Deciding whether somebody needs to go to A&E, can wait for their doctor, or can be
handled at the counter is one of the highest-consequence decisions made in a pharmacy, and
also one of the fastest: sometimes there are thirty seconds and a person waiting. This lesson
draws the exact line for what role AI can play there, because it is different depending on
the stage of the process, and confusing the two stages is where the real risk lives.
The whole thing in one sentence. AI can help you not forget
a question or a warning sign before deciding. It cannot decide for you
whether to refer, because that decision depends on things only you have in front of you — how
the person looks, their tone of voice, thirty years of cases seen — none of which fits into a
written description.
The two questions that are not the same question
| Question | Who can answer it |
|---|
| "What warning signs should I rule out before deciding, for this
symptom?" |
AI helps well here: it is a general-knowledge question — what is known about that
symptom — that does not depend on looking at the person in front of you, and it is
exactly the kind of checklist that helps you not forget anything under time
pressure. |
| "Given this description, should I refer or not?" |
Not here. This is the question that decides, and deciding requires what AI does
not have: seeing it, not reading it. The same written sentence — "bad headache" — can
describe something trivial or an emergency depending on how the person says it, and
that nuance is lost in the text before it even reaches the model. |
Why the line sits exactly there, and not one step further. This is not a
matter of the model "not being good enough yet": the decision to refer needs information that
structurally cannot reach it as text in the time available. However much detail you put into
describing the situation, it is still a second-hand description, filtered through what you
chose to mention — and what you chose to mention is already shaped by what you, using your
own judgement, decided was relevant. Asking AI to decide is asking it to trust your filter
without being able to check it, when the whole point of asking was to check your own
judgement against something external.
Three preparation questions that work well
The part of this lesson that actually adds value is not the warning — "don't ask it to
decide" — it is which questions ARE worth asking AI during preparation, before the case
arrives. Three concrete types:
- "What questions should I ask to rule out something serious for symptom X?"
This is the base question, and it works because it asks for no judgement about a specific
person — it asks for a list, which is exactly what a model builds well from general
knowledge.
- "Which symptom combinations change the urgency of this?" For example,
that a headache with fever and neck stiffness completely changes the standard compared to a
headache alone. This is general information about how symptoms combine, not about a case.
- "What are the usual criteria for 'refer now' versus 'can wait until
tomorrow'?" Also general knowledge, useful to have reviewed, as long as it is
understood as a reference framework and not as the answer to today's case.
All three have something in common: they can be prepared last week, last month, or this very
morning, with nobody waiting. That is the practical test for whether you are on the right
side of the line — if the question only makes sense with the specific case in front of you,
it is no longer preparation, it is the decision itself.
A full example, and what is wrong with it
Somebody arrives at the counter with chest pain. It is exactly the case the bias lesson from
this same level already used as an example of how the model can answer differently depending
on how the person is described — here it is looked at from the triage angle.
What is asked of the AI, before deciding: "Give me the warning signs to
rule out for chest pain at a pharmacy counter, and the general criteria for when to refer to
A&E."
What it returns, and how it is used
A reasonable, useful list: crushing pain radiating outward, sweating, breathlessness,
cardiovascular history, sudden onset, worsening with exertion. It is exactly the kind of
checklist that helps you not skip a question with somebody waiting — the pharmacist uses it
as a mental script of what to ask, not as an answer.
With that checklist already internalised, she asks the person the questions, watches them
while they answer — skin colour, whether they are sweating, whether they struggle to speak
in full sentences — and decides to refer immediately based on what she sees and hears, not
on what she had read beforehand.
And now, why this use worked and another would not have:
- The question put to AI was general knowledge, not about the case. "What
should I rule out?" is a question that does not need to see anybody to answer well. "Does
this specific patient need to go?" does need that, and that is exactly why the second
question was never asked.
- The checklist was used BEFORE the decision, not instead of it. The order
matters: first prepare with AI what to ask, then decide with what you see and hear in
person. Reversing the order — describing the case to AI and asking it to decide — would
have been exactly the mistake this lesson exists to prevent.
- What decided the referral was not in any text. Skin colour, difficulty
speaking, the sense of urgency a person conveys: none of that can be transcribed into a
sentence with the fidelity needed for somebody else — or a model — to decide on your behalf
about it.
- If the result had been borderline, the previous lesson's bias would still
apply. With a borderline case, describing it to a model and asking "should I refer
or not?" risks the same trap you already saw with the chest-pain-by-sex example: the answer
can depend on details of how you phrase it that should not matter clinically and do matter
to the model.
And if your pharmacy is not like that
If you work alone, with nobody to run the case past. This is where it is
most tempting to use AI as a "second colleague" for the decision, and exactly where it needs
resisting most. Use it beforehand — to review the warning-sign checklist from memory, calmly,
at a moment with no pressure — not during, once somebody is already there waiting for a
decision.
If you have a written referral protocol for the most frequent complaints.
The protocol always outranks whatever any AI says. Use the tool only to review the protocol
from memory before a shift, never to decide a case that already has a standard written and
agreed by the team.
If the team is young and has fewer cases seen so far. This is where the
preparation checklist is worth the most, precisely because there is less personal experience
to fill memory gaps under pressure. And for the same reason: the less judgement of your own
you have, the easier the temptation to ask AI to decide rather than just prepare, so it is
worth being explicit with somebody in training about where the line sits.
If you handle a lot of phone queries, without seeing the person. This is
where the line gets tested hardest, because part of the information normally visible — how
they look, their colour — is not available to you either. Even so, the voice still carries
information a text does not — tone, difficulty breathing while talking, pauses — so the rule
holds the same: AI prepares what to ask, you decide with what you hear live.
If the case involves a child or an elderly person who cannot communicate well.
Here the referral threshold drops on its own, with or without AI involved, because the
uncertainty about what is actually going on is greater. The preparation checklist still helps
ask the accompanying person the right things, but the "when in doubt, refer" standard carries
more weight here than in any other case in this lesson.
If you are part of an on-call pharmacy, with more emergency cases than average.
It is worth building, calmly in advance, a small personal list of "preparation questions" for
the five or six most frequent on-call complaints, using AI for the draft and the team's
judgement to refine it. Having it already written beforehand beats generating it with AI on
the spot, with somebody waiting in front of you.
When it does not work first time
I have a borderline case and I'd like to ask AI whether to refer,
even though I know I shouldn't.
That feeling is exactly the sign the case is borderline — and when in
doubt, the default referral standard is to refer. You do not need a second opinion from an
AI for that: the doubt itself is already the answer.
I used the AI checklist and I still forgot to ask something
important.
It happens, and it is the reason the checklist needs to already be
internalised beforehand, not read at the moment of the case: reading a list while somebody
talks competes with actually listening to what they say. Reviewing it at a calm moment,
before it is needed, is what makes it genuinely useful under pressure.
Can I paste the case description to AI to help me think, without
asking it to decide directly?
The risk does not depend on how the question is phrased: as soon as
you describe a specific person's case and ask for any kind of reasoning about severity, the
answer you get back pulls your own judgement toward wherever it points, even if you keep
the final decision. It is better to keep the two things completely separate: the general
checklist, without the case; the decision, without AI.
What if the case is not urgent, just a calm doubt about whether it is
worth seeing a doctor this week?
The line stays the same even as urgency drops: AI can give you general
information about when consulting is usually recommended for a given symptom, but the
decision about this specific person is still yours. What changes with less urgency is not
who decides, it is how much time you have to think about it.
Does this contradict what the Level 1 risk-traffic-light lesson says?
No, it is the same rule applied to a specific case. The traffic light
already classifies "clinical decisions about a real case, with consequences if wrong" as
red: never delegated to AI without full human verification. Triage is exactly that kind of
decision, so this lesson adds no new rule: it applies the one you already knew to the case
where getting it right matters most.
How do I tell whether a question I've thought of is preparation or
already about the case?
Ask yourself whether it would make sense to ask it with nobody in front
of you, last week. If the answer is yes — "what warning signs exist for chest pain?" — it
is preparation. If it only makes sense with this specific person in front of you right now
— "does this person need to go now?" — it is already the decision, and AI has no place
there.
Before moving on
- I distinguish "what to ask" from "what to decide", and I know only the first can be
asked of AI.
- I use the warning-sign checklist BEFORE the case, not during, so it does not compete with
actually listening.
- When in doubt, refer — with no need for AI to confirm the doubt.
- I know describing a real case to AI, even without asking it to decide, can pull my own
judgement along with it.
← Revisit: checking against the SmPC
Next: exercises with synthetic clinical cases →
← Back to the AI School