AI-bilingual: Singapore's 100,000-worker target, in a team of twelve

The short version: the National AI Impact Programme, announced at the 2026 Committee of Supply debates, aims to support 10,000 enterprises and train 100,000 workers in AI skills by 2029, with sector-specific tracks being developed for professions such as legal, accounting and human resources through the relevant professional bodies. The concept is fluency, not expertise: your accountant does not become a data scientist, they become an accountant who knows what to hand over and what to check.

Policy details on this page were checked on 26 September 2026 against the announcing agency's own materials. Schemes, caps and dates change — confirm the current position with the administering agency, and with your own tax or grant adviser, before relying on anything here.

Why "bilingual" is the right word

The common failure in company AI training is aiming at the wrong level. Teach people prompt tricks and it decays in a month. Teach them the technology and they retain nothing they can use on Tuesday. Fluency sits between: the person keeps their domain judgement and adds a working sense of what the tool is good at, where it fails, and which parts of their job it should not touch.

In practice the fluent version of someone's job looks like this. They know which of their recurring tasks is worth handing over. They can tell a plausible-sounding wrong answer from a right one in their own field — which is exactly the check a model cannot perform on itself. And they know which decisions stay with a human regardless of how good the draft is.

That third one is the part that protects you. A team that has internalised where the boundary sits does not need a policy document to stop an agent sending something it should not have.

The three roles to train first in a small company

If you have a dozen people and no training budget worth the name, the order is:

1. Whoever handles inbound. Enquiries, quotes, scheduling, the first reply. This is the highest-volume, most text-shaped work in most SMEs, and the person doing it can tell within a week whether a draft is good enough to send. They are also the natural owner of the agent that drafts.
2. Whoever writes the same thing repeatedly. Proposals, reports, listings, updates, case notes. Their fluency pays back fastest because the work has a template shape they already know.
3. One person who can look at what the agents did. Not a technical role — someone who reviews the week, notices when answers drift, and knows who to tell. Every deployment that is still running a year later has this person, whether or not anyone gave them the title.

Notice who is not on the list: the most technical person in the building. They will be fine. The value is in the people whose judgement the system needs and cannot replicate.

What the national programme covers, and what it does not

The sector-specific tracks under development with professional bodies in accounting, law and related fields are genuinely useful for regulated professions, where the boundary question is a compliance question. If your team sits in one of those professions, the professional body route is worth checking first.

What no national programme does is teach your people your processes. A course can teach an accountant what AI is reliable at; only you can decide which of your firm's five recurring jobs an agent should take on and where your approval line sits. That gap is the whole of the implementation, and it is why training and deployment work better together than in sequence.

We run that combined version — AI training for teams — because a team trained six months before anything is deployed has forgotten most of it, and a system deployed to a team that was not trained gets quietly routed around.

Three things worth teaching that are not on any syllabus

• How to tell the difference between a draft and a decision. The single highest-value habit. An agent produces drafts; a person makes decisions. Teams that blur this either over-trust or refuse to use the thing at all.
• What to do when the output is subtly wrong. Not "flag it" — actually correct the source. An agent that keeps producing the same wrong answer usually has a wrong fact in its context, and whoever spotted it is best placed to fix it. See AI agents for Singapore SMEs for how that loop works in practice.
• Which data does not go in. Under the PDPA this is not an abstract question. A short, concrete list your team actually remembers beats a policy nobody reads — PDPA and AI agents covers the specifics.

How to tell whether the training worked

Not by a completion certificate. Three signals, all observable within a month:

1. People are handing over work unprompted, rather than because they were told to.
2. Someone has caught and corrected a wrong answer without escalating it.
3. Somebody has declined to use an agent for something, and can say why.

The third is the one that tells you the fluency is real. A team that uses AI for everything has not learned the boundary; it has just moved the risk.