Module 6 · Produce real work
Checking figures and spreadsheets
In short
A language model predicts text; it does not calculate. Ask for formulas you can read, test them on rows whose answers you know, then check key figures another way.
After this lesson you can
- Recheck important figures by a different method from the one that made them
- Use AI for spreadsheets: formulas you can read, not finished numbers
- Test a formula on a few rows whose answers you already know
Terms in this lesson
Understand
A large language model is trained to predict the next word, not to do arithmetic. For small sums it is often right. For long tables, layered percentages or averages over hundreds of rows, it can write a figure that looks perfectly reasonable but is wrong, with no warning sign. This is one of the most costly kinds of hallucination, because figures are rarely read critically: people trust neat numbers.
So for work with figures, change the AI's role. Do not ask for finished numbers; ask for a formula or calculation steps you can read. "Total overtime this month is Rp 14,250,000" cannot be checked without redoing everything. A formula that sums the overtime column for rows dated October can be read, understood and run by you in your own spreadsheet, on your real data, on your computer. The bonus: the real data never has to be sent to the model. Column names and two made-up sample rows are enough.
Ask for an explanation of each part of the formula in plain words too. If you cannot explain the formula to a colleague, do not use it yet.
Then test the formula before trusting it. Pick three to five rows whose answers you already know or can work out by hand in a minute, including awkward ones: empty cells, numbers stored as text, dates at month end, zero values. If the formula is right for all of them, it has earned use on the whole table. (To test a prompt, see the lesson on improving in rounds.)
For important figures, the ones that go into reports to your manager, invoices or payroll, add one more check: recompute by another method. The key is the word "another". Running the same formula twice only repeats the same mistake. Another method might be:
- A rough estimate. 30 people, about 10 hours of overtime each, about Rp 40,000 an hour: the total should be around Rp 12 million. If the result is Rp 120 million or Rp 1.2 million, something is wrong.
- Totals from two directions. The sum by section must equal the sum by person. If they differ, a row was missed or counted twice.
- A different tool. Compare the formula with a summary (pivot) table or with a calculator for a few rows.
- A different person. For very important figures, ask a colleague to compute it without seeing your result first.
If two methods agree, you have evidence. If they differ, you have just found a problem before someone else did.
Before
The prompt
Here is the October overtime data, please calculate total overtime per person and the grand total. [30-row table: employee name, date, overtime hours, hourly rate]
What came back
Overtime per person: - First employee: 12 hours, Rp 480,000 - ... Grand total: Rp 14,250,000
After
The prompt
I need a spreadsheet formula to total each employee's overtime cost for a given month. I will run it myself on my data. Columns: A = employee code, B = date, C = overtime hours, D = hourly rate. Made-up sample: K-01 | 03/10/2026 | 2 | 40000 K-01 | 15/10/2026 | 3 | 40000 Give one formula for the total per employee per month, explain each part in plain words, and name three situations that could make it wrong (such as empty cells or numbers stored as text).
What came back
Formula (for the code in F2 and month in G2): =SUMPRODUCT((A2:A500=F2)*(MONTH(B2:B500)=G2)*C2:C500*D2:D500) - (A2:A500=F2): only that employee's rows. - (MONTH(B2:B500)=G2): only rows in that month. - C*D: hours times rate per row, then all are added. What can make it wrong: dates stored as text may yield no month or the wrong one; hours or rates stored as text, such as "40.000", can make the formula return an error or read the wrong value; data from another year in the same month is included unless the year is filtered. Test on the sample: K-01 for month 10 should be (2 x 40,000) + (3 x 40,000) = 200,000.
What changed
- The first version gives figures you cannot check, and sends employee names to the model.
- The formula in the second answer can be read, run on real data, and tested on a sample whose answer is already known.
- The "what can make it wrong" list shows which awkward rows to include in the test. One real trap is right here: this formula does not filter by year.
Common mistake
- Mistake
- Checking a formula's result by running the same formula on another sheet, or asking the same AI to recompute.
- Why it hurts
- The same method repeats the same mistake. The results match, you feel sure, yet the error was copied along. This is exactly what happened in the case at the end of this lesson: a test that repeated the formula approved a wrong formula.
- Do this instead
- Check with a truly different method: a rough estimate, totals from two directions, a hand count of a few rows, or another person computing without seeing your result.
Try
An AI-made formula gives this quarter's business travel total as Rp 48 million. Which is truly a check by another method?
Sorting card
You are testing an overtime total formula. Which rows are most useful in the test?
A row with empty overtime hours
A row where the rate is written "40.000" as text
A row dated 31 October
A row whose result already matched last month's report
A row from last year's October in the table
Ten ordinary rows that look alike
6 cards left
Take home
Template to ask for a formula, not a number
Works with any AI assistant.
Your prompt
I need a spreadsheet formula for ‹hasil›. I will run it myself on my data. Columns: ‹kolom› Two made-up sample rows: ‹contoh› Give one formula, explain each part in plain words, name three situations that could make it wrong, and work out the correct result for the sample so I can compare.
Before an AI figure goes into a report
- I use a formula I can explain, not a finished number
- The formula was tested on rows with known answers, awkward ones included
- Key figures were checked by another method
- My rough estimate matches the result
- The real data stayed on my computer
What went wrong for us
A forced-choice block that pulled scores to the middle
- Context
- We built a composite score from several parts of a personality test, including a forced-choice block (pick the most and least like you).
- What happened
- A forced-choice block always balances out: everyone spreads the same points across all traits, so its average is always the same. Giving it weight in the composite pulled everyone's score towards the middle.
- Impact
- Differences between candidates shrank, making the composite less useful. Found before use.
- Root cause
- Ipsative scores (relative to oneself) were treated like scores that compare across people.
- What we did
- A reviewer recomputed a few scores by hand and saw them all drift to the middle. The composite was changed so that block no longer pulls it.
- The guard we keep now
- Scoring formulas are checked with manual calculations on real examples, not only with tests that repeat the same formula.
The lessonRecompute by hand; a test that copies the formula copies its mistake too.
What it means for your workWhen AI gives you a formula for a score or key figure, work out a few examples by hand; repeating the same formula will not catch its mistake.