Module 1 · Why now, honestly
Where AI helps, and where it does not
In short
AI helps most when it works on material you give it and the result is easy to check. Start with one small, repeating weekly task you know well.
After this lesson you can
- Recognise the kinds of work AI helps with best today
- Judge a task by how easily its result can be checked
- Choose one of your weekly tasks to try with AI
Terms in this lesson
Understand
A large language model is strongest when it works on material you give it. Summarising minutes you paste in, rewriting a paragraph to sound more polite, turning a list into a table, building a report outline from your points, or explaining a term in a document you are reading. In this kind of work everything the model needs is in front of it, and the work is mostly about form and language, which is exactly its skill.
A model is weakest when asked to recall facts without material. Market figures, regulation numbers, latest prices, or a decision made only yesterday. Here the model can hallucinate: write an answer that sounds convincing but has no basis. It is also weak when the task depends on unwritten context, such as the history of a client relationship or the reason an internal rule was made. Context is all the information needed to understand a task, and the model only knows the context you write down.
There is one simple yardstick that is more useful than a list of suitable and unsuitable tasks: how easily you can check the result. A summary of a meeting you attended is easy to check, because you know what was discussed. A summary of a hundred-page regulation you have never read is hard to check, because you have nothing to compare it with. The first might take five minutes to check. The second could take longer than reading it yourself.
The rule: start with tasks whose results you can check faster than doing them yourself. That is where AI really saves time. If checking takes as long as doing, the saving disappears. If you cannot check the result at all, you carry the risk without knowing it.
To pick a first task, look at your past week of work and find a task that meets four conditions. First, it repeats, at least weekly, so your practice gets used again. Second, you know exactly what a good result looks like, so you can judge the output. Third, checking the result is faster than doing it. Fourth, the material holds no personal or confidential data; data is covered in the module "Protect the data".
Examples that often fit: tidying weekly meeting minutes, drafting replies to frequently asked questions, writing a project status summary from your own notes, or turning data from emails into a table. Examples to hold off on: financial reports that go outside, employee assessments, or legal answers for clients. Not because AI cannot help at all, but because a mistake costs a lot and you have not yet mastered how to check it.
Before
The prompt
Please write an analysis of office stationery sales trends in Indonesia this year, with figures.
What came back
The office stationery market in Indonesia is estimated to grow 7.4 percent this year to about Rp 21 trillion. Online sales account for 38 percent, driven by hybrid work policies. Paper and ink remain the largest categories.
After
The prompt
Goal: I need a summary of how our shop's sales changed this week compared with last week. Reader: my manager, reading it in two minutes. Material: a sales table by category for two weeks, pasted below. Limits: use only figures from the table. Do not guess causes that are not in the table. Shape: three points, then one sentence on the category that needs attention. Check: name the table row behind each point. [table pasted here]
What came back
1. Paper rose from 120 to 154 reams (row A4 Paper). 2. Ink fell from 48 to 31 bottles (row Ink). 3. Folders and binders were about the same, 62 and 60 units (row Folders). Needs attention: ink fell the most. The table does not show why, so check stock or the supplier.
What changed
- The first request asks for facts from the model's memory. The figures sound plausible, but you have no quick way to check them.
- With the table pasted in, the model's task becomes working on your own material. Each point can be matched to the table in a minute.
- The model is also allowed to say the table does not show the cause, so it does not invent one.
Try
Sorting card
Sort these tasks by how easily you can check the result.
Tidying minutes of a meeting you attended
Summarising a new regulation you have never read
Rating employee performance from a set of notes
Draft replies to a frequent customer question
Turning an order list from emails into a table
Stating competitors' market share figures
6 cards left
Help me choose a first task
Works with any AI assistant.
Your prompt
I want to start using AI for one weekly task. Here is my list of routine tasks this week: ‹daftar-tugas› For each task, briefly assess: does it repeat, do I know what a good result looks like, is checking the result faster than doing it, and does the material hold personal or confidential data. Then suggest one task to try first and why. Do not suggest tasks that involve personal data.
Common mistake
- Mistake
- Trying AI for the first time on the hardest or most important task, so the result feels impressive.
- Why it hurts
- Such tasks are the hardest to check and the costliest when wrong. You may be disappointed and give up or, worse, trust a result you had no time to check.
- Do this instead
- Start with a small repeating task whose result you know well. Move to heavier tasks once you know where the model usually misses.
Take home
A one-week plan
- Write down one weekly task I will try with AI.
- Write what a good result looks like, before starting.
- Note how many minutes checking took, and compare with doing it myself.
- Note where the model missed, as material for next week.
What went wrong for us
AI features ship switched off
- Context
- Almost every app of ours gets AI features, from summaries to scoring. Their automated tests use fake model answers so they are fast and repeatable.
- What tends to go wrong
- An AI feature can pass every test with a fake model and still give poor results with the real model on real data.
- What it costs if ignored
- Without this rule, users could be the first to see poor results.
- Root cause
- Tests with a fake model test the code around the model, not the model's behaviour.
- How we do it
- Every new AI feature gets a flag that is off by default.
- The rule we keep
- The flag is turned on after the owner checks the real model's results on real data; that check is a release step of its own.
The lessonLet a person see the AI's results on real data before users do.
What it means for your workBefore others rely on AI output, try it on your real work and look at the results yourself; tidy practice examples prove nothing yet.