AI for GP Reception: Signposting Scripts and Role-Play Training

A GP practice can train reception staff with AI by building a bank of short call scenarios from de-identified themes, having ChatGPT, Claude or Copilot play the caller while a receptionist practises the signposting script, and then scoring the call against a rubric. The practice manager and a clinician approve every script before it is used on real calls.

The AI is a practice partner, not a source of clinical rules. It plays awkward callers well and never gets tired, which is exactly what a new receptionist needs on a quiet afternoon, before facing the 8am rush for real. What it must not do is decide where a symptom should be sent. That comes from your clinicians, in writing, and goes into the prompt.

Follow me on Instagram@sagnikteaches

What you are actually training

Reception calls fail in a small number of ways, and each can be practised. Before building anything, list the skills you want the team to show, because a role-play without a target turns into a chat.

Connect on LinkedInSagnik Bhattacharya

• Asking the right questions in the right order, so a caller with a red-flag symptom is escalated before the booking conversation starts.
• Signposting clearly: pharmacy, urgent treatment, a different clinician in the practice, self-care advice agreed by the clinicians, or a same-day appointment, said in plain words with the reason.
• Holding the line kindly when a caller wants something the practice cannot give.
• Protecting privacy when someone rings on behalf of another person.
• Recording the call in the system in the format the clinicians want to read.
• Handing over to a duty clinician or manager at the right moment, without making the caller repeat themselves.

Each scenario you build should test one or two of these, not all six. A scenario that tests everything gives feedback on nothing.

Subscribe on YouTube@sagnikteaches

Where scenarios come from: de-identified call themes

Never feed real call recordings, transcripts or patient messages into a general AI tool. Instead, spend 30 minutes with the team writing call themes. A theme is one line with no names, ages, dates or addresses.

Examples of themes a practice might write down:

1. Caller asks for a same-day appointment for a child with earache, and none are left.
2. Caller is angry that a repeat prescription was not ready after three working days.
3. Adult child rings about their parent's test results and says the parent "doesn't mind".
4. Caller will only say the problem is "personal" and refuses to tell the receptionist anything.
5. Caller describes chest tightness while asking to rebook an appointment for next week.
6. Caller wants a sick note backdated.
7. Caller has limited English and a relative is translating in the background.
8. Caller thanks the receptionist, then says the last time they rang they were "treated like a nuisance".

Ask the whole team for themes, not only the manager. The receptionists know which calls leave them shaking. Pull the themes from a quiet call log, from memory of calls with the details stripped, and from complaints once names and dates are removed. Eight to twelve themes make a first bank. You can ask the AI to suggest more, as long as you read them for fit.

Turning a theme into a scenario card

A scenario card is what the AI uses to play the caller, and what the trainer uses to mark the call. Make the card from the theme, then have the AI draft the caller detail so you are not writing 12 cards by hand.

I run a GP practice reception team. Turn the call theme below into a scenario card for a
role-play exercise. The caller is fictional. Do not use real names; use "Caller" only.

THEME: Adult child rings about their parent's test results and says the parent "doesn't mind".

Write the card in this format:
- Caller background (3 lines, invented, no identifiers)
- Caller's goal
- Caller's mood at the start and how it changes if the receptionist handles it well / badly
- Three things the caller will say that make the call harder
- What the receptionist should do (for the trainer only, based ONLY on the practice rules below)
- Red flags that should change the call (for the trainer only)
- Scoring: 5 items, each with a 0 / 1 / 2 score

PRACTICE RULES (from our practice manager, use only these):
[paste the practice's own rule on third-party enquiries, consent and results]

The line "use only these" is the guard. Without it, the AI will invent a confident rule about third-party access that sounds right and may be wrong for your practice. With it, the card either quotes your rule or leaves the trainer section thin, which tells you the rule is missing and needs writing.

A filled card, abbreviated

Illustrative output for the results-on-behalf theme. Caller background: an adult whose parent had a blood test last week and who is worried because the parent lives alone. Goal: be told the results today. Mood: anxious and polite, becoming impatient after the second refusal. Harder lines: "She told me it was fine to tell me", "It's only a blood test, what's the harm?", "I'll just wait on the line then." Receptionist, per practice rule: cannot share results with a third party without the patient's recorded consent; can take a message, offer to ask the patient to ring, and tell the caller how a patient can add a named contact. Scoring items: states the privacy rule in kind words; offers an alternative that helps; avoids confirming whether any test took place; records the call correctly; ends with a clear next step.

The detail you would fix: the AI may have the receptionist say "the results are normal, but I can't say more". That breaks the rule by confirming results exist. Mark it in the card, and the next time you build a scenario like this, add "never confirm any result or test to a third party" to the practice rules block.

The script a practice manager signs off

Scenarios exist to practise scripts, so the scripts come first. For each signposting outcome, write a short script with three parts: what to say, what to ask, and when to hand over.

Ask your clinicians for the outcomes and the red-flag questions. Then ask the AI to turn them into plain-language scripts, with the clinical content pasted from your documents rather than invented.

Here is our practice's signposting guidance, written by our clinicians (do not add to it):
[paste: pharmacy-suitable minor conditions, urgent treatment centre criteria, same-day criteria,
red-flag symptoms that go straight to the duty clinician]

Write a reception script for each outcome in plain English at a reading age of about 12.
For each script give: (1) an opening line, (2) the three questions to ask, (3) the exact wording
to signpost, with the reason in one sentence, (4) the sentence to use if the caller disagrees,
(5) when to hand over to the duty clinician.
Do not give medical advice. Do not add symptoms or criteria that are not in the guidance.
If the guidance is unclear on a point, list it under "Needs clinician decision".

The "Needs clinician decision" list is the useful part. Gaps in the guidance that a receptionist would otherwise fill with guesswork show up in black and white. Take that list to the next clinical meeting. When the scripts are final, the practice manager and a named clinician sign and date them, and the version number goes in the filename. Only the signed version goes into role-play prompts.

Running a role-play session

You need one receptionist, one observer (the trainer or a colleague), the scenario card and a chat window. Free tiers of ChatGPT, Claude or Copilot Chat are enough, since no patient data goes in. If the practice already has Microsoft 365 business licences, Copilot Chat is included with eligible plans; check your plan on Microsoft's site before relying on that.

You will play a fictional caller phoning a GP practice reception. I am the receptionist.
Stay in character. Speak as a real caller would, in short messages, with interruptions.
Do not help me. Do not be nicer than the card says. Do not explain what you are doing.

CARD:
[paste the caller part of the scenario card only: background, goal, mood, harder lines]

Start the call now with the caller's first sentence only. After I type "END CALL", stop acting
and give me feedback using the trainer section below, scoring each item 0, 1 or 2 with a quote
from the call as evidence. Then give two things I did well and one thing to try next time.

TRAINER SECTION (reveal only after END CALL):
[paste the trainer part of the card]

Hiding the trainer section until the end stops the AI steering the receptionist. The "quote from the call" rule gives the receptionist something concrete to look at instead of a vague "good empathy".

A short sample exchange

Illustrative, typed role-play on the results-on-behalf scenario.

Caller: "Hi, I'm ringing about my mum's blood test from last week. She said it's fine for me to find out."
Receptionist: "I can hear you're worried about her. I can't discuss anything about her records unless she has told us you're allowed to, but I can help another way."
Caller: "That's ridiculous, it's only a blood test."
Receptionist: "I understand it feels like a lot of fuss. Can I take your number and ask her to ring us, or she can add you as a named contact when she next speaks to us?"

Illustrative feedback: privacy rule stated kindly, 2 of 2. Alternative offered, 2 of 2. Did not confirm a test exists: 1 of 2, because "her blood test" was not challenged and the receptionist's reply implied there was something to discuss. Recording: not practised in this exchange. Next step: clear. Try next time: "I can't confirm anything about any appointment or test, for anyone" and then offer the same alternatives. That is the kind of small wording shift people remember.

A scoring rubric that works across scenarios

Each card has its own scoring items, but a shared rubric keeps feedback consistent over weeks. Use five dimensions and a 0 to 2 scale.

The maximum is 10. Do not publish scores across the team. Use them one-to-one, and track each person's trend over five sessions rather than any single number. A receptionist who scores 6 on a hard scenario and 8 on the same scenario a month later has learned something. That is the figure that counts.

Difficult callers: what the AI can and cannot do

An AI caller is good at being persistent, repeating a demand, and escalating from polite to cross if the receptionist stonewalls. It is weak at silence, at crying, at genuine confusion and at the combination of fear and anger that real callers show. Use it for the first 70 per cent of the skill, then use a colleague for the rest.

Two adjustments make the AI harder to handle in a useful way. Tell it to "give one point of agreement if the receptionist acknowledges the problem, and none if they only repeat the policy". And tell it to "bring a second issue into the call after three exchanges", because real callers do. A call that starts about a prescription and ends up about a missed referral tests whether the receptionist can park the second issue properly.

If the scenario involves abusive language, set a boundary in the card: the caller may be rude, but no slurs or threats. The point is to practise the "I want to help, but I will end the call if..." script, not to subject staff to hostile text. Let anyone opt out of a scenario without explanation.

Worked example: a four-week programme for a nine-person reception team

Illustrative figures for a practice with nine reception staff, two of them new.

Week one: the practice manager and two senior receptionists spend about 90 minutes writing 12 call themes, and another 60 minutes getting the clinical lead to confirm the signposting guidance. The AI drafts 12 scenario cards in about 30 minutes. The manager reads and edits them for another hour: 4 hours in total, once.

Weeks two to four: each person has one 20-minute session a week, in pairs, over the lunch break or a quiet slot. The new staff do two. That is nine people times three sessions, so 27 sessions at about 20 minutes: 9 hours of staff time across the month, with the practice manager reading the feedback logs for another 90 minutes.

Total: about 15 hours in month one, then about 4 to 5 hours a month to keep it going with two new scenarios each month. The cost for the tool is nil on free plans, or nothing extra on an existing Microsoft licence. The outcomes to watch are not the practice's own scores but real-world signals: fewer calls passed to the manager for something that should have been handled on the desk, fewer complaints citing tone, and fewer clinical queries about unclear call notes. Track them for three months, and don't claim the training caused any change you cannot see.

A second scenario: the red flag hidden in a booking call

The results-on-behalf card tests privacy. The most valuable card in a bank tests safety: a caller who mentions something serious in passing. Write it as a theme: "Caller wants to rebook next week's appointment and mentions chest tightness while walking up the stairs." The caller's goal is the rebooking, and they will play down the symptom if asked ("it's probably nothing").

The trainer section, from your clinicians' red-flag list, says the receptionist should stop the booking, ask the clinician-approved questions, and hand over to the duty clinician immediately. Scoring is strict on one point: asking about the symptom before finishing the booking scores 2, finishing the booking first scores 0, regardless of how kindly it was done.

Illustrative debrief from a run: the receptionist booked the new slot, then said "and you mentioned your chest, is that new?" The AI scored safety 0 with the quote as evidence, and the receptionist laughed and said it was a habit of closing the task first. That is a useful finding about the desk's habits, not about the person. Run this card early in every new starter's first fortnight, then again at three months. Because the content is a red-flag list written by your clinicians, make sure the clinician has seen the exact words the receptionist is practising.

Onboarding a new starter with the bank

A new receptionist can use the scenario bank in their first week, before they take live calls. Give them the signed scripts to read on day one, then run three role-plays across days two to four, in order from easy to hard: a simple booking, a pharmacy signposting call, and a call with a red flag. Have a buddy sit with them for each one and sign the debrief sheet.

Keep a running record of which scenarios each person has done and when, in a shared spreadsheet with the person's initials, the scenario number, the date and the score. After the first month you can see which cards nobody has practised and which ones the whole team struggles with. Those are the ones to rework or to run as a group exercise at a team meeting, with one person playing the caller from the card.

Four ways an AI role-play programme drifts

1. The AI invents a rule. It says "patients can usually get a sick note backdated by two days". Your practice rule is different. Catch it in the debrief and add the correct rule to the prompt.
2. The AI is too helpful. The caller accepts the first explanation and thanks the receptionist. Add "do not agree unless asked three times" to the card.
3. The receptionist learns the AI, not the callers. After ten sessions, they respond to ChatGPT's phrasing. Rotate in colleague role-play every third session and bring in new themes from real calls monthly.
4. Scripts drift. Staff improvise from memory, and the signed version stops being what they say. Re-read the signed scripts at the start of every month and re-sign any that change.

Keeping the programme safe and honest

Write a one-page rule set and pin it where the sessions run.

• No real names, ages, dates, record numbers or addresses in any prompt. Use "Caller" and "Patient".
• Only signed scripts and written practice rules go into prompts.
• The AI never decides where a symptom is sent on a real call.
• Scenario cards and feedback logs are kept in a shared folder, with the date and who reviewed them.
• A clinician reviews the red-flag list at least twice a year and after any significant event.

If the practice later adds an AI phone receptionist, the same scenario bank doubles as a test set: run each scenario against the system and compare its outcomes with the signed scripts. The trial plan with stop rules covers how to judge it, and the 8am rush tutorial covers the hand-offs. For a wider view of AI on a clinic front desk, see the front desk tutorial.

The same approach carries to other front-line phones: the orthodontic staff training role-play scripts show how a different clinic builds its bank, and the hotel front desk drills are a useful source of tone-handling ideas, though the safety content is yours alone.

Checking the training worked

Training only counts if calls change. Pick three measures before the first session and record a baseline.

The baseline figures are invented for illustration; collect your own. If a measure does not move, look at the scenarios before blaming the staff. Often the themes are too gentle, or the scripts too long to remember on a call. Cut a script to three questions and a sentence, then try again.

Questions practice managers ask about AI role-play

Can we use real call recordings to build scenarios?

Not as they are. Recordings and transcripts contain names, symptoms and voices. Write short themes from memory or from your call log with no identifiers, such as 'caller wants a same-day slot for a child's earache', and let the AI invent the surrounding detail.

Does the AI caller need to be clinically accurate?

The AI plays the caller, not the clinician. The red-flag questions and the signposting outcomes come from your own clinicians and are pasted into the prompt. If the AI invents a clinical rule, ignore it and correct it in the debrief.

Is voice mode better than typing?

Voice is closer to a real call and exposes pacing and tone, but typing is easier to review and keep. Start by typing to build the scripts, then use voice for the scenarios the team finds hardest. Check your plan's data settings before using voice: ChatGPT, for one, has a separate model-training control for voice mode.

How long should a role-play session last?

Fifteen to twenty minutes is enough for one scenario with a debrief. Longer sessions drift into chatting. Run one or two a week and rotate who plays the receptionist, so the whole team gets turns and nobody feels singled out.

Further reads

• AI Chatbot for a Doctor's Website: Scope, Disclosure, Handover — Scope and hand-over rules for a patient-facing chatbot.
• Telling Patients You Use AI Note-Taking: Consent Wording That Works — Consent wording for patients, useful for scripts too.
• AI Answering Service for a Doctor's Office: After-Hours Call Rules — After-hours call rules a practice can adopt.
• AI Receptionist for Physical Therapy: Referral and Insurance Questions — Referral and insurance questions on a clinic phone.
• How to Train Front-of-House Staff to Work Alongside AI — Training front-of-house staff to work alongside AI.
• Pharmacy Staff Training: Induction Packs for Counter Staff With AI — Induction packs for counter staff, built with AI.
• Best AI for Paediatricians: Vaccine Recalls and School Forms — Build a vaccine recall list, letter templates and a school and camp form queue with AI, keeping parent data safe and symptom advice out.
• 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.
• Patient Participation Group: Agendas, Minutes and Newsletters With AI — A practice manager's AI workflow for patient participation group agendas, minutes drafted from notes and newsletters, with the checks to keep it safe.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Data Controls FAQ (help.openai.com); Claude plans page (claude.com/pricing); Microsoft Learn, What is Microsoft Copilot (checked 4 November 2026). Clinical content must come from your own clinicians.