Cost Per Task: Which Work Should a Singapore SME Actually Hand to AI This Quarter?

Which work should a Singapore SME hand to AI this quarter? The tasks that are high in volume, low in judgement, and checkable in under a minute — draft replies to repeat enquiries, first-pass quotation line items, meeting notes turned into action lists, product descriptions, messy data tidied into a consistent format. The test is not whether AI can do the task. It is whether your fully loaded cost per task — licence, usage fees, and the minutes a person spends checking the output — lands below what the task costs you today, without creating a data or compliance problem. Most SMEs skip that arithmetic, buy a tool on a monthly seat, and six months later cannot say whether it paid for itself.

What does cost per task actually mean?

Three numbers, and the third is the one everybody forgets.

• Licence, spread over real volume. A S$30 per month seat used for 400 tasks costs about 8 cents a task. The same seat used for 20 tasks costs S$1.50. Volume is what makes the tool cheap, not the price on the pricing page.
• Usage fees. Per-message, per-document or per-token charges on top of the seat. Small individually, and easy to lose track of when three people are using the same account.
• Human review time. If a staff member on a loaded cost of roughly S$25 an hour spends 90 seconds checking each output, that is 63 cents a task — usually the largest line of the three.

Add them up and compare against the task as it runs today. A supplier enquiry reply that takes 8 minutes to write from scratch costs about S$3.30. The same reply drafted by AI and checked in 90 seconds costs about S$0.70. That gap, multiplied by a few hundred a month, is a real saving. But reverse the review time — a task where checking the output takes as long as doing the work — and the saving is zero no matter how impressive the demo was.

Which tasks pass the test in a 10 to 30 person company?

Four patterns pass reliably:

• Repeat enquiries with a known answer. Lead times, delivery areas, minimum order quantities, warranty terms. The content already exists somewhere; AI is assembling it into a reply, not inventing it.
• First drafts of structured documents. Quotation line descriptions, tender response boilerplate, job ads, SOPs. A human finalises price, scope and commitment — the AI removes the blank page.
• Format conversion. Voice notes to written instructions, meeting recordings to action lists, a supplier PDF into spreadsheet rows. Errors here are obvious on sight, which keeps review fast.
• Volume tidying. Deduplicating a customer list, standardising product naming, translating listings into a second language. Tedious, high-volume, verifiable in bulk.

Which tasks fail, and why do they look so tempting?

They fail because review cost is high, or because being wrong is expensive:

• Final pricing and commercial commitments. Checking a quoted price means recalculating it, which means doing the task anyway.
• Anything with a regulator attached. Hiring decisions, grievance handling, employment terms. Under the Workplace Fairness Act regime, what you need is a documented, defensible human reason — not a screened shortlist nobody can explain.
• Work that lives in one person's head. If the knowledge was never written down, AI will produce a confident, plausible, wrong version of it. Fix the documentation gap first.
• Anything involving personal or customer data in a consumer-grade tool. NRIC numbers, staff records, customer databases and unreleased pricing do not belong in a free public chatbot. That is a PDPA exposure, and your appointed data protection officer owns the consequences.

How do you stop the review step eating the savings?

Set the checkpoint deliberately, per task, in writing:

• Draft only. AI produces, a human rewrites. Correct for anything commercial or sensitive.
• Review then send. A human reads and approves before it leaves the company. This is the default for customer-facing work.
• Spot check. Output goes out, a sample is audited weekly. Only earn your way here — after roughly 100 outputs at a 95 percent acceptance rate on a narrow, unchanged task.

Track one number while you pilot: the share of outputs that needed material correction. If it falls, tighten the checkpoint and the cost per task drops with it. If it stays flat after a few weeks of prompt and template adjustment, the task has failed the test — stop, and put the licence somewhere else.

What has to be in writing before the first task moves?

One page is enough, and it needs four things: which tools are approved; what data may never be entered into them; who reviews each category of output before it leaves the company; and where the output is stored so it does not sit only in a chat history. Add one line on exit — can you export your prompts, templates and history if you leave? Answer it before you commit, not at renewal.

What does a realistic Q4 sequence look like?

October: pick exactly two tasks that pass the test, and write down today's baseline — volume per month and minutes per task. Without a baseline you have no argument in December. November: run them and record review time and correction rate. December: decide. Keep what beats its baseline, kill what does not, and delete the accounts you killed before they auto-renew.

One caution on timing: stay monthly through the pilot. Annual plans get sold hardest in Q4, and committing twelve months to a tool you have used for three weeks converts a reversible experiment into a sunk cost you will defend rather than measure.

The SMEs getting value from AI this year are not the ones using the most of it. They are the ones who can name the three tasks they moved, the cost per task before and after, and the one they tried and dropped.

Frequently asked questions

How much should a 20-person SME budget for an AI pilot?
Two to four seats on a business plan, roughly S$60 to S$250 a month all in, plus a few hours of someone's time to write the prompts and templates. If a pilot needs a five-figure commitment before you have measured anything, it is not a pilot — it is a purchase.

Do we need a formal AI policy for a company this small?
Yes, but one page, not a manual. The exposure that matters is staff pasting customer or employee data into consumer tools, and that risk exists whether or not you have bought anything — your team is almost certainly already using these tools on personal accounts.

What is the single most common mistake SMEs make here?
Automating the task that annoys the owner most rather than the task with the highest volume. Annoyance is not the same as cost. Count how many times a month it happens before you decide it is worth automating.