Productivity is evidence. It is not yet a verdict.

Three in ten Singapore firms have adopted AI. About seven in ten of those report improvements in worker productivity.

That sounds like a verdict. It isn’t one yet.

On 9 September, Singapore’s Ministry of Trade and Industry published two answers that belong together. The first described early signs of progress from AI adoption: higher reported worker productivity, more redesigned roles and new AI-related jobs, rather than broad headcount reduction. It also referred to an MTI study in which firms using AI saw higher revenue and total employment, with stronger gains among firms that had deepened their capabilities.

The second answer was more cautious. Asked how productivity improvements from Government-supported AI investments are measured, MTI said there is not yet enough data to assess the overall impact. It also said AI’s contribution cannot be isolated cleanly from demand conditions, workforce skills, management practices and wider digital transformation.

That isn’t a contradiction. It is the operating problem.

A signal is not the same as a result

Tool adoption is activity. Time saved is an output. Reported productivity is evidence. None of them, on their own, tells you whether the business improved.

A team may complete a task faster while creating more review work downstream. A customer-service tool may shorten handling time while reducing resolution quality. A commercial team may produce more proposals while winning fewer of the right ones. Every headline measure can improve while value moves somewhere else in the system.

That does not make the early productivity figures unimportant. It makes the questions around them more important.

Leaders do not need perfect attribution before acting. They do need enough evidence to distinguish useful change from activity that merely became easier to count.

Four lines for every material AI initiative

For each material initiative, I would want four lines on one page.

1. The baseline. What happened before the workflow changed? Record the time, cost, quality, volume or commercial result that matters before the new tool becomes normal.

2. The business outcome. What should improve beyond usage or time saved? This might be margin, revenue quality, conversion, resolution, error reduction, cycle time or capacity released for higher-value work.

3. The guardrail. What must not get worse while the headline number improves? Faster output is not progress if accuracy falls, risk accumulates, customers repeat themselves or senior people inherit more checking.

4. The decision date. When will someone decide whether to scale, redesign or stop? Name the person with authority to make that call and the evidence they will use.

The fourth line matters most. A review without an owner authorised to act is reporting, not governance.

The standard is a defensible decision

AI will rarely be the only thing changing inside a business. People learn. Demand moves. Processes are redesigned. Management attention shifts. That is why waiting for a perfectly isolated causal answer can become its own form of avoidance.

The practical standard is simpler: can the organisation explain what changed, show why the result matters, identify what could have been damaged and make a dated decision from the evidence?

If it can, imperfect evidence can still support a responsible choice.

If it cannot, a dashboard full of time saved will not rescue the operating model.

Productivity is valuable evidence. It becomes a business result when it survives the questions around cost, quality and consequence.

Which measure would make you stop an AI initiative even if it saved time?