AI Receptionist for a GP Surgery: The 8am Rush and Hand-Offs

Yes, an AI receptionist can answer and sort some calls during a GP surgery's 8am rush, provided it has enough concurrent capacity and a tested route to staff. It should handle routine information and approved administrative bookings. Symptom requests, possible emergencies, distress, safeguarding concerns and any uncertain call must reach a person; the AI must not clinically triage.

Start with the existing phone data, not a vendor demonstration. The relevant question is how many callers arrive together, how long they wait and what happens when a transfer fails. NHS England's modern general practice model aims to move away from a first-come, first-served 8am queue, using the same care-navigation process across telephone, online and walk-in requests. A phone agent must fit that process rather than invent a parallel route.

Follow me on Instagram@sagnikteaches

Count the calls that actually collide at 8am

Export four weeks of call records from the surgery's cloud phone system. Keep five-minute arrival counts from opening until the morning peak ends, rather than using only the day's total. Record abandoned calls, callback requests, time to answer, transfers, repeat calls and the number of staff answering. Remove patient identifiers from the working spreadsheet. The five-minute peak shows the concurrency the agent needs; the daily average conceals it.

Connect on LinkedInSagnik Bhattacharya

For an illustrative surgery, 52 calls arrive from 8:00 to 8:05, 31 from 8:05 to 8:10 and 19 from 8:10 to 8:15. If an agent occupies a line for three minutes on average, the first window's 52 calls represent roughly 156 call-minutes arriving in five minutes. One simultaneous agent channel can serve only five call-minutes in that window. The arithmetic does not prove that 32 channels are needed, because callers and calls spill into the next window. It does show why a single impressive demo call says nothing about the rush. Ask the supplier for a written concurrent-call limit and show what callers hear when it is reached.

Subscribe on YouTube@sagnikteaches

Separate requests before deciding what to automate. In an illustrative sample of 100 calls, perhaps 28 ask about appointments, 22 seek test results, 18 concern prescriptions, 12 request letters, 10 ask about opening arrangements and 10 are unclear. Those are planning figures, not a claim about a typical surgery. A request for test results may require identity checks and clinical interpretation, so it is not automatically a simple information call. Have staff code a real sample and review the categories with the clinical lead.

Compare the agent with features already in the phone system: queue position, callback and real-time call data. NHS England's model names these cloud-telephony features as the way to manage peaks in demand. An AI answer may improve the first contact but still leave the same team facing all the work at 8:20. If the aim is access, measure time to a useful response and resolution, not just the proportion of calls an automated voice picks up. The wider AI receptionist booking explanation covers basic agent behaviour; surgery access needs stricter boundaries.

Make a map of permitted and prohibited work

Write an approval matrix before connecting a phone number. Put each call type into one of three lanes: answer from approved practice information, capture for a named administrative queue, or transfer to a person. The first lane may cover opening times and how to submit a routine request. The second may cover a repeat-prescription request that staff will verify, if the surgery's process permits it. The third covers symptoms, unexpected results, medication problems, urgent requests and uncertainty. No lane should invite the AI to decide severity or to reassure a patient that a symptom can wait.

As an illustrative script rule, a caller saying, “I need to speak to a doctor because my new tablets make me feel faint,” goes to the clinical hand-off queue, even if the call began as a repeat-prescription enquiry. The agent should acknowledge the request, avoid clinical advice and transfer. A caller saying, “What time do you close today?” may receive the approved opening information. A caller saying, “Can I get the result of my scan?” should not receive an invented or unverified explanation. These examples expose the flaw in a menu that categorises by the caller's first sentence only.

Agree what “transfer” means. A live transfer rings a staffed destination while the caller remains connected; a task or message is an asynchronous hand-off. Label both honestly. A symptom caller must not hear “you are connected to the clinical team” when the product has merely saved a transcript for review. Define which destination answers during the peak, who watches failed transfers and what the agent says if every person is busy. Preserve the practice's established urgent-care instructions, approved by its clinical lead, without having the AI create advice.

A useful test card for each lane contains the caller's words, allowed response, destination, maximum waiting behaviour, record created and owner. For example: “appointment for a new rash” → capture minimal details and live transfer to the care-navigation team; if unavailable, join the existing callback queue according to the practice's rule; do not offer a slot or assess the rash. Another card: “change my address” → identity-safe administrative route; do not read out the address on file. The card is short enough that reception can use it when checking transcripts.

Do not equate total triage with automation making clinical decisions. NHS England's guidance distinguishes gathering structured information at first contact from assessment and prioritisation. The agent can ask which kind of help is requested and collect a callback number. The practice must decide which clinical information may be captured, who sees it and how quickly. For fuller clinical governance on a different technology, see the NHS practice scribe assurance checklist; a scribe and a phone receptionist need separate reviews.

Build the hand-off before the greeting

First nominate a reception owner and clinical owner. The reception owner keeps the service directory and escalation schedule current. The clinical owner signs off wording for clinical requests and failure states. Then configure the destinations: live reception, care-navigation queue, administrative task queue and the existing out-of-hours message. Use the practice's phone and appointment systems as the source of truth. An AI's general training should never supply practice-specific opening times or available slots.

For every transfer destination, test four states: answered quickly, busy, unanswered and unavailable because the phone system or internet connection has failed. A voice agent that can transfer during a quiet Tuesday demonstration may fail when 20 callers ask for people at once. Run a supervised load test using test callers and no real patient data. Record whether calls are cut off, whether the original request remains visible and whether staff can call back. If the vendor cannot describe concurrency, queue behaviour and incident support, keep the pilot off the main number.

Use plain consent and identity wording. An illustrative opening is: “You are speaking with an automated receptionist. I can help with practice information or pass your request to the team. You can ask for a person at any time.” The practice should agree the exact wording with its privacy and clinical leads. For a booking or record-linked task, identify the caller through the practice's approved process; do not assume that a familiar phone number proves identity. Do not invite a caller to recite a long medical history to an unverified service.

Ask the supplier how calls, recordings, transcripts and summaries are stored, who can see them, how they are deleted and whether they are used to train models. A data protection impact assessment, or DPIA, is the documented assessment of privacy risks and controls. The ICO's AI guidance, which it says is under review after recent changes to data-protection law, states that most uses of AI involve processing likely to be high risk, which legally requires a DPIA, and that the judgement is made case by case. The surgery's information-governance lead should determine the requirement and approve any processor agreement before live calls. The scribe vendor security questions provide a useful starting list, but adapt it to call recording, routing and telephone subcontractors.

Imagine a caller asks for a human in another language or has difficulty hearing the agent. The correct failure mode is a prompt human route, not repeated attempts to force a scripted response. Test speech recognition with accents, quiet speakers, background noise, pauses and callers using an interpreter. A phone-only pilot also needs an accessible route for people who cannot use the agent. NHS England explicitly keeps telephone and face-to-face access alongside digital routes; the AI should not be a gatekeeper to either.

Ask the named GP phone agents the same written questions

Few products are built for surgery phones, and each describes its safety boundary differently. QuantumLoopAI markets EMMA as an AI receptionist built for NHS GP surgeries. Its site says EMMA handles hundreds of calls simultaneously, integrates with consultation tools such as Accurx and does not provide clinical advice or triage: urgent medical enquiries are directed to emergency and urgent-care numbers while EMMA captures details for follow-up. It also says the product meets DTAC, the NHS Digital Technology Assessment Criteria. Treat each statement as a claim to evidence, and ignore the headline savings figures on the page until your own trial produces numbers.

Accurx, whose online-request tools many surgeries already use for total triage, lists an AI phone assistant called Accurx Bell among the features coming to its total-triage toolkit. The page says it will handle routine inbound appointment requests and send a structured summary to the practice's triage inbox, and it invites practices to register interest in piloting. That design keeps the phone request inside the existing care-navigation queue instead of creating a parallel route, which is the property this tutorial argues for. Confirm what the pilot version actually does before planning around it.

Send every shortlisted supplier the same five questions and ask for written answers. What is the concurrent-call limit, and what do callers hear when it is reached? Which call types does the agent complete, capture or transfer, and can symptom-related calls be forced into the transfer lane? How does a failed transfer show up, and to whom? Where are recordings, transcripts and summaries held, for how long, and are they used to train models? Which clinical safety and information-governance documents will you hand over before we sign? A supplier that answers the last question with a badge on its website has not answered it.

Clinical safety has its own paperwork. DCB0129 is the NHS clinical risk management standard for manufacturers of health IT systems; DCB0160 is the matching standard for organisations that deploy and use them. Ask the supplier for its DCB0129 documentation, including the clinical safety case report, then complete the surgery's own DCB0160 documentation for this call flow: a safety case, hazard log and monitoring plan that cover the hand-off rules and failure states, signed off by the clinical lead. Keep it beside the DPIA. One addresses patient harm, the other privacy risk, and a phone agent needs both before it takes live calls.

Run a controlled morning trial with a stop rule

Begin with an administrative or overflow slice: for example, send only overflow calls to the agent from 8:00 to 8:30 for five working days, with reception able to disable forwarding immediately. Keep an agreed maximum number of concurrent calls and a live staff monitor. The exact threshold must come from the phone data and supplier capacity, not from a generic percentage. If the practice cannot monitor clinical transfers during those minutes, choose a quieter period for the first live test.

Make a before-and-after sheet with comparable mornings. Columns should include incoming calls, calls answered by a person, AI-handled calls, abandoned calls, callbacks completed, transfer failures, incorrect bookings, repeat calls within a day and the time a clinical request spent awaiting human review. Record the day, staffing level and any phone outage. A reduction in abandonment is useful, but a rise in repeat calls or delayed clinical messages may cancel the apparent gain.

Here is an illustrative calculation. A surgery receives 80 peak-window calls; 24 are sent to the agent. Of those, nine receive approved information, six leave administrative requests, seven transfer to people and two disconnect before an outcome. The agent has answered 24 calls, but only 15 were resolved or captured in the permitted lanes. The seven transfers still consume staff time. The two disconnects need review against call recordings and any repeat calls. Report the denominator and the actual outcomes, rather than “30% of calls automated”.

Set a stop rule before the trial. For example, disable the agent immediately if a clinical request is falsely marked complete, an emergency phrase is not transferred under the approved rule, a transfer fails without staff visibility or the wrong patient's information appears in a call record. Those thresholds are illustrative governance choices, not clinical advice. Give the duty manager authority and a one-step switchback to the established phone queue. Log the event, preserve evidence safely and ask the clinical lead to decide when testing can resume.

Two quieter examples matter too. A caller asks whether a routine letter is ready: the AI should not disclose status without the practice's identity process, but may create a callback task. Another asks to cancel tomorrow's appointment: the agent can do so only if the booking integration writes to the correct record, confirms the change and shows it to staff. If it merely says “cancelled” after making a note, that is a failed booking. See the booking-error recovery steps for a correction log that does not rely on a caller complaining.

Decide whether the call agent earned a wider role

At the end of the trial, ask reception to review a sample from each lane, not just the vendor's successful examples. Include the first calls after opening, the longest calls, every clinical transfer, every caller request for a person and every hang-up where the caller rang again. The clinical owner reviews the symptom-related sample. Score whether the agent captured the right words, chose the approved destination, accurately described what would happen next and left a usable record.

Keep the outcome categories strict: resolved with approved information; administrative request accurately captured; successfully handed to a person; unclear or incomplete; unsafe. A call that the agent answered and then stranded in a queue belongs in incomplete. A call that provided unsupported advice belongs in unsafe even if the caller sounded satisfied. Compare these results with a human answering service if the issue is overflow rather than automating information, using the answering-service comparison as a cost framework.

A second illustrative review may find that the agent answered opening-hours questions correctly but misheard two medication names. The answer is not to add drug vocabulary and let it discuss medicines. Remove that topic from automated handling and route medication calls directly. Another review may find that the transfer destination went to voicemail during staff breaks. Fix staffing and routing before changing the voice script. The operational bottleneck often sits after the AI's answer.

Ask the supplier to show changes in a controlled test environment and to document who can edit prompts, call flows and knowledge. A tiny wording change can move a symptom call into the wrong lane. Require a version log and repeat the high-risk test cards after each update, including model updates. Give staff a simple way to report a suspect call and tell callers how to reach a person. These controls matter more than a polished voice.

Use a short acceptance sheet that makes disagreement visible. For each of 30 test calls, write the expected lane before placing the call. Give one point for an accurate explanation of the next step, one for a complete and correctly identified task, one for the correct destination and one for a successful human connection where required. Mark any unsupported clinical assessment as an automatic failure regardless of its point total. For example, an illustrative “I need my blood result and I feel worse” call may capture the right callback number but still fail if it offers a routine result callback instead of the approved clinical hand-off. Keep the transcript and the reviewer’s reason with access limited to the practice team.

Test what happens if two people describe the same request differently. One caller says “I need another prescription”; another says “I have run out of my tablets and feel unwell”. The first may be an administrative request under a practice-approved process; the second needs a person. The point is not to teach the model medicine. It is to test whether a broad prescription category swallows language that should trigger a hand-off. Ask a receptionist to write the expected route without seeing the AI result, then compare. That avoids adjusting the answer after the fact to make the vendor appear correct.

Plan for changes in the roster. The destination for a clinical hand-off at 8:05 may differ from the one at lunchtime or on a day with fewer staff. Put a named owner in charge of that schedule and make the agent fail closed: when the approved person is unavailable, it should use the practice’s established fallback, not improvise a new contact or promise a response time it cannot control. In an illustrative shift change, a transfer goes to an unstaffed extension for 12 minutes. The log should identify the missing roster update and the caller’s eventual outcome, not simply count the attempted transfer as a success.

Finally, budget staff time for reviewing the pilot. If 24 calls reach the agent each morning and the team samples ten of them at roughly three minutes per review, that is 30 minutes a day before clinical review of flagged calls. This is an illustrative labour estimate, but it makes the pilot's hidden cost visible. The review burden may shrink after safe routing is established; it should not disappear while the software is making or receiving changes. Put that time beside the subscription quote and any telephone charges when deciding whether the agent is worth extending.

If the test meets the practice's safety and access criteria, expand one call category or time window at a time. Continue weekly sampling after launch because call mix, staffing and opening information change. If abandonment falls but clinical hand-offs are late, narrow the agent to information calls. If the queue itself is the problem, improve callback and total-triage workflow before purchasing more AI capacity. The purchase decision rests on documented patient access and safe completion, not the count of greetings delivered.

Questions before switching on the surgery phone agent

Can the AI allocate urgent appointments?

It should not determine clinical urgency or decide the appointment type from symptoms. It can collect the caller's request and contact details, mark a clinical hand-off, and connect or queue the caller according to a practice-approved rule. A clinician or trained care-navigation process decides priority and the suitable service.

What if the caller cannot use an online form?

Keep the telephone and in-person routes open. The phone agent may capture the same minimum request fields as the online form, but a caller who cannot use it or cannot communicate effectively with the agent needs a human route. Test that transfer with real staff before launch.

Should every 8am call reach the AI first?

Usually begin with overflow or a limited administrative queue. Compare the results with the existing cloud phone queue, callback and online request routes. Expand only after clinical hand-offs, message accuracy and accessibility meet the thresholds agreed by the practice's clinical and operational leads.

Further reads

• How to Write Call Scripts and Escalation Rules for an AI Receptionist — Build general transfer language before tailoring clinical rules.
• How to Set Up an AI Receptionist Without Losing Callers — Review the general phone and booking setup steps.
• AI Scribe for GPs: Costs, Free Tiers and the NHS Pilot Limit — Assess documentation tools separately from phone access.
• Medical Practice AI Budget: What to Fund First and What It Costs — Compare phone work with other practice investments.
• How to Measure an AI Receptionist's Return in the First 90 Days — Track whether improved answers create useful outcomes.
• AI Voice Agents for GP Practices: Costs and Safety Checks — How voice-agent pricing is built for a GP practice, with checked list prices and a worked sum, plus the safety checks before go-live.
• GP Surgery Phone System: Cloud Telephony With AI Call Handling — Cloud telephony against legacy lines for a GP surgery, where AI call handling fits, and the safety and privacy paperwork to finish before go-live.
• Healthcare AI Call Centre: When a Group Practice Should Centralise Calls — Five tests for a two-to-six site medical or dental group deciding between one central AI-assisted queue and a receptionist at each site.
• AI for GP Reception: Signposting Scripts and Role-Play Training — Build a bank of reception call scenarios from de-identified themes, then practise signposting and difficult calls with an AI caller.
• GP Practice DNA Policy: Wording, Reminders and Letters With AI — A did-not-attend policy in plain English: what to define, sensible thresholds, who is protected, reminder timing and letters AI can draft.
• GP Surgery Website: Plain-English Pages and Wording With AI — A page-by-page list for a GP surgery website, with prompts and samples for plain-English wording that keeps clinical advice off the page.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: NHS England, Modern general practice model; NHS England, Guidance on the use of AI-enabled ambient scribing products in health and care settings (for DCB0129 and DCB0160 documentation); DCB0160 clinical risk management standard (archived copy of 24 September 2026, live page behind a bot check); ICO, Guidance on AI and data protection (archived copy of 21 September 2026); QuantumLoopAI website; Accurx total triage page. Checked 4 November 2026.