AI School · Level 1 · Practical case · First steps, no risk

A colleague shows you the salary ChatGPT gave them: checking it

Léelo en español →

Reading time: about 11 minutes

The situation. A pharmacy technician on your team shows you their phone: "I asked ChatGPT what I should be earning this year and it says less than my payslip. Am I being overpaid, or is the machine wrong?" They are asking you because you are the one who knows about this.

This is the most common situation of all and the best one to learn from, because the right answer exists, is published and fits on one line. You do not need to know anything about AI to solve it: you need to know where the right figure lives and refuse to accept any other. What you practise here works for any number a chat hands you, whether it is a salary, a deadline or a contribution rate.

What you need to hand: Your colleague’s phone with the conversation open, this year’s pay table, which is published in the official gazette (the Spanish version of this case links to the table as this site keeps it) and five minutes. Nothing from their payslip goes into the chat: checking a table does not need anybody’s payslip.

Step by step

  1. Read the question they asked, not just the answer

    Before looking at the figure, look at what was asked. Nine times out of ten the problem starts there. "How much does a pharmacy technician earn?" is a question the model cannot answer well, because four things are missing: which collective agreement (the national one is not the Bizkaia one or the Asturias one), which exact job category, which year, and whether they mean base salary, gross monthly pay or take-home pay.

    A model will never tell you "I am missing information" unless you ask it to. It fills the gaps with whatever is most likely and answers in exactly the same confident tone as if it knew. So the first question for your colleague is not "what did it tell you?" but "what exactly did you ask it?".

    If the question was vague, the figure is worthless even if it happens to match the right one. That is already half the lesson: a correct answer to a badly asked question is a lottery win, not a fact.

  2. Ask again with all four details, and ask for the source

    In a new conversation — so it does not drag in what came before — ask the full question. Notice that it contains no personal data: no name, no length of service, no pharmacy. You do not need any of that to know what a table says.

    I work in a community pharmacy in Spain.

I want the MONTHLY BASE SALARY for the job category pharmacy technician
under the National Collective Agreement for Community Pharmacies
for the year 2026.

Give me the figure and, directly underneath, the exact reference of the
official gazette resolution it comes from (number BOE-A-…).
If you do not know the table for that specific year, tell me so plainly
instead of giving me the one for another year.

    The last line is the important one. Without it, what usually happens is exactly what happened to your colleague, and you will see it in the example below.

  3. Do not believe the reference: open it

    If it gives you a gazette number, do not look at it: open it. Any of these three things can happen, and all three are common:

    The reference exists and the figure is in it. Perfect — but then what you rely on is the gazette, not the chat. The chat saved you looking for it, which is not nothing.

    The reference exists but it is from another year. This is the most treacherous case, because everything looks right: the document is real, it is the right agreement, and the figure is in there. It is just the table for a year that is no longer yours.

    The reference is perfectly shaped and leads nowhere, or leads to a resolution from another sector. An invented gazette number looks exactly like a real one. That is why you do not look at it: you open it.

  4. Compare with the right table, not with another answer

    This is the most common mistake when checking: asking another chat. If ChatGPT gave you one figure and Gemini gives you another, you know no more than before; you know less, because now you have two numbers and no way of choosing. And if they agree, that means nothing either: they may both have read the same old text.

    For the national agreement in 2026 the table is published in resolution BOE-A-2026-4220, and in it the base salary of a pharmacy technician is €1,507.27 a month, paid 14 times a year. That is the figure to compare against, and the one printed in the resolution itself, which anyone can open and read.

    If you work under a different agreement — Bizkaia, Gipuzkoa, Asturias, Girona-Tarragona-Lleida — the figure is different and so is the table. Each one has its own published version, and that is the one that counts for your payslip.

  5. Explain the difference to your colleague, with no payslip involved

    With the right figure in front of you, the conversation with your colleague changes. It is no longer "the machine says one thing and the payroll office says another": it is "the table says this, and your payslip has to reach at least this".

    If their payslip is higher, they are not being overpaid: they may receive supplements, a consolidated seniority payment or a voluntary improvement. If it is lower, look for an "advance on the agreement" line or an absorbable supplement that brings the total up — some payroll offices pay a lower base and top it up with an allowance. Only if, with everything added, it still falls short, is there something to claim.

    And this is where the AI stops having any role. The review of the actual payslip is done by them with the table, or by the two of you together. What you do not do is paste their payslip into a chat so it can "check it": it carries their name, their ID number and their social security number.

A full example, and what is wrong with it

This is what a model replied to the full question from step 2 — with all four details — but without the last line asking it to say so if it does not know that year’s table. It is shortened and not polished.

What it replies

"Under the National Collective Agreement for Community Pharmacies, the monthly base salary for a pharmacy technician in 2026 is €1,477.71, paid in 14 instalments (12 monthly payments and 2 extra payments).

This figure comes from the pay tables published in the official gazette for the 25th Agreement. The resulting gross annual salary would be approximately €20,688.

Bear in mind that supplements such as seniority or night work may be added to this amount, depending on each case."

And now, what is wrong with that reply:

  1. The figure is real, but it is last year’s. €1,477.71 is exactly the technician’s base salary in the 2025 table. The 2026 figure is €1,507.27. It has not invented a number: it has taken a true one and put a different year on it. It is the hardest failure to spot, because if you search for it you will find it "in the gazette".
  2. The source is vague on purpose. "The pay tables published in the official gazette for the 25th Agreement" sounds like a reference and is not one: there is no resolution number, so there is nothing to open. A source you cannot open is not a source, it is a sentence.
  3. The arithmetic is right, and that is why it convinces. €1,477.71 × 14 is €20,687.94, and "approximately €20,688" checks out. A correct calculation on a wrong figure gives a wrong result that looks verified.
  4. It talks about seniority as if there were three-yearly increments. The national agreement removed them; only a personal guarantee supplement remains for people who were already receiving them. It is a detail, but it is the kind of detail a model repeats because it has read it a thousand times in old texts.

Look at the difference: €1,507.27 − €1,477.71 is €29.56 a month, around €414 a year over 14 payments. That is exactly the 2 % rise from one year to the next, and it is exactly what your colleague saw as "extra" on their payslip. The payslip was right and the chat was a year behind.

Now repeat the question with the last line. What usually happens is that it says it cannot guarantee the table for that specific year and points you to the gazette. That reply looks worse — it gives you no figure — and it is far better: it has not given you a wrong one.

And if your pharmacy is not like that

If you work under a regional agreement. Ask about it by its full name and be twice as suspicious. There is much less published text about the Bizkaia or Asturias agreements than about the national one, so the model is more likely to mix figures up. In Asturias, seniority is paid every five years; if it talks about three-yearly increments, it is thinking of a different agreement.
If you work in Barcelona. The Barcelona agreement was annulled and the national one has applied since January 2025. Plenty of old texts still describe the Barcelona agreement as current, and a model trained on them will give you its tables with complete confidence. If the answer does not mention the annulment, it does not know the current situation.
If the question is about take-home pay. Then even the table is not enough: take-home pay depends on income tax withholding, which is personal, and on the employee’s social security contribution, which in 2026 is 6.48 %. A chat knows nothing about anybody’s family situation. For an approximate net figure, use a proper calculator with your own numbers, one that does not send them anywhere.

When it does not work first time

I tell it the year is wrong and it gives me a different figure.
Do not accept that one either. When you correct it, a model tends to agree with you and produce a new figure that sounds corrected, not necessarily the right one. To check, you go to the table; arguing with the chat checks nothing.
It gives me the right gazette number and the right figure.
Good, but keep opening the gazette for the first few times. What you are training is not trusting that it got it right this time: it is the reflex of opening the source before repeating the number to anyone. That reflex is what protects you the day it gets it wrong with the same composure.
My colleague insists that "AI does not get numbers wrong".
Show them the example above. The 2025 figure instead of the 2026 one is real, it is in a gazette, and the annual calculation is done correctly. It did not make an arithmetic mistake: it got the year wrong. That is what a model does, and you cannot tell by reading.
I cannot find the table for my agreement anywhere.
Ask the chat where it is published, not what it says. "In which official bulletin are the pay tables for the community pharmacy agreement in [your province] published?" is a question for orientation, and it saves you time even though you then have to go and find it yourself. What you do not do is keep the figure it gives you.

Before you use what it gave you

Last year’s figure. It is the most common failure with any data that changes every year: pay tables, contribution bases, prices. There are far more texts from last year than from this one, and the model repeats what it has read most.
A source without a number. "According to the gazette", "according to the agreement", "according to current regulations". If it does not carry a reference you can open, it is not a source. Ask for one, and if it does not have one, it should say so.
Mixing agreements. Increments where there are none, payments that do not apply, allowances from another province. The less well known an agreement is, the more likely the model is to fill in with what it knows about another one.
Now, with your own. Try it with your own category and your own agreement, but without your payslip: ask for this year’s base salary using the template from step 2, open the reference it gives you and compare it with the table. If it is right, note down the gazette number for next time. If it is wrong, now you know what a failure looks like when it does not look like one.

← All practical cases Back to the school →