ChatGPT, Claude, Copilot or Gemini · Practical walkthrough

Resolve conflicting units before asking AI for a comparison

One team reports tickets closed in September. The other reports tickets opened per week. Ask AI which team handles more, and it will tell you, confidently and in one line.

Task by Task6 min read + practice

What you'll make

a short list of questions to ask before comparing two numbers, and a comparison you can trust once they’re answered.

What you'll need

the numbers you want to compare, with whatever labels came with them, and your work AI account (ChatGPT, Claude, Copilot or Gemini). About 10 minutes.

Example result.

A: 340 tickets closed, 1–30 September (a month, closed)
B: 85 tickets opened per week, average (a week, opened)
Not comparable yet. To compare, I need one of:
 - tickets B closed in September, or
 - tickets A opened in September

1. Ask what each number measures

Before any arithmetic, ask the AI to describe each number: what it counts, over what period, for whom, and whether it was observed or estimated. The usual mismatches are period (week or month), measure (opened or closed, ordered or delivered), people (FTE or headcount) and money (with or without VAT).

Before comparing anything, describe each of these numbers: [numbers].
For each, say what it counts, the period, who or what it covers, and whether it's observed or an estimate.
Then say whether they can be compared as they are. If not, list the questions I'd need answered first.
Don't fill in a missing fact with what sounds typical.

[numbers] are the figures exactly as they came to you, with their labels.

Why this works: describing each number first exposes the mismatch before it disappears into a calculation. “Don’t fill in a missing fact” stops the AI from assuming four weeks in a month or that opened means closed.

What you should see: one line per number, a clear “comparable” or “not comparable yet”, and a short list of questions.

2. Get the missing facts, or label an estimate as an estimate

Send the questions to whoever owns the data. If you can’t get an answer and still need a rough figure, you can ask for an estimate, but give the assumptions yourself and keep the result labelled.

I can't get [missing fact] yet. Estimate [question] using only these assumptions: [assumptions].
Show each calculation step with its units. Label the result as an estimate,
say which assumption it depends on most, and keep the original numbers unchanged.

[missing fact] is what you couldn’t get; [question] is what you want to know; [assumptions] are the ones you’re willing to stand behind.

Why this works: an estimate with named assumptions can be checked and corrected. An estimate with hidden ones just looks like a fact.

What you should see: a short calculation with units on every line, and the word “estimate” next to the result.

3. Compare like with like, and say what it doesn’t show

Once both numbers count the same thing over the same period, the comparison is simple. Ask for it with units, and for one line on what it leaves out, such as team size or ticket difficulty.

Both numbers now count [measure] over [period]: [numbers].
Give the difference in units and as a percentage of the smaller number.
Then add one sentence on what this comparison doesn't tell me.

[measure] and [period] are what both numbers now share, such as “tickets closed” and “September”.

Why this works: the comparison is now honest, and the “doesn’t tell me” line stops a workload figure turning into a performance verdict.

What you should see: a difference with units, a percentage and one plain limit.

Check the result

Both numbers cover the same period, or the conversion is shown with its units.
Both count the same thing: opened with opened, closed with closed.
Any assumption, such as “4.3 weeks in a month”, came from you, not from the AI.
A chart or headline doesn’t present an estimate as a measured result.

The usual failure: the AI quietly converts one number and declares a winner. It multiplies a weekly figure by four, ignores that one team counts opened tickets and the other closed, and says “the teams handle the same workload”. Spot it by asking where the “× 4” came from, and whether both numbers count the same thing.

Your answer compares [number A] with [number B] as if they measure the same thing.
Go back to what each one counts and its period, remove the comparison,
and list the questions I'd need answered before comparing them.

Worked example: two support teams, two kinds of count

Fictional example. A service manager wants to know which support team is busier.

Team A: closed 340 tickets in September
Team B: opens about 85 tickets a week (average from its dashboard)
Question: which team handles more?

First answer (wrong): “Team B handles 85 tickets a week, which is 85 × 4 = 340 a month. Both teams handle the same workload.”

Repair: the repair prompt above, then the step 1 prompt.

Repaired answer: the example result at the top of this page: two different measures over two different periods, and two questions. Team B’s lead replied that it closed 298 tickets in September. Step 3 then gave: “Team A closed 42 more tickets than Team B in September (340 against 298), about 14% more. This doesn’t account for team size or ticket difficulty.”

The “× 4” was invented. September has 30 days, about 4.3 weeks. At 85 a week, Team B would have opened about 364 tickets (85 × 30 ÷ 7), not 340.
Even 364 isn’t comparable with 340. Tickets opened is incoming work; tickets closed is finished work. A team can open plenty and close few.
One email to Team B’s lead for September’s closed count answered the real question. The answer wasn’t “the same”: A closed 42 more.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Convert the period and the measure separately

Changing weeks into months doesn’t make “opened” mean “closed”, and it doesn’t make headcount into FTE. Settle what each number counts first, then convert the period with the real number of days. Don’t apply one multiplier to every row of a table.

Keep a data dictionary for numbers you use every month

If the same report comes round every month, write down what each field counts, its unit, its period and where it comes from. Give that to the AI with each new export. See data dictionaries.

Prices have their own traps

For supplier quotes, the mismatch is often per user against per pack, monthly against annual, or a fee that’s missing. The supplier price guide covers that case.