Score a task on how easy it is to check, what an error would cost, and how well it fits what models do well.
Some work tasks go well when you hand them to an AI assistant and some go badly. Usually you only find out which after the damage is done. A short scoring method lets you sort tasks before you start. It uses three questions: how quickly you can check the output, what an error costs, and whether the task fits what models do well. None of them needs technical knowledge.
Score each question from one to five. On every question, a higher score means the task is better suited to AI. A task that scores high on all three is a strong candidate to hand over.
Take Marcus, an account manager at a freight forwarding company in Changi (an invented example). On a typical Monday he writes a status update to each of his six clients and prepares a short volume report for his manager. He also answers a complaint about a delayed shipment and drafts notes for a quarterly review. After lesson 8.1 he knows AI will do some of these tasks well and fumble others, and that the difference will not be obvious. He needs a quick way to tell which is which beforehand.
Ask about checking first, because everything else depends on it. Module 6 showed that any AI output can contain errors that look exactly like correct content. You only catch those errors if you check. The checking score is how quickly you can verify the output compared with producing it yourself. The rule is that a task suits AI when you can check the result faster than you could produce it.
A draft email to a client you know well might score a four or five on checking. Reading it takes a minute, and you can see whether it says what you meant. A summary of a meeting you attended scores about the same, because you know what was said. An analysis you would have to redo to verify might score a one or two. Say Marcus asks an assistant which clients are shipping less this quarter. The only way to verify the answer is to go through the figures himself. The AI saves him little there and adds risk, so the task scores low on checking.
For the second question, picture the output going wrong and ask what happens. Error cost is that consequence. A high cost means a low score.
An awkward sentence in an internal note costs almost nothing. Someone frowns and you fix it, so it scores a five. A wrong figure in a client report costs more: it can damage trust, lead to a bad decision, or take a week of emails to correct. A wrong statement in a legal letter, a regulatory filing, a price quote, or anything involving someone's money or health costs more still, and scores a one or two.
Cost depends on where the output goes as well as on the task itself. A rough contract summary for your own understanding is low cost. The same summary pasted into a client email is high cost, because someone else will now rely on it. When you score, think about who will read the output and what they will do with it.
The third question, fit, draws on the whole course. Language models are strong at language-based work: drafting, summarising text you provide, rewriting in a different tone, reformatting, translating between plain and formal writing, and brainstorming options. These strengths match lessons 3.2 and 4.1, where the likeliest text is usually useful text.
Models are weaker at the following kinds of work, and tasks like these score low on fit:
fresh facts, unless the model can search (lesson 4.4) exact figures and references (module 6) long documents where one buried detail matters (module 5) judgement about people, such as whether a client is about to leave or whether a candidate suits the team
A task can mix both. The wording may fit well while the facts must come from your own records. Marcus scores his status updates five on fit. His volume report scores two, because it depends on exact numbers. The complaint reply falls somewhere in between. The wording fits, but the facts about the delay must come from his own records.
Put the three scores side by side and you have a quick ranking. Tasks that score high on all three are the best candidates to hand over.
The scores measure only accuracy and effort, and some tasks score well on all three but still belong to you. Before you hand over a candidate, check it for three reasons to keep it.
Accountability applies when your name, signature or professional licence is on the output. You must be able to stand behind every line, and saying the AI wrote it will not help. Confidentiality covers client data, staff records, and anything covered by an agreement or by the PDPA. That material may not be yours to paste into an outside tool, so check your company's policy first. Relationships cover condolence messages, difficult conversations with staff and personal apologies. These need your own words, because the recipient will be able to tell.
Marcus decides to keep the complaint reply himself. The facts were simple, but the client had been with the company for years and he wanted the apology to sound like him. He keeps it for the relationships reason.
Lesson 8.3 turns this ranking into three clear zones. Before then, try the scoring on five tasks from your own week. It goes faster than you would expect, and the ranking is often surprising.
Pick five of your tasks. Score each from one to five on checkability, cost of error and fit, then see which ones come out on top.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).