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

Your week through the traffic light: what you give the AI and what you do not, task by task

Léelo en español →

Reading time: about 11 minutes

The situation. You have read the Level 1 traffic light and it seemed sensible. But on Monday morning, with a queue at the door, you cannot remember any colours: only that there are twenty things to do and an AI open on the computer. What is missing is not the theory, it is having applied it once to your week.

This case is the practical close of Level 1. You will list what you actually do in a normal week, sort each task with the five traffic-light questions, and use AI for one very specific thing: to argue against you. Along the way you will see something worth knowing early: when you ask an AI to sort which tasks can be handed to it, it tends to say nearly all of them. Not out of bad faith, but because almost everything written about AI is enthusiastic.

What you need to hand: Half an hour of peace, paper or a note on your phone, and the lesson on when not to use AI open in another tab. The list contains no names of patients or colleagues: "replying to a patient’s WhatsApp about their treatment" is a task; you do not need the patient’s name to sort it.

Step by step

  1. Write down the week as it is, not as it should be

    Take a sheet of paper and write down everything you did last week that involved writing, reading, looking something up, calculating or deciding. Not what a pharmacy does in theory: what you did. It usually comes to between fifteen and twenty-five things, and it helps if there is a bit of everything: the counter, the back room, the team and the management side.

    Write them as concrete verbs. Not "patient communication", but "replying to a WhatsApp asking whether they can take ibuprofen with their treatment". Not "admin", but "checking a technician’s payslip before signing it". The more concrete the task, the easier it is to sort, because the colour almost never depends on the subject: it depends on where the output goes.

    A common trap is leaving out small tasks because "I would never ask an AI that". Write them down anyway. They are exactly the ones that end up in the chat one busy day without anyone thinking about it.

  2. Sort them yourself first, without the AI

    Before asking a model anything, colour each task yourself. For the ones you are unsure about, use the five questions: does it leave my screen without me checking it? is there an identifiable person? does it depend on an exact figure? would I know if the answer were wrong? is it a clinical decision? One answer on the wrong side is enough for it not to be green.

    Do it quickly, without overthinking: what you want is your first instinct, because that is what you will use on Monday in a hurry. Afterwards you will mark where you got it wrong, and that is what teaches most.

    Usually you end up with few greens, quite a few ambers and a handful of reds. If nearly everything comes out green, go back to the fourth question — would I know if it were wrong? — which is the one most people skip.

  3. Ask the AI to argue against you

    Now the model comes in, but with a very precise role: it is not the referee, it is devil’s advocate. You give it your sorting and ask it to find where you have been too permissive. That is the direction in which you want it to err, if it errs at all.

    I work in a community pharmacy in Spain. I have sorted my tasks for the
week into a traffic light to decide which ones I can do with help from AI:
GREEN (no risk), AMBER (with a source and a check), RED (not done with AI).

My criteria: whether the result leaves my screen without being checked,
whether there is an identifiable person, whether it depends on an exact
figure, whether I would spot a mistake, and whether it is a clinical decision.

Here is my list:
paste your list with your colour next to each task

Look ONLY for the tasks where you think I have been TOO PERMISSIVE
(green that should be amber, amber that should be red) and explain the
specific risk in one line. Do not suggest moving any task towards green.

    The last line is deliberate. If you let it suggest changes in both directions, nearly all its energy will go into telling you that you are being too cautious, because that is what it has read in a thousand articles about productivity.

  4. Compare, and keep whatever surprises you

    Put the two sortings side by side. Where they agree you learn nothing; where they differ you do. For each task the model moved down a colour, ask yourself whether its reason is good. Sometimes it will be, and it will have pointed out something you had not seen: an email that ends up forwarded to a patient, an internal text that ends up pinned to the counter.

    And sometimes its reason will be generic — "it is always wise to check" — and will not apply to your case. Discard it without guilt. What you are after is not agreement with the machine, but the two or three specific places where your Monday-morning instinct would have let you down.

    Note those two or three tasks separately. They are your personal "watch out" list, and it is far more useful than the generic traffic light because it is built from your own week.

  5. Write your pharmacy’s rule in one line per colour

    The last step turns the exercise into something the whole team can use: a sheet with three lines. What can be done with AI without asking anyone, what can be done with AI but gets checked before it goes out, and what is not done with AI in this pharmacy.

    Write it using tasks from your list, not abstract categories. "Green: drafting internal notes, organising information from a catalogue, getting ideas for Monday’s training." "Amber: any text going to a patient or a doctor; any figure." "Red: nothing with data about an identifiable patient; no decisions about a treatment."

    If you work with other people, this sheet matters more than your own judgement, because if everyone applies their own, the most relaxed person decides for everyone. The lesson on professional responsibility explains why the signature is still the pharmacy’s, whoever wrote the text.

A full example, and what is wrong with it

This is what a model replied when, instead of the prompt from step 3, it was given the list without colours and asked "sort these tasks by whether I can do them with AI". These are five of the twenty tasks from a real week. It is shortened.

What it replies

"Green — Replying to a patient’s WhatsApp asking whether they can take ibuprofen with their treatment: AI can draft a clear, empathetic reply in seconds.

Green — Preparing Monday’s training on sunscreens: ideal for AI.

Green — Checking a technician’s payslip before signing it: AI can verify the calculations and spot errors.

Amber — Writing the poster for the autumn campaign: it is worth checking the tone.

Green — Summarising the professional body’s circular about the change to the on-call rota: AI summarises documents with great accuracy."

And now, what is wrong with that reply:

  1. The patient’s WhatsApp is red, not green. It is a question about a specific person’s medication: there is an identifiable patient on the other end, the reply leaves your screen and reaches them, and the decision is clinical. That AI can "draft an empathetic reply" does not make it green: it makes it a risk with good spelling. Here, at most, AI helps you write a sentence moving the conversation to the counter or the phone.
  2. The payslip is red twice over. It carries a name, an ID number and a social security number, so it does not get pasted; and it also depends on exact figures that a model can get wrong in the same confident tone. "Verifying the calculations" is exactly what it does not do well. That is what the published pay table is for.
  3. The circular is amber. Summarising is a good task, but a circular about on-call duty contains dates, times and obligations: exactly the precise details a summary swallows. Summarising it with AI is fine; organising the rota from the summary without reading the circular is not.
  4. The only amber is the one with the least risk. The autumn poster stays inside the pharmacy until it is printed, and its risk — health promises — is well covered in the lesson. That it is the only thing marked with caution says a lot about the criterion being applied: it looks at the type of text, not at where it goes.

Four out of five sortings too permissive, all in the same direction. That is no accident: the model reasons about whether AI can do the task, which is nearly always yes, rather than about what happens if it does it badly, which is the traffic-light question.

That is why step 3 does not ask it to sort: it asks it to look for your mistakes in one direction only. Used that way, its biases work in your favour.

And if your pharmacy is not like that

If you work alone in a small pharmacy. The list will be shorter and more mixed, because you do everything. That is exactly why it helps: with nobody to check you, the three-line sheet is your second pair of eyes. Read it before using AI for anything new, at least for the first few weeks.
If you own the pharmacy and want to roll this out to the team. Do the exercise yourself first, then repeat it in a meeting with everyone’s list. Tasks change with the role: the technician in the back room does not have the same week as whoever runs the admin. The final sheet has to cover everyone’s tasks, or whoever does not see themselves in it will apply their own judgement.
If AI is already used in your pharmacy with no rules at all. Do not start by banning things: start by asking what each person uses it for and run it through the traffic light together. There will nearly always be one or two red tasks someone was doing without thinking, and they are much easier to fix when the team has found them itself.

When it does not work first time

Almost everything comes out amber and I lose heart.
That is to be expected, and amber is not bad: it is the working zone. It means "yes, with a source and a check". What helps is knowing the source for each amber, so that checking is not a vague idea but a specific place to look.
The model moves nearly everything down a colour, even though I asked only for the permissive ones.
Then it is being generic. Ask it, for each move, to say concretely what would happen if the task went wrong and who would notice. Discard the ones it cannot make concrete.
I do not know which colour to give a task.
Apply the fourth question on its own: would you know if the answer were wrong? If not, it is not green, however innocent it looks. That one settles most doubts.
My colleagues say this is over the top.
Show them the example above, especially the WhatsApp and the payslip. Hardly anyone argues that those two are red once they see them written down; the problem is that on a busy Monday nobody stops to think. The sheet exists for that.

Before you use what it gave you

It sorts by type of text, not by where it goes. "Drafting" sounds green to it, so a WhatsApp to a patient looks the same as an internal draft. What decides the colour is what happens if it goes wrong.
It confuses "it can do it" with "it can be handed over". A model can do almost anything; the traffic-light question is whether you can afford for it to do it badly. Its answers nearly always address the first.
It is optimistic by default. What has been written about AI in recent years is mostly enthusiastic, and that shows in how it sorts. Use it to look for your permissive mistakes, never to give you permission.
Now, with your own. Make the list today, while last week is still fresh. Sort it, run it through the prompt from step 3 and write the three-line sheet. Put it where you will see it — next to the back-room computer works well — and in a month do the exercise again: you will find some tasks have changed colour because you now know where their source is.

← All practical cases Back to the school →