AI for pharmacy owners: the 17 areas

The ordering-agent lesson and the website one are two out of seventeen. This is the map: what can genuinely be automated in a pharmacy, in what order, and what you need to have for each.

The whole thing in one sentence. In none of the seventeen is the AI the one doing the calculating. Arithmetic does the arithmetic — an average, a deviation, a margin — and the AI does what arithmetic cannot: explain it, prioritise it and write it up. An agent that "decides" with no arithmetic underneath is a text generator with an agent's name.

Where to start

Not with the flashiest. With whatever you already have as data and costs you most hours: usually ordering and dead stock, because they come out of the same file and the saving shows in the first week. The financial and team ones come later, once you trust the first.

The map: seventeen areas

Buying and holding stock

Automated ordering Ready

Proposes what to order and how much; you approve. The quantity comes from arithmetic, the AI explains and prioritises.

Sales per line + current stock

Dead stock Coming

What has not moved in months, how much money that is, and what to do before it expires.

Stock + last sale per line

Purchasing terms Coming

Comparing what each wholesaler offers against what you actually sell, not against the catalogue.

Purchase invoices

Demand forecasting Coming

The real seasonality of YOUR pharmacy: flu, hay fever, the neighbourhood, the month.

Two years of sales per line

Suppliers Coming

Shortages, lead times and substitutions by supplier — the conversation you have with data in front of you.

Order and shortage history

Making money from what you already sell

Sales analysis Coming

What changed this month and why, on one page. Not the listing: the change.

Sales by category and period

Margins Coming

Where you actually make money and where you are working for free. There are usually surprises.

Sales + cost price

Financial analysis Coming

Staff cost, rent and purchases against turnover, month by month.

Your accounts

The team

Opening hours and on-call Coming

Which hours you declare and how the on-call duty fits in: that is an opening decision, not a staffing one.

Declared hours and the on-call calendar

Team rotas Coming

Who covers each hour, respecting the agreement, holidays and rest periods — with the annual hours in front of you.

Staff, hours and holidays

Team training Coming

Preparing a 20-minute session on what actually goes wrong at your counter.

What you get asked and get wrong

Protocols and documentation Coming

Writing the protocol everyone says you should have and nobody has time to draft.

How you do it today

Meetings Coming

Agenda, minutes and actions — with who does what, by when.

The recording or your notes

Being found and being come back to

Your pharmacy website Ready

Building it with AI knowing what the rules require and what you may not claim.

Whatever you want to offer

Social media Coming

A calendar you can actually sustain, without claiming properties you may not claim.

Your health calendar

Marketing and campaigns Coming

Campaigns built on what you already sell, not on what the average sells.

Sales by category

Customer emails Coming

Repeat reminders, seasonal campaigns and follow-up, without wearing anyone out.

Consents and purchases

What it takes to build any of them

The four pieces are always the same, and three of them are not AI:

A file you can export. Almost every PMR system exports CSV. If yours does not, that is the first problem to solve, and no model solves it.
Arithmetic you wrote yourself. The average, the cover, the margin. Ten lines of spreadsheet you can check.
A threshold you decide, not the model. How many days of cover, what margin is too thin, how many units count as dead stock.
And only then the AI, to explain the result, rank it by what matters and write it into something readable in a minute.

The ordering agent lesson is exactly this, done end to end on one case, with its simulator. If you are going to build any of the other sixteen, start there: the skeleton repeats.

And the rule for not wasting time. Before automating something, ask how many times a month you do it and how long it takes. Below an hour a month, automating costs more than doing it — and that sum is done before you start, not after.

Where to start: the three axes that decide it

Seventeen areas is a list, not a plan. And picking the wrong first one is the most common way of abandoning this within three weeks — not because AI fails, but because people start with what impresses rather than with what hurts. Every area scores on three axes:

Axis	The question	How to measure it without guessing
Hours	How much of your time does it eat per month?	Time it for a week. You will underestimate: things always take longer than you remember
Risk	What happens if it gets it wrong?	Green: you see it and fix it. Amber: it costs money. Red: it affects a patient
Data	Do you already have it somewhere?	If it has to be typed in, that work counts too — and often eats the entire saving

The first area you build is always the same combination: many hours, green risk, and data you already have. Not the most ambitious, not the one that demos best. And there is a deeper reason: the first one is not a project, it is the learning. When the second takes half the time, you will know the first did its job.

The axis most people skip is the third, and it is the one that kills the most projects. "Profitability analysis by category" sounds excellent, and then it turns out the data lives in three places, two of them on paper. Before deciding anything, check that you can export a CSV of what is needed. If you cannot, that is not your first area no matter how well it scores on the other two.

The three questions before automating anything

Are the steps always the same?

If they are, you do not need AI: you need a formula or a script. Cheaper, faster, always the same answer, and testable. Putting a model where there was a rule is paying for uncertainty.
Is anybody going to review the result?

If nobody is going to look at it, then that result has to be safely irreversible — meaning it cannot touch anything. And if it is going to touch something, somebody has to approve it. There is no third option, and skipping it is the origin of nearly every scare.
What happens on the day it fails?

Not if: when. If the answer is "we do it by hand that day", go ahead. If it is "we grind to a halt", you need the manual route written down before switching anything on.

And a fourth question that is not optional in a pharmacy: does this touch patient data? If the answer is yes, it stops being a productivity project and becomes something else — with retention periods, access control and documentation. Not that it cannot be done: it is that it is not where you start, and hardly any of the seventeen areas that genuinely save hours need it.

The data you already have and are not using

Nearly every area on the list feeds on things your dispensing system already stores. The bottleneck is almost never the model: it is that nobody has ever exported that file.

What you already have	What it unlocks
Sales by product and month	Ordering, turnover, seasonality, what to stop buying
Stock and minimums	Ordering and stockouts
Expiry dates	What will expire unsold — money, directly
Hours worked per person	Rotas and keeping within contracted hours
Transactions by hour of day	When you genuinely need staff, which is almost never when you think
What people ask you at the counter	Nobody records it, and it is the most valuable data of all: it is what content to write and what service to offer

That last row is a zero-cost task almost nobody does. A notebook on the counter for two weeks, writing down the question that keeps repeating. That tells you which pages to write, what to put in the chatbot and which service is missing — and there is no tool, with AI or without, that can give you that data if nobody ever wrote it down.

What an owner should NOT delegate

Not delegated	Why
Anything going to a patient with your name on it	Responsibility does not transfer. Not professionally and not in practice
Deciding what to build	It is the one part that requires knowing your pharmacy, and it is in no model
Reviewing anything that touches money	Because the mistake is expensive and gives no warning
Knowing why a figure is that figure	An owner who does not know where a number comes from cannot defend it to anybody — or argue it with their accountant

The second row is the most misread. Asking a model for ideas is fine and usually produces useful ones; what it cannot do is set the priority. A model does not know there are three care homes in your neighbourhood, that your technician leaves in June, or that you have spent two years fighting with a supplier. Everything that makes your pharmacy yours is outside the model's context, and it is precisely what decides what is worth doing.

A 90-day plan that actually finishes

Weeks	What	How you know it is going well
1-2	Pick ONE area and time what it costs today	You have a measured hours-per-month number
3-4	Export the data and look at it. No AI yet	You know whether the data is usable — which is where half of these die
5-8	Build it, dry: it produces the result and you compare it with what you do by hand	The differences shrink every week
9-12	Use it for real, noting what you correct	You stop correcting. At that point it is a tool

What does NOT fit in those 90 days: a second area. The temptation to start the next one while the first is half done is enormous and is exactly how you end up with two half-things and neither in use. One finished beats two halves, always — and the second will take a third of the time precisely because the first is done.

And the measure of success is not "we have AI". It is that there is one concrete thing that used to take you three hours a month and now takes twenty minutes, and that you can name it without thinking. If after three months you cannot name it, the project has not worked — however well built whatever exists happens to be.

The seventeen, scored

With the three axes in front of you, the list stops being a list and becomes an order. This is not a universal recommendation — your pharmacy scores differently — but it is the honest starting point, and above all it is the format in which they should be scored:

Area	Hours/month	Risk	Data?
Drafting the order	Many	Amber	Yes
Expiry management	Medium	Green	Yes
Staff rotas	Many	Green	Yes
Opening hours and out-of-hours	Few	Amber	Yes
Website content	Many	Green	Not needed
Social media	Many	Green	Not needed
Counter questions (chatbot)	Medium	Amber	Partial
Sales and turnover analysis	Medium	Green	Yes
Profitability by category	Few	Green	Usually not
Payroll and employment terms	Medium	Amber	Yes
Contracts and HR paperwork	Few	Amber	Yes
Team training	Medium	Green	Not needed
Suppliers and comparing terms	Few	Green	Partial
Pharmaceutical care (support, not decisions)	Medium	Red	Partial
Screening and campaigns	Few	Red	Partial
Patient-facing reports	Medium	Red	Yes
Regulation and alerts	Few	Amber	Not needed

Read it looking for the "many hours + green + data yes" row, which is where you start. Two come out: staff rotas and website content. Not ordering — which is amber because it costs money — and nothing clinical, which is red by definition. What most people would want to build first is exactly what should not be first.

And notice the three red rows: all three are clinical. They are not there to be dismissed — they are legitimate areas and they are built on this very site — but to make clear that there AI supports and does not decide, always with a pharmacist present. Scoring them red is not fear: it is recognising that the cost of an error there is not measured in hours.

What it really costs, and in what currency

Everybody asks about the model cost, which is the irrelevant part. These are the four currencies this is actually paid in:

Currency	How much	Who pays it
The model	Pennies a month for nearly everything	Nobody notices
Your time building it	From two evenings to three weekends per area	You, and it is the real cost
The time spent reviewing	For the first months, almost as much as it saves	You again — and it only falls if you note what you correct
The cost of an error	Depends on the risk axis	The pharmacy. Which is why you start with green

The third row is what makes people abandon projects and the one nobody tells you about. In the first weeks, reviewing what the tool produces costs almost as much as doing it by hand, and it feels like it is not working. That is normal and temporary: what decides whether it falls is not time, it is noting every correction and fixing the missing rule. Without that notebook, reviewing never gets cheaper and the project dies of boredom.

The four questions that separate a serious offer from a brochure

You are going to be offered this. Under another name, with a very good demo and a monthly price. These four questions require knowing nothing about technology and separate a serious offer from one that is not:

The question	What you are listening for
"Where is the data stored and who can see it?"	A specific answer. "In the cloud, it's secure" is not an answer
"If I leave, do I take my data with me?"	If there is no export, what you are buying is a cage
"Show me how it gets things wrong"	The best of the four. Somebody who knows their product knows where it fails and will show you. Somebody who says it does not fail has not used it
"Who answers if this gets something wrong with a patient?"	You will. What matters is that they say so plainly

The third gives you the most information and is the least asked. Every AI product gets things wrong; it is a property of the tool, not a manufacturing defect. Somebody who shows you their bad cases and what they put in place so those cases do no harm is showing you they have thought about this. Somebody who answers that "the model is very good" is showing you the demo, which is a different thing.

And none of the four is about technology. They are the same questions you would ask any supplier of anything that touches your pharmacy: where is my stuff, how do I leave, what happens when it fails and who answers. The product having AI inside changes not one of them.

Before calling this learned

I know arithmetic goes underneath and the AI on top, in all seventeen.
I start with what I already have as data, not with the flashiest.
I know which data each area needs.
I work out whether it pays before automating, not after.

← The ordering agent, step by step
← Back to the AI School

AI for pharmacy owners: the full map