How Can AI Help GP Practices? Eight Admin Jobs Beyond the Scribe

A GP practice can use AI beyond the scribe for eight admin jobs: letter drafts, coding suggestions, patient messages, rota planning, complaint-response drafts, policy and audit documents, inbox sorting and call summaries. Each is safe only with a named person checking the output against the record or source before it leaves the building.

Clinical AI reviews dominate the search results, yet most of a practice manager's week is not clinical at all. It is letters, messages, rotas, complaints and paperwork for inspections. Those jobs are about language, structure and consistency, which is where current assistants help and where a mistake is usually caught by a reviewer who knows the practice.

Follow me on Instagram@sagnikteaches

Ground rules before any of the eight

Three rules apply to every job, and a practice that sets them once can use them for the lot.

Connect on LinkedInSagnik Bhattacharya

• Keep identifiable data out where a placeholder will do. Work from templates with [patient name], [date] and [condition] and fill them in your own system. Where a job genuinely needs record content, use a tool your data-protection lead has approved and that is covered by a suitable contract.
• A named reviewer per job. Write the role, not "the team".
• No clinical judgement from the assistant. It drafts and sorts; clinicians decide.

On tool choice, business plans such as Microsoft Copilot (Microsoft now uses that name for what was Microsoft 365 Copilot), ChatGPT Business, Claude Team and Gemini in Workspace do not train on business content by default, per the vendors' pages as of November 2026; Microsoft's privacy page says prompts and responses are not used to train its foundation models. That is a starting point for your data-protection assessment, not a substitute for one. If the practice already holds Microsoft 365 business licences, Copilot Chat (Basic) is included with eligible licences at no extra cost. Signed in with a work account, it carries Microsoft's enterprise data protection and works in Outlook and Teams on the message or chat you have open, or on a file you upload; it does not search your mailbox or shared files on its own. Check what you already have before buying anything.

Subscribe on YouTube@sagnikteaches

1. Letters: referral covers, recall letters and standard replies

Practices send the same kinds of letter repeatedly: recall invitations, medication review reminders, referral covering notes, replies to insurance or employer requests. An assistant can write a clean version from a template and a short list of facts. The clinician, or the manager for non-clinical letters, reads and signs.

The saving comes from consistency, not speed alone. A practice with five different people writing "we would like to invite you for a review" ends up with five tones. One saved prompt gives you one.

Draft a letter inviting a patient to a long-term-condition review. Use these facts only:
reason: annual review of [condition]; please book a 20-minute appointment;
call [number] or use [online booking route]; bring a list of current medicines.
Warm, plain English, under 130 words, reading age about 11. No clinical advice.
Leave [patient name] and [date] as placeholders.

An illustrative output starts "Dear [patient name], we would like to invite you for your annual review of [condition]" and ends with the booking details. Check step: the reviewer confirms the condition term is right for this patient group and that no advice has crept in. For anything clinical, such as a letter explaining a test result, the GP writes or edits every sentence.

2. Coding suggestions: a second pair of eyes, not an autopilot

Clinical coding turns a consultation into structured codes for the record and for reporting. Some systems and add-ons now suggest codes from note text. Used well, this is a prompt for the clinician or coder: "did you mean to code this as X?". Used badly, it silently adds wrong codes that distort the record and affect reporting and payments.

The safe pattern is suggestion-only: the tool proposes, the clinician accepts or rejects, and nothing is written to the record without a click. Before turning any suggestion feature on, ask the supplier three questions in writing: what data does it see, where is it processed, and can suggestions be switched to review-only? Then audit a sample.

A sample audit: say the practice reviews 50 suggested codes a month. It counts how many were accepted unchanged, edited or rejected. If more than one in five is wrong, switch the feature off until the supplier explains why. Check step: a clinician or trained coder accepts every code, and a monthly audit of a random sample sits in the practice's quality file.

3. Patient messages: templates for your most-sent replies

Online consultation and message systems create a steady stream of replies: "we have received your request", "please book a blood test", "your medication is ready", "we are unable to help with that online; please call". An assistant helps you write a library of good replies once, which staff then reuse.

Take the five most common reply types and write each with the assistant, then give each a code name in your message system. Say a practice sends about 60 of the "please book a test" message a week and each is currently retyped in two minutes. A saved template, adjusted in ten seconds, cuts 120 minutes to about 10, saving roughly 110 minutes a week. For a patient-facing chatbot, a different tutorial covers the scope and handover rules: AI chatbot for a doctor's website.

Check step: every template is approved once by a clinician and a manager, with a review date. Individual messages that mention symptoms or results go to a clinician, not a template. A realistic mistake is a template that says "no appointment needed" and is reused for a patient who does need one.

4. Rota and cover planning

Practice rotas mix GP sessions, nurse clinics, admin cover, leave and training days. An assistant can build a first draft from a table of rules and availability, answer "what if Dr B is absent on Thursday?", and write the email that asks for cover. It is poor at the final version, because long tables breed arithmetic slips.

Give the assistant the rules in one block: minimum number of GP sessions per day, nurse cover for each clinic type, reception cover at 8am and 6pm, training afternoons. Ask it to list any rule it cannot meet. Then count sessions per person in a spreadsheet. Check step: the practice manager totals sessions and checks every rule box by hand before publishing.

An illustrative failure: a draft that looks balanced but leaves Wednesday afternoon with one receptionist because the rule sat in an earlier message. The fix is a single "rules" document pasted into every request.

5. Complaint responses: structure and tone, never facts

Complaints demand a careful, empathetic reply within a time limit, and drafting one from a blank page is stressful. An assistant can give you a structure: acknowledge, summarise what the complainant said in their words, describe what the practice has checked, set out the outcome, explain how to take it further. The practice manager and the clinician involved supply every fact.

The rule is that the assistant never writes a fact about what happened. It writes the frame, and you fill the facts in from your own investigation. A sample instruction: "Draft a complaint response using only the points below. Where a point is missing, write [FACT NEEDED]. Do not speculate about causes, apologise for matters not listed, or admit liability." An illustrative output has clean paragraphs with gaps marked, which the manager fills.

Check step: the clinician involved and the manager both read the final text. Avoid pasting the complaint itself into a general assistant; summarise it with identifiers removed. Ask your indemnity provider or a qualified adviser if there is any doubt about wording that could affect liability.

6. Policies, audit packs and inspection preparation

This is the job with no patient data and the biggest hidden time cost. Infection control policies, chaperone policies, significant-event templates, fire-safety checks and staff induction packs all need writing and refreshing. An assistant can turn your rough notes into a clean policy, produce a checklist from a policy, and compare two versions to show what changed.

Say a practice has 28 policies and each needs an annual review. If a manager spends 90 minutes per policy, that is 42 hours. With drafted updates and a diff against last year's version, it might be nearer 45 minutes each, or 21 hours. Illustrative numbers again, but it is the kind of saving a practice can test on three policies before committing.

Check step: the policy lead reads the whole document against how the practice really works, because a policy that describes steps nobody takes is worse than none. Tie it to current guidance by reading the source yourself; do not rely on an assistant's memory of what a regulation says.

7. Inbox sorting and task summaries

Practice inboxes carry letters from hospitals, lab results, insurer requests and supplier mail. Document-management features that read and route incoming letters are appearing in some systems, and a general assistant can summarise long non-clinical correspondence into two lines and an action. Clinical letters and results belong in your clinical system's own workflow, with a clinician reading the source document.

The safe use is for non-clinical admin: a 12-page supplier contract, a long email chain about a building repair, a set of training-provider terms. Paste the text without personal data and ask for a summary with the three decisions needed and a date for each. Check step: read the original for any date, number or obligation, because summaries drop the awkward clause.

8. Call summaries and after-hours handling

Reception handles the heaviest load at 8am, when everyone calls at once. Call-handling AI can answer routine questions, take messages and pass urgent calls to a person, and call summaries can turn voicemails into a sorted list. This job is covered more fully in what an AI answering service costs a doctors' practice and after-hours call rules.

The design decision that matters is the urgent route. Chest pain, breathing difficulty, severe bleeding, suspected stroke and similar phrases should hand over to a person, or to the emergency instruction your practice uses, at once, and no AI should try to assess them. Check step: a weekly listen to a sample of twenty calls, counting any urgent call that was not escalated, with a target of zero.

A practice manager's first quarter, mapped in hours

Here is an illustration with invented numbers. Say a practice with eight clinicians and 12 admin staff spends about 73 hours a week on the seven jobs in the table below (inbox sorting is folded into letters here). The manager maps each job against two questions: how many hours, and how bad is an error?

The plan puts the internal, low-risk jobs first so the team learns the review habit on work where mistakes cost little. By month three the practice has a half-page AI-use note, a small library of approved templates and a monthly audit log. The estimated saving is a day or two of admin a week, not a headline number, and the real gain is that the practice manager stops writing the same letter forty times.

The blood-test message: one template, one afternoon

Before: the practice sends a "please book a blood test" message about 60 times a week, and four different people phrase it four ways. Some mention fasting, some do not, and patients phone to check.

After: the manager asks an assistant for a standard version with a short checklist (book online, whether fasting is needed, where to go, how long results take) and a clinician approves the wording once. The system holds it as a template. Calls about it fall because the answer is in the first message. Illustrative wording: "Hello [patient name], your clinician has asked for a blood test. Please book online or call [number]. [If fasting is needed: Please do not eat from midnight.] Results usually take [X] days." The bracketed parts are filled by the sender from the request. The check is that the fasting line is added only when the order says so.

One recall template stretched across every condition

Say a practice builds a recall template for a long-term-condition review and reuses it for every condition. The wording says "bring your home readings", which suits blood-pressure patients and confuses those with a different condition. Nothing is unsafe, but twenty patients phone to ask what readings, and reception loses an afternoon. The cause was a template approved once and applied too broadly. The fix is one template per condition group, each approved by the clinician who leads that area, and a review date on each.

A filled-in checklist for reviewing any AI draft

Reviewers work faster with a short list than with a vague instruction to "check it". Pin this beside the monitor:

1. Facts: every date, number, name and phone number matches the source in your system.
2. Additions: is there any sentence that is not in the facts you supplied, especially advice or reassurance?
3. Tone: would you be comfortable if this letter were read out in a complaint hearing?
4. Placeholders: no [square brackets] left, and none filled with the wrong patient's details.
5. Sign-off: the right person's name and role are on it.

Time one real batch. If reviewing a drafted letter takes longer than writing it, the prompt or the facts sheet needs work, and you should not scale the job up yet.

A sample prompt for a patient message, and what to fix in the output

Here is a prompt for the "unable to help online" reply, which practices send daily:

Write a short message for a GP practice's online request system. Situation: the request
describes a problem that needs an examination, so we cannot deal with it online.
Say we are sorry, ask them to call [number] after 8am or book at [link], and say that if
they feel very unwell or it is an emergency they should use the urgent route in [instruction].
Under 70 words. No diagnosis, no reassurance about the symptoms.

An illustrative output reads: "Thank you for your request. We are sorry, but we cannot deal with this online and need to see you in person. Please call [number] after 8am or book at [link]. If you feel very unwell or it is an emergency, [instruction]." The fix a reviewer makes: the draft sometimes adds "do not worry", which reassures about symptoms nobody has assessed. Delete it, and add the line to the prompt's "do not" list so it stops recurring.

What message templates and policy drafts are worth a year

Take two jobs from the table above. Message templates: 14 hours a week today. If templates cut that by a third, you save about 4.7 hours; at an admin cost of, say, $20 an hour fully loaded, that is about $93 a week or $4,850 a year, against a tool cost of $20 to $25 a seat a month for a few users on ChatGPT Business or Claude Team (both list $20 billed annually and $25 monthly, minimum two seats). Policies: 6 hours a week, saving perhaps half, so 3 hours a week and roughly $3,100 a year at the same rate. Neither calculation includes the value of fewer phone calls and calmer mornings, which is real but harder to count. Use your own hourly cost and your own logged hours rather than these figures.

The AI-use note: half a page that settles most arguments

Write it once. It lists approved tools, which jobs they may be used for, what data never goes in, who reviews each output and how often you audit. It also says that clinical decisions remain with clinicians and that any significant AI-related error is recorded as an incident. Share it with staff and revisit it every quarter. When a new tool or a keen member of staff proposes something new, the note gives you a quick answer: it is either inside the rules, or it needs a short assessment first.

If you want to compare scribe, phone and message tools as one decision, one suite or separate tools sets out the trade-off, and dictation or ambient notes covers the scribe question this tutorial deliberately leaves aside. If you would like someone to map your workload against these eight jobs, that is what my AI implementation consultation covers.

Questions practice managers ask

Can we use ChatGPT or Copilot with patient information?

Not casually. Patient information is special-category data, so use only a tool your data-protection lead has approved, covered by a contract that fits your obligations. For most admin jobs below you can avoid the question entirely by working from templates and placeholders, with the identifying details added in your clinical system afterwards.

Who is responsible if AI-drafted wording is wrong?

The person who sends it. An AI draft that a clinician or manager reviews and signs is their document, so the review step is a real control, not a formality. Write down who reviews each job and what they are checking, and keep that in your practice's AI-use note.

Where should a practice start?

With the job that has no patient data and a low cost of error: draft policies, rota first drafts or the structure of a complaint response. Move to patient-facing messages once your review habit is solid, and to anything touching coding or clinical content last, with the clinician checking every line.

Further reads

• Cliniko for Doctors: Fit, Cost and the AI Tools That Sit Around It — A practice-software example and the AI tools around it.
• What Is an AI Scribe and How Does It Work in a Consultation? — How a scribe works in a consultation room.
• Is an AI Scribe Worth It for a Small Private Clinic? — Whether a scribe pays for itself in a small clinic.
• How a Small Clinic Can Use AI at the Front Desk Safely — Front-desk rules for a small clinic.
• 10 Admin Tasks a Small Clinic Can Hand to AI This Month — Ten more clinic admin tasks and their checks.
• AI Scribe for GPs: Costs, Free Tiers and the NHS Pilot Limit — A GP scribe budget with free and paid options, hidden review time and the four-month pilot boundary.
• CQC and GP Practices: Build an Evidence Folder With AI — Organise a GP practice's CQC evidence by key question and quality statement, with an AI-built index, policy review log and mock-inspection summary.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Learn (What is Microsoft Copilot; Data, privacy and security for Microsoft Copilot), OpenAI ChatGPT business pricing page, Claude pricing page and Google Workspace generative AI privacy hub, all checked November 2026; check your own practice-software and coding-system documentation.