How to Use AI in a Community Pharmacy: Front Counter and Back Office

A community pharmacy can use AI for words, not decisions: drafting services posts, answers to routine counter questions, rota messages, staff briefings and SOP first drafts, all read by a person before use. Anything that identifies a patient or influences what is dispensed, checked or advised on a medicine stays out of the tool.

The line is easier to hold when you draw it before the first prompt. Think of two columns on the whiteboard in the dispensary. The left column holds jobs where a wrong sentence costs a few minutes to fix. The right column holds jobs where a wrong sentence could reach a patient's medicine. AI lives in the left column only. A quick test for a borderline job: if a wrong sentence would need a pharmacist to put it right, it belongs on the right.

Follow me on Instagram@sagnikteaches

The two-column rule on the dispensary whiteboard

Write it down and pin it up. The left column, "words about the business", covers opening hours, how a service works, what the shop sells, staff messages, policy documents and training notes. The right column, "anything about a patient or a medicine", covers who is taking what, interaction questions, dose queries, label wording, clinical checks, complaints involving a named person and any prescription image.

Connect on LinkedInSagnik Bhattacharya

The reason for a hard split is practical rather than moral. A general AI tool can sound equally confident whether it is right or wrong. A pharmacist can spot an odd sentence about a dosing interval because they know the answer; a counter assistant often cannot. And the moment a patient's name or medicine list enters a general chat window you have created a data-protection problem that a tidy policy cannot undo.

Subscribe on YouTube@sagnikteaches

Regulators have said much the same. One pharmacy regulator's position statement on AI, published in April 2026, says pharmacists, pharmacy technicians and pharmacies must meet its standards when AI is used, that professionals stay personally accountable and must be able to justify their decisions, and that owners and superintendents should carry out due diligence on any AI tool and be able to show it at inspection. It also says the regulator does not assess the clinical safety of AI tools, so that check falls to the owner. Read your own regulator's current guidance; it takes a few minutes.

A week in a pharmacy with AI doing the left column

Here is an illustrative independent pharmacy with a pharmacist owner, two part-time pharmacists, three counter staff and a delivery driver. The numbers below are invented to show the arithmetic, so replace them with your own timings from a one-week log.

On those figures the owner gets back about 225 minutes a week, nearly four hours. It is a saving on writing and admin only. No figure in the table involves dispensing time, and none should. The real test is whether each draft needed fewer corrections as you refined your prompts, so keep a tally of edits for two weeks.

Promoting services without making health claims

Services promotion is the easiest win because the content is about the shop, not a patient. A pharmacy might promote a flu vaccination service, a blood pressure check, a travel health consultation, a smoking cessation programme or a delivery round. The risk is claims: promising results, naming a medicine in a promotional post, or implying a service is available when it is not.

Give the AI the facts and the guardrails, and ask for options rather than a finished post.

Write three short Facebook posts for a community pharmacy announcing
our seasonal flu vaccination service.

Facts you may use:
- Walk-ins Tuesday to Thursday, 10am to 4pm; bookings other days.
- Takes about 10 minutes including the pharmacist's questions.
- Eligibility is checked by the pharmacist on the day.
- Cost: [price] for patients who do not qualify for a free vaccine.

Rules:
- Do not mention any brand of vaccine or any medicine.
- Do not promise protection, prevention or any health outcome.
- Do not say who is "eligible"; say "the pharmacist will confirm".
- Friendly, plain English, under 60 words each, British spelling.
- End each with how to book: [booking link].

A realistic output (illustrative) is three tidy posts, and the typical fix is in the second: it will often say something like "protect yourself this winter". That is a protection claim, so you edit it to "book your seasonal flu vaccination", which is a statement about the service. Keep a one-page "claims we do not make" list next to the prompt and paste it into every promotional request. Over a month, the list becomes your house style.

For seasonal campaign ideas beyond a single post, the tutorial on independent pharmacy marketing for flu and travel campaigns goes further. Here the point is narrower: AI speeds up the wording, while the owner decides what the shop can legitimately say and offer.

Counter queries: building an approved answer library

Counter staff answer the same non-clinical questions all day. "Are you open on the bank holiday?" "Do you deliver to my street?" "Can I collect for my mother?" "How do I order a repeat?" "Do I need an appointment for the blood pressure check?" These are policy questions with fixed answers, so write them once, have the pharmacist approve them, and let AI help you phrase them for the counter, the website, a text message and a phone script.

Ask for the same answer in four forms.

Below are our approved facts about repeat prescription collection.
Turn them into: (1) a 40-word counter script, (2) a text message under
160 characters, (3) a website FAQ answer, (4) a voicemail greeting line.
Do not add any fact that is not in the list. If something is missing,
write [CHECK] instead of guessing.

Facts:
- Collection is from 9am the working day after the request.
- Bring photo ID if collecting for someone else.
- We cannot give medicine details over the phone to anyone but the patient.

The "write [CHECK] instead of guessing" line matters. Without it, models tend to fill gaps with plausible policy, such as "we hold prescriptions for 28 days", and a wrong collection rule produces an upset patient at the counter the next morning. With it, you get a flag you can resolve. For an independent pharmacy thinking about automating the phone side, the tutorial on how independent pharmacies use AI to shorten phone queues covers when a voice agent is sensible and what it must hand to a human.

A second example from the counter: a patient asks whether the shop can post a medicine abroad. You do not ask AI. That is a regulatory and clinical question for the pharmacist. The test for sorting questions is a single sentence for the team: "If the answer would be different for a different patient, it is not for the library." Opening hours do not vary by patient. A question about a rash does.

Rota and staffing messages that people actually read

Back-office messages are the second easy win. A pharmacy rota involves pharmacist cover, which has legal weight in some places, so keep the cover decision with the owner. What AI can do is draft the message, the swap request and the reminder.

Suppose the Thursday pharmacist is off sick and a locum is booked for 9am to 1pm. The owner needs a message to the team, a note for the door and a text to the delivery driver. The prompt:

Rewrite these notes as three messages, in a calm, direct tone:
1) to the whole team (WhatsApp group), 2) to the delivery driver,
3) a printed notice for the door, in case of changed services.
Notes: Thursday, [name] is off. Locum pharmacist covers 9-1.
Afternoon: [name 2] is in from 1. Flu service closed Thursday morning,
bookings moved to Friday. Driver starts the round at 2pm, not 11am.
Do not name the reason for the absence.

The last line is a small privacy habit worth keeping: staff health information does not belong in a group message, and the instruction stops the model "helpfully" adding it. Check names and times against the rota before sending. A common slip is a 12-hour mix-up, such as "1pm" turning into "1am" in a rewritten sentence. It is rare, but at 8am on a busy morning you will not catch it by eye unless you check.

One more rota example: holiday cover over the festive weeks. Give the AI the dates the shop is open, the reduced hours and the nearest alternatives you have already agreed, and ask for a window poster, a text template and a line for the phone greeting, all saying the same thing. Consistency across those three is where small shops slip, because each is written at a different moment by a different person. Check every date and every hour against the poster you actually printed, and have a second member of staff read the final version aloud before it goes up.

SOP first drafts: from your steps to a clean document

Standard operating procedures are the heaviest piece of admin in a pharmacy, and AI is good at structuring them. It is not good at knowing your legal requirements, so the superintendent or owner reviews every draft against the real rules. The workflow in writing pharmacy SOPs with AI goes through the whole process; here is the short version for one procedure.

Say the shop wants a written procedure for a delivery that cannot be handed over because nobody is home. The owner dictates the steps they follow in plain words: the driver calls the number, waits five minutes, tries once more on the way back, never leaves medicines with a neighbour, and returns the bag to the shop. The AI turns that into a numbered SOP with purpose, scope, responsibilities, steps, records and review date.

Turn my notes into a pharmacy SOP with these headings: Purpose,
Scope, Responsibilities, Procedure (numbered), Records kept, Review
date. Keep my steps exactly; do not add steps or legal requirements.
If something looks missing, list it under "Questions for the owner"
at the end instead of adding it to the procedure.

The output is a clean document plus a short list of questions, perhaps "What is the record kept when a delivery returns?" and "Who is told if the same patient is missed twice?". Those questions are valuable, because they are the gaps in your process. You answer them, then redraft. Then comes the human review: read it with the dispensary team, change what the real day looks like, date it and file it. If the AI added a legal requirement you did not write, delete it, because you cannot verify it from a chat window.

Staff briefings and meeting notes

Two more back-office uses sit comfortably in the left column. First, short briefings: a new seasonal service, a changed supplier, a new till procedure. Paste your own bullet points and ask for a one-page note a counter assistant can read in two minutes. Second, meeting notes: type or paste your own rough notes (no recordings of conversations about patients) and ask for actions with owners and dates.

A before and after shows the difference. Before: "ordering. new wholesaler Monday. check delivery note. tote labels. ask [name] re returns." After: a three-item list with a named owner and a due date for each, and a line saying which items are still undecided. Leave undecided items undecided; models like to resolve them for you, so add "do not invent decisions" to the prompt.

Replying to online reviews without naming anyone

Reviews are a left-column job with one trap. A patient who writes a review has not waived their confidentiality, so the reply must never confirm that they are a patient, what they collected or when. AI is useful here because it writes calm replies when a person is irritated, and it can be told the rules once.

Draft a reply to this public review of our pharmacy. Rules: do not
confirm the reviewer is a patient or customer; do not mention any
medicine, service used, date or staff name; thank them, say we take
feedback seriously, give the shop's general contact route, and keep
it under 70 words. Review: "[paste review text]"

Take an illustrative one-star review complaining about a long wait on a Saturday. A weak first draft says "We are sorry your prescription took so long on Saturday". It confirms a prescription and a day, so you cut it to "We are sorry the wait was longer than it should have been. Please contact us at [shop contact] so we can look into it." The fix takes ten seconds and is the pattern for every reply: apologise for the experience, never describe the visit.

Put the rule beside the prompt in your shared Project and add a second one for positive reviews, where the temptation runs the other way. A reply that says "great to see you again for your regular collection" is just as revealing as a complaint reply.

What never touches dispensing decisions

Keep this list pinned beside the whiteboard. It is deliberately blunt.

• No patient names, dates of birth, addresses, insurance or health-service numbers, or prescription images in a general AI tool.
• No question about a particular medicine, dose, interaction, contraindication or symptom put to a general chatbot to decide what happens next.
• No AI-generated label wording, counselling script for a specific patient or clinical note filed without pharmacist authorship.
• No AI check of a prescription, no "second pair of eyes" on a dispensing error, no AI input into who gets a service.
• No sending a draft to a patient that nobody has read.

The tutorial on which pharmacy tasks should never be handed to AI expands the list with reasons. One thing worth stating: a mistake made with AI is still your mistake. A pharmacist cannot defend a wrong answer by pointing at a chatbot, and the regulator's statement quoted above says so.

Choosing the tool and the plan

For left-column work you do not need a pharmacy-specific product. A general assistant on a paid business plan is enough, and the key is how it treats your content. On their pricing pages in November 2026, ChatGPT Business and Claude Team both say business content is not used for training by default, while on consumer plans training is controlled by a privacy setting ("Improve the model for everyone" in ChatGPT, "Model Improvement" in Claude) that you should check and switch off. Both business plans start at two seats, so a one-owner shop pays for two or uses an individual plan. Both also offer shared Projects, which is where your claims list, approved answers and tone notes can live.

Do not treat a business plan as permission to paste patient data. Whether a vendor will sign a business associate agreement or a data-processing contract, and which of its products that covers, is a separate question that depends on your rules and the plan. OpenAI also offers ChatGPT for Clinicians, free at launch, to verified pharmacists, but only in one country so far; it is a different product with its own terms and an individual BAA; the tutorial on AI tools for pharmacists and what the rules allow covers where it fits.

A two-week pilot with a pass mark

Pick one task from the table and run it for two weeks. Day one, write the prompt and the "claims we do not make" or "facts" block. Days two to ten, use it for real work, and keep a simple tally: minutes taken, number of edits, any error that nearly went out. Days eleven to fourteen, decide.

A sensible pass mark: at least 30% of the time saved, fewer than three material edits per draft after the first week, and no near-miss that reached a patient. If you fail on the third, stop and tighten the prompt. If you fail on time saved, the task is too small to bother with. Then add the next task.

An illustrative pilot: the owner times the weekly services post. Week one, 25 minutes with the AI against 40 before, with five edits for tone and one claim removed. Week two, 15 minutes and two edits, because the claims list is now in the prompt. That is a pass, so the owner moves on to the rota message.

Five slips from a pharmacy's first month of AI drafting

• The helpful guess. A model fills a gap in your facts with something plausible, such as an opening time. Fix: tell it to write [CHECK] for anything missing.
• The creeping patient detail. A counter assistant pastes a real complaint to "make it sound kinder". Fix: use placeholders and rewrite the reply by hand with the real names in your own system.
• The invisible draft. A message goes out because it "looked fine". Fix: every outbound text has a named reader who ticks it off.
• The shared login. One account for the whole team means no one owns the output. Fix: individual seats and a shared Project.
• The stale SOP. A drafted procedure is filed and forgotten. Fix: a review date in the document and a diary entry.

Three numbers that show the drafting is paying off

Three measures are enough. Minutes spent on admin per week (from your log). Number of corrections per draft (from your tally). Staff confidence: ask the counter team once a month whether the library answers helped. If minutes fall and corrections fall while confidence rises, extend. If any patient-related slip occurs, pause and review the rule, not the person.

If you would like help choosing where to start, that is what my AI implementation consultation is for. Otherwise, copy the prompts above, pin the two-column rule, and begin with the weekly services post.

Questions pharmacy owners ask before starting

Can the counter team paste patient details into ChatGPT?

No, not on a standard plan. Anything that identifies a patient, a prescription or a medicine history stays out of general tools. Write prompts with placeholders such as [patient] and [medicine], and put the real detail into your own system afterwards.

Does AI replace the pharmacist's check on a query answer?

No. AI drafts wording for questions about opening hours, services, prices and how a service works. Anything that touches a specific medicine, dose or symptom goes to the pharmacist, and a draft is never sent unread.

How long does a first pilot take?

Most owners can run a two-week pilot on one task, such as the weekly services post or the rota message, in about an hour a day. Judge it on minutes saved and corrections needed, then add a second task.

Should we tell patients we use AI?

If a patient could reasonably believe a person wrote a message or recorded a call, say so plainly. For drafts a member of staff reads, edits and sends, a line in your privacy notice may be enough, but at least one pharmacy regulator now expects professionals to be transparent about their use of AI, so check your regulator's current guidance.

Further reads

• Pharmacy Automation Companies: A Buyer's Checklist Before You Sign — Questions to ask before signing with an automation vendor.
• AI Agents for Pharmacies: Phone, Chat or Refill Requests First? — Phone, chat or refill requests: which agent first.
• AI for Pharmacies: The First 90 Days for an Independent Owner — A longer plan once the first jobs work.
• Pharmacy Staff Training: Induction Packs for Counter Staff With AI — Induction packs for new counter staff.
• Pharmacy Near Miss Log: Finding Patterns With AI, Not Blame — Find patterns in near misses without blame.
• Does AI Stock Ordering Pay Off for an Independent Pharmacy? — Whether AI stock ordering pays off for a small shop.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help centre (ChatGPT for Clinicians), GPhC position statement on the use of AI in pharmacy (April 2026), ChatGPT Business and Claude Team pricing pages, OpenAI and Anthropic consumer privacy settings (checked November 2026).