You will be able to explain why some tasks should not be automated even when they are repetitive.
A parent in Tampines writes to her son's tuition centre. He is unhappy in his Secondary 3 maths class and she is thinking of taking him out. The message is courteous and slightly strained, and it closes by asking about refunds. She and the other people below are invented examples.
Mei Ling at the centre rings her that same afternoon, and the boy stays enrolled. Suppose the centre had sent a canned answer such as "thanks for your enquiry, here is our timetable". The parent would have felt nobody had listened, and the family would most likely have left. No tool available today would have chosen to make Mei Ling's call, and you would not want one to.
Plenty of jobs score as repetitive on the sheet you filled in but still belong with a person. Sometimes that person does the whole job. Sometimes software does the groundwork and a person checks it before anything goes out to a client, supplier or customer. Four signs tell you which tasks fall into this group, and they lead to three labels you can put on any task.
A message can arrive every week through the same form, in the same layout, and a person still has to handle it. The test is whether a good answer changes with who sent it, the history between you and their mood. A long-standing client complaining is one case. Others are a colleague raising a sensitive question about pay, or a supplier running late for the third time. The written words give you only some of what you need. The rest is in the sender's head, and only someone who knows them can work it out.
Software still has a supporting role. It can record each message, mark it as sensitive and move it to the top of the queue. An AI step can also sort incoming messages and prepare a draft reply that a person then edits. Module 5 covers how to do this. The person who knows the relationship decides what actually gets sent.
In lesson 1.1 you priced each task by how often it happens, how long each run takes and what an error costs. For something you do twice a year, that sum seldom favours building an automation. Count the upkeep during the idle months as well as the time to set it up.
Arif handles admin at a small renovation firm in Ubi. Once a year he gathers staff details for the company's insurance renewal, and he reckons it takes about three hours. To automate it he would need a form and a sheet, the fields mapped between them, and testing, which could easily take a day. The automation would then sit idle for twelve months. During that time the form tool gets updated, a connection expires and the insurer changes its spreadsheet. Arif would then spend an hour repairing it before it had run even once.
A checklist in a shared document, with each step written out and last year's file attached, suits Arif better. It is usually the right tool for any task this rare.
Some errors cost almost nothing. If an email gets the wrong label, you fix it in two seconds. Others are much harder to take back: money paid into the wrong account, a file removed from a shared drive, a legal notice served on a tenant, or a message sent to your entire customer list.
Tasks like these need a person involved even when software does most of the work. The software gets everything ready, such as entering the payment details, writing the notice or building the recipient list. Then it stops until someone approves, and only after that does the step that cannot be reversed go ahead. The approval wait takes seconds. A payment sent to the wrong place can mean days of phone calls, and you might never get the money back.
This pattern sits between full automation and doing the job by hand, and it will come up often as you go through your list. It only protects you if the approver reads what is in front of them, so the step has to be designed so people don't just click past it. Lesson 7.3, "Logs and human review checkpoints", explains how to set that up.
To apply the write-the-rule test, put the rule the automation would follow into plain sentences and check for any "it depends" left in them. "When a form arrives, add it to the sheet and send the confirmation" has none, so it passes. "Reply to enquiries in the right tone" fails, because no one can define the right tone for every case. "Approve expense claims that look reasonable" also fails, because it doesn't say which amounts, categories or receipts count as reasonable.
If a vague part remains, try to replace it with something definite, such as a limit, a required attachment or a fixed set of categories. The expense rule could become: approve claims below an amount the finance team sets, with a receipt attached, in four named categories. If you can do that, the task may be worth automating. If you can't, don't automate it. When a person cannot write a rule down, the software cannot follow it either. It will guess, and you will have no way of telling when a guess is wrong.
Each task gets exactly one of three labels:
Automate when the inputs are stable, the steps are fixed, there is a clear finish, and an error is easy to undo. Automate with a human check when the preparation is routine but the final action is expensive to get wrong or affects a relationship. The software prepares and a person approves. Keep manual when the task is rare, depends on judgement, or cannot be written as a rule.
People use the middle label less than they should. A task dismissed as "too important to automate" is often nine routine steps and one decision, and only the decision needs you. Before you keep a whole task for yourself, look at its steps one at a time.
For every task, choose one label and write one reason for it, naming whichever of the four signs applies. Go back to the ten tasks you listed in lesson 1.1. The activity below asks you to give each one a label and a single reason, which will show you where the real automation work in your week sits.
Go back to your ten tasks and mark each as automate, automate with a human check, or keep manual, with one reason for each.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).