At five in the morning, Rina already has her laptop open in the corner of her coffee shop. Before the first beans hit the roaster, she pastes last week’s sales figures into a template and asks a language model to estimate what she needs to buy today. The whole routine takes ten minutes. Doing it by hand, she says, used to take nearly an hour with a calculator.

Rina is not an outlier. In the past three months, at least six micro-business owners interviewed for this story have started using generative AI for work that once felt too technical: forecasting stock, drafting shift schedules, writing product descriptions, and answering customer questions during the lunch rush.

Starting with the most boring task

The pattern is remarkably consistent. Nearly everyone began with a repetitive, low-risk job — not with anything grand like a full financial model. A leather-goods maker uses a model to draft product descriptions in three languages. A café uses one to summarise customer reviews every week.

“The rule is: never trust it straight away. I still check the numbers,” says Dimas, who runs a home-catering business and now uses a similar tool to plan his daily ingredient order. He admits he got it wrong twice, when the model invented sales figures that never existed.

These tools speed up the work, but they don’t replace what an owner knows about their own shop.

Saving time is not the same as saving money

Cost remains the sticking point. Free tiers limit the number of requests and the quality of output. Paid plans range from a few dollars to tens of dollars a month — a reasonable line item for a café turning over millions of rupiah a week, but a heavy one for a vendor working on thin margins.

Small-business researcher Ardi Kusuma argues the real impact is not the direct saving but the hours freed up. “If two hours a day can go back into serving customers or finding a better supplier, that is the actual value,” he says.

What to watch

  • Customer data. Avoid pasting phone numbers or addresses into public tools without a clear policy.
  • Verify the numbers. Language models sound confident even when they are wrong; cross-check against real sales records.
  • Dependence. Keep the templates and manual procedures so the business still runs when the service goes down.
  • Tone. Customers still want a human reply, especially when something has gone wrong.

For Rina the decision stays simple: as long as the tool removes the boring parts and she can still check the result, it has a place in her kitchen. “But the day I stop understanding the numbers on the report, I’ll stop using it,” she says.