that starts with a planning conversation, not a training booking. Rebuilding the process is digital transformation work, not training work, and the two are bought differently. We would rather scope the process with you first than sell you the tenth that was never the problem — and if the diagnosis genuinely does come back as a capability gap, our HRD Corp claimable training and the Anchor Academy course structure are there for it.