About 86% of Singapore workers say they use AI at work. As businesses, only 23.4% of SMEs have adopted it, against 70.4% of larger firms, according to IMDA’s Singapore Digital Economy Report 2026, released on 5 October. The figures come from different surveys and years, so they do not measure the same firms. The pattern is still familiar. Someone uses AI to write a reply, then has to find the customer record, check the status and copy the answer into another system. That remaining work is why admin keeps piling up.

To work out where AI fits, follow one recurring job from start to finish. Look for a repeatable step with reliable information and a result someone can check.

ASME president Ang Yuit said SMEs struggle to find the time, budget and manpower to plan AI automation, as reported by the Straits Times.

You can start that planning with work your team already knows.

Find the work between the tools

Ask whoever handles the admin: “What do you still copy, check or chase every week, even when you use AI?”

Watch them complete a recent job. Write down what starts it, where the information comes from, what they do and what counts as finished. Include waiting for approval and correcting mistakes. A polished AI draft may sit in the middle of a largely manual process.

For each candidate, record how often it happens and the time spent, including checking. Pick a job that recurs often enough to matter and has a clear finish. “Keep the business organised” is too broad. “Prepare follow-up drafts for unanswered quotes” gives you something to assess.

Check whether it suits automation

Before choosing a tool, ask:

  • Are the steps predictable? Someone should be able to explain the usual path and when it needs an exception.
  • Is the information dependable? Find the current record, rate card or policy. Conflicting versions need sorting first.
  • Can someone check the result? The reviewer needs evidence, such as the source message and record, alongside the draft.
  • What happens if it is wrong? Start where an error can be caught before a customer commitment or a change to records.

If the job follows fixed rules, check whether your existing software already handles it. AI may suit interpreting varied messages or drafting text. A reminder based on a date may only need ordinary automation.

If every case needs negotiation, leave that decision with a person. You might still automate gathering the background information.

An illustrative example

Imagine a small maintenance business following up unanswered quotes.

A staff member checks the quote register, searches the inbox for a reply, then asks ChatGPT to draft a follow-up. They still have to match the customer, confirm the quote is current and remember to send it.

A connected workflow could gather eligible quotes and recent replies, then prepare follow-up drafts with links to both. The staff member reviews each draft before sending. Rejected quotes and customers already in discussion go back to the person handling them.

That defines a useful job with a clear boundary. It also exposes a dependency: if nobody updates the quote register, automate that process only after fixing the records.

Test before connecting it

Use anonymised past examples, including an ordinary case, missing information and an unusual request. Write down the expected result before trying a tool. Keep customer details out of tools your business has not approved.

Compare the output with what actually happened. Count checking and correction time as part of the work. If reviewing takes as much effort as doing the job, narrow the task or leave it manual.

A wrong customer match, stale record or confident guess can turn a routine follow-up into an awkward conversation. Keep sending and record changes behind human approval during a trial, with someone responsible for exceptions. Our guide to limits for AI using business tools explains how to enforce those boundaries.

Your next step this week

Choose one recurring job and write its start, information sources, steps, finish and exceptions on a page. Walk through past cases with the person who does it. You can make that assessment without hiring anyone or buying another subscription.

Want a second view on a recurring admin job? Tell us what still happens by hand, and we can work through where automation would help.

Sources: IMDA, Singapore Digital Economy Report 2026, Straits Times reporting of ASME’s comments. The method is practical guidance, not a survey finding.