Which Tasks Can You Safely Hand to AI? A Triage Framework for Singapore SMEs

A task is safe to hand to AI when it passes four tests: the output is reversible (a mistake costs minutes, not a client or a penalty), verifiable (a human can tell it is wrong in under a minute without opening another system), low-sensitivity (no NRIC, payroll, health or contractual data leaves your controlled environment), and unowned by a regulator (nobody outside your company will ask who signed it). Drafting, summarising, reformatting, first-pass classification and internal search pass all four. Anything that files, pays, hires, terminates, prices a contract or speaks to a customer without review fails at least one. Triage every task against those four tests and the automation roadmap writes itself.

Why does the tool-first approach keep failing?

The common pattern in a 10-to-30 person Singapore company is that someone buys a licence, the team is told to use it, and six months later two people use it daily, nobody can say what it replaced, and the finance lead has quietly discovered that staff have been pasting customer lists into a free consumer chatbot because the approved tool was slower.

That happens because the decision was framed as a tool question instead of a task question. Tools change every quarter. The tasks in your business barely change at all. If you triage tasks first, you can swap vendors in 2028 without redoing the thinking, and you can tell a staff member exactly where the line is instead of hoping they guess correctly.

What do the four tests actually mean in practice?

Reversibility. Ask what happens if the output is wrong and nobody notices for a week. A badly drafted internal meeting summary costs a correction. A wrongly computed CPF contribution costs a back payment, an amended submission and a conversation you do not want to have. Reversibility is not about how often the model is wrong; it is about what the wrong case costs you.

Verifiability. A task is verifiable if the person reviewing it can spot an error from the output alone. A quotation summary is verifiable: the reviewer knows the job. A reconciliation across 400 line items is not, because checking it properly means redoing it. Tasks that are hard to verify quietly convert into tasks nobody checks.

Data sensitivity. Under the PDPA you remain the data controller regardless of which service processed the data. If a task cannot be performed without NRIC numbers, salary figures, medical details, bank details or signed contract terms, it does not move to a consumer-grade tool at all. It either stays human or it moves to a tool with a business agreement, no training on your inputs, and an administrator who can revoke access.

Regulatory ownership. Some outputs have a named human attached by law or by contract. GST returns, CPF submissions, employment decisions, tender declarations and safety sign-offs all belong to a person. AI can prepare the working. It cannot be the signatory, and from 2027 the documentation trail behind hiring decisions in particular needs to show human reasoning, not a generated paragraph.

Which tasks should a Singapore SME hand over first?

These clear all four tests for most businesses and usually return time within the first fortnight:

• First-draft written material — supplier emails, service descriptions, SOP drafts, job ad copy for a human to edit and own.
• Summarising documents you already hold — long supplier contracts, tender specifications, meeting notes, where the reader knows the subject well enough to catch a wrong summary.
• Reformatting and extraction — turning a PDF delivery order into structured rows, normalising inconsistent product names, converting a messy spreadsheet into an import template.
• First-pass triage — sorting an inbox or WhatsApp queue by intent and urgency, with the routing visible and a human handling the exceptions.
• Internal question answering — a searchable layer over your own SOPs and policies so new staff stop asking the owner.
• Code and formula assistance — spreadsheet logic, report queries, small scripts, where output is tested before it matters.

Which tasks stay with a human?

Keep these on a named person, with AI used only to prepare the working:

• Anything that files or pays — GST and corporate tax submissions, CPF, payroll runs, supplier payment releases.
• Hiring and termination decisions, including the written reason behind a rejection. The decision and its record must be defensible as human reasoning.
• Pricing and contract terms on anything material. A generated quotation that undercuts your margin is binding the moment the client accepts it.
• Unreviewed customer communication on complaints, disputes, credit and anything involving a named individual's personal data.
• Final tender and compliance declarations, where a wrong statement is a disqualification rather than an error.

What about the middle category?

Most tasks land in the middle: valuable, but failing one test. The useful move is to fix the failing test rather than ban the task. A task that is sensitive can often be de-identified before it goes anywhere. A task that is unverifiable can be made verifiable by sampling, by having the system show its source, or by splitting it into a step a human checks and a step that does not matter. A task that is irreversible can be made reversible by inserting an approval before anything is sent, filed or paid.

That reframing is where most of the real gain sits, because the high-value tasks in a small business are almost never the clean ones.

How do you run this triage in one afternoon?

List every recurring task in one function — start with the function where the owner is the bottleneck. Against each task, mark the four tests as pass or fail and note the weekly hours. Everything with four passes goes into a do now list, ordered by hours saved. Everything failing only one test goes into a fix the test list with the specific control needed. Everything failing two or more stays human and gets written down as such, because an explicit no is what stops staff from improvising with a consumer chatbot.

Then put the result in writing. The triage list is the practical half of an AI use policy: not a statement of principles, but a named list of what is handed over, what is prepared-then-reviewed, and what never leaves a human's hands. Review it quarterly, because the tool capabilities move and the list should move with them.

Frequently asked questions

Can AI handle our bookkeeping if a human reviews everything?
Data entry, document extraction and transaction categorisation are reasonable to hand over, since errors surface at reconciliation. Submission and payment release stay human. The distinction is between preparing the numbers and being accountable for them.

Our staff already use AI for everything. Do we pull it back?
No. Audit what they are actually doing, approve what passes the four tests, and provide a sanctioned tool for the rest. Prohibition without a working alternative produces shadow usage you can no longer see, which is the worse position under the PDPA.

How much of a task should AI do before we call it automated?
Treat anything short of end-to-end as assistance, not automation, and account for review time honestly. A task where AI does 80% and a human spends 20 minutes checking is still a genuine win, but it has not been removed from anyone's week.

If you want this triage run across your operations with the controls specified task by task, that is a half-day exercise we do with Singapore SMEs regularly. The output is a list you can hand to your team on Monday.