Which Process Should an AI Agent Run First? A Picking Guide for Singapore SMEs

The first process you hand to an AI agent should pass four tests: you can describe it on one page, it runs at least twenty times a week, a wrong output is visible within a day, and undoing that wrong output costs under a few hundred dollars. For most Singapore companies of ten to twenty people, that points to quotation follow up, matching supplier invoices against delivery orders, or first line enquiry triage. Payroll, statutory filing and anything that moves money out of your account are not first candidates. In the quarter before the 1 January 2027 cost changes, payroll in particular is the worst place in your business to run an experiment.

What makes a process safe to hand over first?

Four properties, and all four have to be present. Three out of four is how pilots quietly die.

• Written down. Not in someone's head. If the only copy of the process is in your operations manager's memory, you are not automating a process, you are automating a guess.
• Repetitive. Twenty times a week is roughly the floor where the setup effort pays back inside a quarter. Below that, the person doing it manually is cheaper than the person maintaining the agent.
• Reversible. A badly worded follow up email is embarrassing. A wrongly released payment is a bank recall, a supplier conversation and a week of your time.
• Observable. Someone must be able to look at yesterday's output and say whether it was right. If nobody can tell for a month, you will discover the failure through a client complaint.

Add a fifth, non technical requirement: one named person who owns the output. Not the vendor, not the committee. A person whose name is on it.

Which processes do Singapore SMEs usually start with?

These four come up repeatedly because they sit in the sweet spot of high volume and low judgement:

• Quotation follow up. Every quote older than five days with no reply gets a drafted follow up, with the specifics of that job pulled in. Your salesperson approves or edits. Most SMEs lose more revenue to silence than to losing on price.
• Enquiry triage and routing. Incoming web and email enquiries get read, classified, tagged with urgency and routed to the right person with a suggested first reply. This typically removes the worst delay in the whole sales cycle, which is the enquiry that sat unread over a weekend.
• Supplier invoice matching. The agent compares each invoice against the purchase order and delivery order, flags the differences, and prepares the exceptions for a human. It does not approve and it does not pay.
• Monthly report assembly. Pulling the same numbers from the same three places into the same format, with the agent noting anything that moved more than ten per cent. The judgement stays with you, the copying and pasting does not.

Notice what these have in common. In every one, the agent prepares and a human releases.

Which processes should wait?

Leave these alone for now. Payroll and CPF configuration, because the rates and thresholds change on 1 January 2027 and the correct values have to be entered and checked by a human who understands them. Statutory filings, because the deadline is immovable and the penalty lands on the company. Anything that sends money. Anything that commits you contractually. Anything that moves personal data to a service you have not assessed. Pricing decisions, because an agent will cheerfully be consistent about a margin you did not intend to offer.

None of these are permanently off limits. They are simply the wrong place to learn.

How do you tell if the process is actually written down?

Run this test before you spend a dollar. Ask the two people who do the work to write the steps down separately, without conferring, and compare the two documents.

If they differ in any meaningful way, you have found your real problem, and it is not an AI problem. Two people have been doing the job two different ways, and the output has been inconsistent for as long as that has been true. An agent will not resolve that ambiguity. It will industrialise whichever version you happened to describe, at a volume no human was producing before.

Fix the process on paper first. That alone usually pays for itself, and it makes the automation step roughly twice as fast.

What does the first 30 days look like?

Four weeks, with the leash shortened only when the numbers earn it:

1. Week one, shadow mode. The agent produces its output and nobody acts on it. You compare its work against what your staff actually did. Nothing it produces reaches a customer.
2. Week two, approve everything. Every single action waits for a human click. Slow on purpose. This is where you find the failure patterns.
3. Week three, approve by exception. Low value and low risk items go through automatically. Anything above a threshold you set still waits.
4. Week four, measure and decide.

Capture four numbers from day one, because you cannot reconstruct them later: minutes per item before and after, the error rate, the share of outputs a human edited before releasing, and the running cost per item. The edit rate is the one that tells the truth. If your staff are rewriting more than one output in five after 30 days, the process was not ready, and the honest move is to stop, fix the written process and start again.

What should the first one cost?

Set a hard monthly ceiling before anything goes live, so a loop or a bad prompt cannot run up a bill overnight. For a single process at SME volumes, model and service costs usually land in the low hundreds of dollars a month, with the real money going into setup and the process clean up around it. Government support exists for certain pre approved digital solutions, so check the current listings before you self fund, but do not let a grant decide which process you pick. Grant fit is a payment question. Process fit is a results question.

Judge the result in money or hours, not in enthusiasm. One process, one owner, one number that moved. Then do the next one.

Frequently asked questions

Do we need a developer in house to run an AI agent?
No, but you need someone who will read the audit log every week and who has the authority to switch it off. That can be your operations manager. What you cannot do is leave it unowned because the technology appeared to work on day one.

Is our customer data safe if an agent reads our email?
It depends entirely on which service you use and what you agreed to. Before you connect anything to your mailbox, confirm in writing where the data is processed, how long it is retained and whether it is used for training. Under the PDPA the obligation to your customers stays with you, not with your vendor.

Should we wait for Budget 2027 before committing?
No. Process clean up and a single low risk pilot cost little and are useful whatever February brings. Hold back the larger spend, the multi process rollout and anything that depends on a specific grant until the announcements are out.