Preparing the patient conversation | AI School for Pharmacists

AI prepares the conversation; it does not have it. That distinction sounds obvious written down and disappears the moment somebody pastes a script into their phone and reads it out in front of the person. What it does do well — and it saves real time — is turn a drug list into the five questions worth asking, in order.

And the script is not recited: you read it beforehand and pick three questions. Text written by an AI sounds like written text, and the moment it sounds like a questionnaire the person starts answering what they think you want to hear — which is exactly the useless answer.

The questions that do not work

They are the ones almost everybody asks, including the ones an AI writes if you do not say otherwise. And they fail the same way every time: they have a socially correct answer, and that is the one you get.

You ask	You get	Ask instead
"Are you taking them all right?"	"Yes"	"Which one do you forget most?"
"Do you ever forget?"	"Now and then"	"This week, how many times?"
"Do you take anything else?"	"No"	"What about for sleep? For pain? Anything from a health shop?"
"Does it agree with you?"	"Fine"	"Have you noticed anything since you started it?"
"Do you understand how to take it?"	"Yes"	"Talk me through how you take it."

The last one is the most useful of the five. "Talk me through how you take it" cannot be answered with a yes, and in twenty seconds you find out whether they split it, take it with coffee, skip it at weekends, or have been on double the dose for two months.

A prompt that prepares and does not replace

Here is a medication list. I am a community pharmacist and I have
five minutes with the person, at the counter.

Give me FIVE open questions, in the order I would ask them, to find
out what they actually take and how they take it. No yes-or-no
questions.

For each one, tell me in a line what I am looking for with it.

Do not give me advice for the patient and do not interpret their
treatment: the questions only.

That last paragraph is what changes the result. Without it the model hands you questions and an unrequested piece of clinical advice, and the second reads exactly as well as the first.

Three things only you can do

Stay quiet. The silence after a question is what makes somebody add the sentence that mattered. No script includes it.
Watch their face. "Yes, yes, I take it" said to the floor is data, and it is on no list.
Change plan. If something big turns up on the second question, the other three are irrelevant. A script does not abandon itself.

And the obvious one, skipped here more than anywhere: you prepare the conversation with AI without identifying anyone. "82-year-old woman on these nine drugs" prepares just as well as with the name — and the name does not add one extra question.

Afterwards

Whatever you agree goes in writing and goes with the person: what changes, from when, and what to watch for. The private area generates that sheet with the warnings already built in — and what you may claim on it is lesson 4 of Level 2.

A full example, and what is wrong with it

The prompt above has a final line that looks like padding: "do not give me advice for the patient or interpret their treatment: just the questions". This is what comes out without that line, with everything else the same.

What is asked for: five questions for a five-minute conversation with an older person taking nine medicines, at the counter.

What it replies

1. Are you taking all your medication exactly as your doctor prescribed?

2. Why do you think you sometimes forget a dose?

3. Do you understand what each of your medicines is for?

4. Have you noticed any adverse effects?

5. Are you taking any other medicines besides the prescribed ones?

I should also mention that with nine drugs there is very likely inappropriate polypharmacy. It would be worth reviewing whether any can be withdrawn, particularly the omeprazole, which is often continued without a current indication. It would also be advisable to stress the importance of therapeutic adherence to them.

And now, what is wrong with that reply:

Four of the five are answered with "yes". Numbers 1, 3, 4 and 5 are closed questions with a socially correct answer, which is the one you get: yes, yes, no, no. You asked for open questions and got closed ones, and not out of disobedience — that is how questions are written in the texts it learned from. Compare them with the table above: number 3 is literally "do you understand how to take it?", which is the lesson's own example of what does not work.
Number 2 accuses before it asks. "Why do you think you sometimes forget?" takes it as given that they forget and asks them to account for it. At a counter that puts the person on the defensive by the second question, and from then on the remaining three get answered to look good. It is the most expensive failure in the whole interview and it does not show when you read the script: it shows on the face of the person hearing it. "This week, how many times?" asks the same thing without accusing.
The order starts with the most uncomfortable one. You asked for "the order in which I would ask them" and it put adherence first, which is precisely the one to ask once there is some trust. A five-minute conversation opens with the easy, concrete things — what they take, how they take it — and the delicate part arrives once the person is already talking. Order is not presentation: it is part of the technique, and the model orders by topic importance because that is what an index does.
And the closing paragraph is what the missing line prevented. Nobody asked for an interpretation of the treatment and there it is: a judgement on a patient it has not seen, with one drug named and a withdrawal recommended. And the worst of it is that it sounds reasonable — omeprazole continued without a current indication is a genuine classic. But it has been said about a list of nine names, without knowing why any of them was prescribed, and said before the conversation that existed to find that out.

That closing paragraph is also the stickiest. Anybody who reads it before the interview walks in already looking for the omeprazole, and from then on the conversation stops doing what it was for: an interview with a hypothesis inside it only confirms the hypothesis. Which is why the line that forbids it is not politeness towards the model — it is what protects the one step in the whole of Level 3 that a machine cannot do.

And checking the script before you keep it takes ten seconds: read it and cross out every question that can be answered yes or no. If fewer than three are left, ask again and say explicitly "no question may admit yes or no as an answer".

And if your pharmacy is not like that

If the person who comes in is the carer, not the patient. The questions change entirely, not the tone: the carer knows what happens at home — whether they choke on it, whether breakfast gets skipped, whether some days they do not get up — and does not know how the person feels. "Talk me through a normal medication day at home" gets more in thirty seconds than five adherence questions. Say so in the prompt, because otherwise you get the patient script with the pronouns swapped.

If the person has trouble hearing or expressing themselves. Open questions are still better and they need breaking up: "talk me through how you take them" is too big and gets answered with "fine". It works better to go by time of day — "what do you take when you get up?", "and with lunch?" — because each one is a single thing to remember. Ask the model for that and the script changes shape. And budget ten minutes, not five.

If there is little privacy at the counter. Then half these questions cannot be asked, and that is not a logistical detail: nobody admits skipping doses with three people behind them. The practical move is to ask the model for two scripts — a two-question one for the counter and a full one for the consultation area — and use the first to decide whether the second is worth it. An interview in the wrong place returns polite answers, which is worse than not asking.

If this is your first time doing an interview like this. Take three questions and one person, not five and a plan. And of the three, make one "talk me through how you take them", which returns more than anything else in this lesson. What surprises you the first time is not what you learn about the treatment: it is how much people talk if you let them — and no script prepares that, you only discover it by staying quiet after asking.

If the conversation is on the phone. You lose the one thing that was never in the script — the face — and that changes what can be asked: without seeing the person, you have to ask them to narrate rather than to confirm, because a "yes" over the phone carries no extra information. "Talk me through what you have taken today, from when you got up" works far better than any adherence question. And say at the start how long it will take: on the phone, three seconds of silence reads as the line dropping.

When it does not work first time

I ask for open questions and it keeps giving me closed ones.

Give it the criterion instead of the adjective: "no question may be answered with yes, no, fine or sometimes". "Open" is a label the model recognises and half-applies; an explicit ban on those four answers is checkable and it obeys it. And if one still slips through, no harm done — crossing them out as you read the script is part of the method and takes ten seconds.

I ask properly and I get "yes, yes, all fine".

Then try the technique that is not a question: silence. Ask, and do not fill the gap. Three awkward seconds make most people add the sentence that mattered — "well, the one for my sugar I do skip sometimes". It is the one thing in this lesson no model can write and the one that changes the result most. If nothing comes even so, do not push that day: try again next time they come in, now that they know what it is about.

They tell me an enormous amount and I run out of time.

This is the good problem, and it resolves the opposite way to how it looks: do not cut them off, note it and book them. "This matters — can you come back on Thursday and we will go through it properly?" turns an overflowing conversation into a real review, and the person comes back — they have seen somebody listening. What does not work is trying to resolve it all standing up: you come away with half a story and the feeling you have done neither thing.

Can I keep the script on my phone and read it in front of them?

Read it beforehand and put it away. The moment the person sees you reading, the conversation becomes a questionnaire and the answers change — they start answering what they think is wanted, which is exactly what is no use. Three questions can be carried in your head, which is why the lesson says pick three and not five. If you need to write something down, write the answer, not the next question.

Before you call this learned

I ask for open questions and forbid clinical advice in the same prompt.
I read the script beforehand and pick three; I do not recite it.
I ask "talk me through how you take it", not "do you understand?".
I prepare without identifying anyone.
What we agree goes out in writing with the person.

← Previous: medication review Next: checking against the SmPC →