Which Pharmacy Tasks Should Never Be Handed to AI?

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Which Pharmacy Tasks Should Never Be Handed to AI?
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for Which Pharmacy Tasks Should Never Be Handed to AI?

Never hand AI the clinical check of a prescription, the final check before a medicine is handed over, dosing or interaction advice for a specific patient, decisions on medicines sold under pharmacist supervision, controlled drug records, emergency supply decisions, safeguarding concerns or the handling of a dispensing error. Each needs a pharmacist's judgement and personal accountability.

It helps to be precise about what "AI" means here. Pharmacies already rely on validated software: patient medication records that flag interactions, barcode checks, dispensing robots. Those are tested systems with known rules. The risk this list addresses is general-purpose AI, the chat assistants and writing tools that produce fluent, plausible text and can be confidently wrong. They're useful for a lot of pharmacy admin. They shouldn't be anywhere near the decisions below.

Follow me on Instagram@sagnikteaches

A three-question test for any pharmacy task

Before any task goes near an AI tool, ask:

Connect on LinkedInSagnik Bhattacharya
  1. Could a wrong answer harm a patient? Not "probably won't", but could.
  2. Does it need judgement about this particular patient? Their history, their other medicines, what they told you at the counter.
  3. Do the law, your professional standards or your SOPs name a responsible person for it? If a named person must sign, check or record it, that person can't delegate the thinking to software.

A yes to any of the three means AI doesn't make the decision. It may still help around the edges, for example by drafting the SOP that describes how the task is done, but the task itself stays human. The ten tasks below all fail the test, most of them on all three questions.

Subscribe on YouTube@codingliquids

Clinical decisions that stay with the pharmacist

1. The clinical check of a prescription

Deciding whether a prescription is appropriate for this patient, at this dose, with these other medicines, is the core of the job. A chat assistant asked "is this dose right?" answers from general patterns, without the patient's record, and can miss the dangerous exception.

The classic example is methotrexate, which for many conditions is taken once a week. A prescription or label that says "daily" is a known, serious error. An assistant asked to "tidy up the directions" on a methotrexate label could produce clean, confident text that repeats the wrong frequency, or "corrects" a weekly dose to daily because daily is the more common pattern in its training. Fluent output is exactly what makes this dangerous.

2. Interaction and contraindication judgements for a patient

A patient on warfarin asks whether they can take an over-the-counter painkiller for a sore back. The answer depends on their record, their bleeding risk and what else they take, and it comes from a pharmacist using validated references, not a chat window. General AI tools sometimes invent interactions that don't exist and miss ones that do, and they present both with the same confidence.

3. Dosing advice, especially for children, kidneys and pregnancy

Weight-based children's doses, adjustments for reduced kidney function and medicines in pregnancy are where small errors have big consequences. A realistic failure: an assistant converts a milligram dose to millilitres using the wrong strength of liquid, producing a volume that's double what it should be. The arithmetic looks right; the input assumption is wrong. Doses for a named patient are worked out and checked by a pharmacist, every time.

4. Responding to symptoms and supervised sales

When someone describes symptoms at the counter, or asks for a medicine that can only be sold under a pharmacist's supervision, the value is in the questions: who's it for, how long, what else are they taking, any warning signs. Red flags, such as chest pain described as indigestion or a headache with a stiff neck, need a person to notice them. A chatbot on your website or a counter assistant checking an AI app instead of asking the pharmacist removes the safeguard the sales rules exist to provide.

Records and legal accountability

5. Controlled drug records and balance checks

Controlled drug registers must be accurate, made at the time, and completed by the authorised person. AI transcription of a handwritten entry, or an assistant "reconciling" a running balance from a spreadsheet, introduces exactly the errors the register exists to prevent. If a balance doesn't match, the investigation is a human one, and the record of it should be too.

6. Emergency supply decisions

Supplying a prescription medicine without a prescription, at a patient's request, involves conditions set by law and a professional judgement about need, previous prescribing and risk. An AI tool can't interview the patient, check the record with the same care, or carry the responsibility. The decision and the record of it stay with the pharmacist.

7. The final check before a medicine is handed over

The last check (right patient, right medicine, right strength, right directions, right counselling) is the barrier that catches everything earlier. Validated barcode scanning supports it; a general AI tool reading a label photo and saying "looks correct" does not replace it. If the final checker starts relying on an AI's "OK", the check has quietly stopped happening.

Conversations that depend on trust

8. Safeguarding concerns and suspected misuse

A regular customer buying codeine-containing painkillers every few days, a vulnerable adult who seems frightened of the person with them, a teenager asking for emergency contraception with an adult waiting outside. These need a trained person's judgement and your safeguarding procedure. Nothing about them should be typed into an AI tool, even anonymised, because the details that make a concern real are also the details that identify people.

9. Disclosing and investigating a dispensing error

When an error reaches a patient, the conversation (what happened, the apology, what happens next) must be honest, personal and led by the pharmacist. AI can help structure an anonymised incident report afterwards. It must not write the explanation to the patient or decide the root cause, because a plausible-sounding cause ("busy period, human error") can stop the team finding the real one.

10. Translating medicine instructions for a patient

It's tempting to paste directions into a free translation or chat tool for a patient who doesn't read your language well. The risk is real and well known: in Spanish, for instance, "once" means eleven, so an English "take once daily" left partly untranslated can be read as "take eleven daily". Machine translation of dosing instructions can also drop negations ("do not take with alcohol" becoming "take with alcohol"). Use professionally translated patient information where it exists, pictogram labels, and a spoken check with the patient, ideally with a qualified interpreter.

The grey zone: where AI helps if a pharmacist checks

Plenty of pharmacy work passes the three-question test once it's framed properly. The difference is that AI drafts or suggests, and a named person decides.

TaskWhat AI doesWhat stays human
Writing SOPsTurns your bullet points into a structured draftPharmacist checks every step against practice and law
Stock orderingSuggests quantities from sales and seasonalityBuyer approves orders; shortages handled by staff
Phone queueAnswers opening hours, "is my prescription ready?"Any clinical question goes to a person
Plain-English leafletsRewrites approved text at a lower reading levelPharmacist checks nothing changed in meaning
Staff rotasDrafts from availability and cover rulesManager confirms pharmacist cover is legal and safe
Counter-staff training quizzesWrites questions from your protocolPharmacist checks the answers are right

The SOP workflow in particular saves hours; writing pharmacy SOPs with AI shows it step by step. For the phone side, how independent pharmacies use AI to shorten phone queues covers where the automated answers stop.

The pharmacist's check in the grey zone isn't a formality, and three illustrative drafts show why. First, an SOP. The prompt: "Turn these bullet points into an SOP for handing out completed prescriptions: confirm identity with name and address; check bag against the prescription; counsel on new medicines; pharmacist decides on any query." The draft was clear and well laid out, and it included a step nobody wrote: "If the patient cannot attend, the prescription may be collected by any adult who can give the patient's name and address." That's the model filling a gap with something plausible, and it contradicts how many pharmacies handle collection by a representative. The pharmacist deleted it and wrote the pharmacy's actual rule, and the prompt for future SOPs now ends with "Add no steps that aren't in my bullet points; list anything you think is missing under QUESTIONS instead."

Second, a leaflet rewrite. The approved text: "Do not take more than 8 tablets in 24 hours. Leave at least 4 hours between doses." The plainer version the assistant offered: "You can take up to 8 tablets a day, about every 4 hours." It reads more easily and it's worse. "A day" is vaguer than "24 hours" to someone who takes a dose at 11pm, and "about every 4 hours" drops the words "at least", turning a minimum gap into a rough schedule. The pharmacist kept the original two sentences, which were already plain, and let the assistant simplify only the paragraph about side effects. Compare meaning line by line, not readability.

Third, the phone queue. A caller asks whether their prescription is ready, and an early version of the bot's reply read out the medicine's name. Anyone can ring and give a name, so that's a confidentiality leak, and for some medicines a sensitive one. The fixed reply says only "Your prescription is ready to collect", and any question about the medicine itself ("can I take it with my antibiotics?") gets "I'll pass that to the pharmacist, who'll call you back today" with no attempt at an answer.

One independent pharmacy's written line

Here is how an illustrative independent pharmacy, with one pharmacist, two dispensing technicians and three counter assistants, might put this into practice. After a month of trying AI tools on admin, the pharmacist writes a one-page rule, reads it through with the whole team and pins it in the dispensary. Filled in, it looks like this:

AI at our pharmacy: what we use it for, and what we never do
We use our business AI account for: drafting SOPs and staff notices; rewriting leaflets into plainer English (pharmacist checks before use); stock order suggestions (buyer approves); rota drafts; replying to non-clinical emails (opening hours, delivery days, collection times).
We never use any AI tool for: checking prescriptions; doses or interactions for any patient; deciding on supervised sales; controlled drug records; emergency supplies; anything about safeguarding; talking to a patient about an error; translating directions for a patient.
Patient details: never typed into any AI tool, including our business one.
If unsure: ask the pharmacist before using AI, not after.

With that line in place, the pharmacy might reasonably expect to recover four or five hours a week of admin: SOP reviews that took an afternoon now take an hour, the weekly order takes 30 minutes instead of 90, and the inbox is cleared before opening. The pharmacist's clinical time doesn't shrink, and that's the point. The cost is a business plan seat; ChatGPT Business, for example, is $25 per seat a month (or $20 billed annually) with a two-seat minimum. The general version of this reasoning for other sectors is in when not to use AI in your business.

When a patient arrives with a chatbot's answer

A newer problem: patients increasingly ask a chatbot first. Someone comes to the counter with a phone showing an AI's advice to "double your dose if the pain continues", or that two of their medicines "can't be taken together" when they can, with the right spacing. How the team responds matters for trust.

A sensible approach is to treat it like any other piece of information the patient brings: read it, don't dismiss it, then give your own professional view and the reason. Something like: "I can see why that's confusing. These tools give general answers without your record. For you, with the other tablets you take, here's what I'd recommend, and here's why." Counter staff should pass these straight to the pharmacist rather than trying to settle the disagreement themselves. And it's worth noticing which questions keep coming up, because a short, pharmacist-approved leaflet on the three most common ones can answer them before the patient searches.

Keeping the line from creeping

The risk isn't usually a deliberate decision to let AI check prescriptions. It's drift: a busy Saturday, a counter assistant asking a chatbot about a dose "just to double-check", a technician using an AI app to read a messy handwritten prescription. Three habits keep the line where you drew it:

  • Give staff a sanctioned tool for admin, so they don't reach for free apps on their own phones. Business plans don't train on your content by default, but the "no patient details" rule still applies.
  • Brief every new starter on the never list in their first week, with the reasons, not just the rules.
  • Ask at the monthly team meeting whether anyone used AI for something not on the approved list. Make it safe to say yes, because the answer tells you where the pressure is and whether the admin tools are covering enough.

Here's the kind of answer that meeting surfaces, in an illustrative case. A technician admits that on a busy Saturday she photographed a hard-to-read handwritten prescription and asked a free app on her phone what it said. The app read the strength as 10mg; the prescriber had written 5mg. The pharmacist's clinical check caught it because the dose didn't fit the patient's record, so no harm reached the patient, but two rules were broken at once: patient details went into an unapproved tool, and a dispensing step leaned on AI. The response that works is practical, not punitive: an unclear prescription is always a query to the prescriber, a standing note on the dispensary wall says so, and the pharmacist adds the incident to the anonymised near-miss log so the team can see the pattern rather than just the person.

Professional standards and pharmacy law vary, and change. When your regulator or professional body publishes guidance on AI, check your one-page rule against it, and take advice if something on your approved list looks borderline.

Further reads

Sources: published medication-safety alerts on accidental daily methotrexate dosing; research and reporting on errors in machine-translated prescription labels; vendor documentation for business AI plans (checked September 2026). Pharmacy law and professional standards vary; check your regulator's current guidance.

Want a clear AI line drawn for your pharmacy?

On a 1:1 call we'll list the tasks your team does each week, mark which ones AI can support and which it never touches, and set up the safe ones in tools you already have.

Book a 1:1 call with me