GP Surgery Phone System: Cloud Telephony With AI Call Handling

Most GP surgeries should move to a cloud phone system built for primary care, then add AI call handling only as overflow for the busiest hour. The cloud system gives you queuing, callbacks, recording and call data; the AI takes simple, scripted calls. Finish the clinical safety and data-protection paperwork before any patient talks to it.

The order matters. Buying the AI first leaves it sitting on a phone system that cannot route calls to it, report on it or hand over cleanly. Fix the phone system's queue and callback behaviour, measure the 8am peak for a fortnight, then decide how much of that peak a machine could take safely.

Follow me on Instagram@sagnikteaches

Why the phone system is the real decision

A surgery's phone problem is mostly a 30-minute problem. Calls arrive in a spike after opening, the lines fill, patients hear engaged tones or wait, reception cannot book and answer at once, and the same patients call again. Everything you buy should be judged on what it does in that half hour.

Connect on LinkedInSagnik Bhattacharya

A second pressure is the analogue network. Openreach, which runs the network most phone providers use, gives 31 January 2027 as the date all traditional phone lines go digital. A practice still on analogue lines, or with alarms and card machines plugged into them, will need a digital route anyway, so the choice is often "which digital system", not "whether to change".

Subscribe on YouTube@sagnikteaches

Cloud telephony against legacy lines, on the points that matter

I read two suppliers' own pages on 4 November 2026. X-on Health describes Surgery Connect as a cloud phone system for primary care used by over 3,500 GP surgeries, scalable from a single practice to larger groups. Babblevoice describes a hosted system for primary care "with no contract commitments", made up of a phone system, patient self-booking by handset keypad, a desktop softphone that integrates with EMIS and SystmOne, call recording and dashboard reporting. It states it is assured on the NHS England Advanced Telephony Better Purchasing Framework. Note that Babblevoice's keypad booking is rule-based automation, not AI; ask any supplier which of its features actually use AI.

Neither page prints prices. That is normal here: you get a quote. The section on cost below gives you a way to compare quotes fairly.

Queuing and callbacks before AI

Most of the benefit that people credit to AI at the front door comes from queue design. Three settings do the work.

• Queue position and expected wait. A caller told "you are fifth in the queue" hangs on longer than one who hears nothing.
• Callback instead of waiting. The caller keeps their place and is rung back. X-on's reporting product names callback success rate as a tracked metric, which is the figure to watch.
• Time-based routing. Routine requests (repeat prescription status, test results, forms) go to a separate queue or digital route from urgent calls, so a form query does not sit ahead of a worried parent.

An illustrative sum. A surgery receives 260 calls between 8:00 and 9:00 and answers 140. Most of the 120 who give up ring again, so the following hour carries about 90 repeat calls on top of its own. Turning on callbacks for callers who have waited over four minutes might let 70 of those 120 hang up for good and be rung back once. That is 70 fewer repeat calls, from a setting, not an AI. Measure your own numbers before you quote anyone's.

Where AI call handling fits

AI call handling is the layer that answers some calls without a person. Two shapes exist. One sits in front, greeting every caller and asking what they need. The other works as overflow: calls that would otherwise queue past a threshold, or arrive when everyone is busy or after closing, go to the assistant.

One supplier's range shows how the pieces split. X-on's Surgery Assist is an AI-powered care navigation assistant for the practice website: a chatbot that answers practice questions, signposts services, shows the live call queue and takes callback requests, so patients find the next step "before they pick up the phone", with translation in 60+ languages. X-on's home page says it cuts admin workload by up to 50%; that is the vendor's own figure, a marketing claim to test, not a promise. The phone-side product is separate. X-on's Voice Agents for General Practice handle routine phone requests: callers choose between reception and the agent, and requests land in the practice's existing workflow for staff to review. The page lists the inbound agent as available now and an outbound appointment-reminder agent as coming soon, and describes the technology as a self-certified Class I medical device, which is one more reason the safety paperwork below applies. For a deeper walk through the hand-off and 8am pattern, AI receptionist for a GP surgery: the 8am rush and hand-offs covers it.

Jobs that suit overflow

1. Taking a callback request and the reason in one line.
2. Giving opening hours, the process for repeat prescriptions and where to find forms.
3. Booking or cancelling a routine appointment type you have defined in advance.
4. Telling callers that a phone line is for the practice and directing them elsewhere for an emergency.

Jobs to keep with people

• Anything where the patient describes symptoms and expects advice. An assistant should not triage; it should recognise possible urgency words and pass to a person immediately.
• Safeguarding, bereavement, distressed callers and complaints.
• Any request that depends on reading a record.

A hand-off rule you can write down

The assistant must transfer to a member of staff, or give the emergency message, when the caller:
- mentions chest pain, breathing difficulty, bleeding, a child who is very unwell, thoughts of self-harm, or anything they call an emergency
- asks for a clinician, a result, or advice about symptoms
- is distressed, confused or says the assistant is not helping
- asks twice for a person
- speaks a language the assistant cannot confirm it understands
If no one is free to take the transfer, take a callback request marked URGENT and tell the caller how to reach urgent services now.

Recording notices and what an AI adds

Recording calls helps with complaints and training. The rules in outline: tell the caller before recording starts, say why, keep recordings only as long as the stated purpose needs, and limit who can play them. Babblevoice describes its recording as an audit trail of all the surgery's calls, stored on servers it says comply with the NHS Data Protection & Security Toolkit and ISO 27001; confirm certificates and scope with a supplier before relying on a sentence on a web page.

An AI assistant adds three questions a plain recorder does not:

• Does the assistant keep its own audio, a transcript, or only a summary?
• Where is that stored, for how long, and who else processes it (the speech engine, the language model, a cloud host)?
• Does the opening message tell the caller it is an automated assistant and that the call is recorded?

An illustrative opening message: "You have reached the surgery. This is an automated assistant. Calls are recorded for patient safety and training. If this is an emergency, hang up and call the emergency number now. Otherwise, tell me what you need, or say 'person' at any time." Put the emergency line first, before anything else the caller has to listen to.

DCB0160 and the DPIA for an AI add-on

Two documents come up before an AI tool goes live in a surgery. NHS England's published guidance on AI-enabled ambient scribing states that safety requirements apply to all digital products used in the NHS, that suppliers verify safety under the DCB0129 standard, and that the deploying organisation completes the DCB0160 documentation, which it lists as a safety case, hazard log and monitoring framework, along with a Data Protection Impact Assessment. That guidance is about scribes, not phone assistants, so I would not claim it names call handling. The same logic fits any AI that touches patient access, and a clinical safety officer should confirm scope in writing for a phone assistant.

What to ask the supplier for

• Their DCB0129 clinical safety case report.
• Their data flow: what audio and text are processed, where, by whom.
• Their retention and deletion settings.
• A list of subprocessors.
• A statement of what the assistant must not do, and how it enforces that.

What the practice writes itself

• The DCB0160 safety case and hazard log, owned by a named clinical safety officer. A realistic hazard-log line: "Caller describing chest pain is not transferred. Cause: urgency phrase not recognised. Control: keyword list, immediate transfer rule, weekly test calls, monthly review of transfer logs."
• The DPIA covering call audio, transcripts, recordings, patient identifiers and the lawful basis.
• Patient-facing wording: a notice on the website and a short message on the line.

The scribing guidance also says to engage information governance support early and to include any "invisible processing" in the DPIA. A phone assistant that listens to every call is a good example of it.

Cost: comparing quotes on the surgery's own numbers

I could not find published prices for either supplier, so no figures appear here that I could not source. Instead, use a sheet that makes quotes comparable. Request every supplier to fill in the same rows:

An illustrative comparison in dollars (your quotes will be in your own currency). Quote A: $300 a month flat for the phone system plus $0.40 per AI call. Quote B: $450 a month with AI overflow included up to 1,000 calls, then $0.30 each. At 600 overflow calls a month, A costs $540 and B costs $450. At 1,800 calls, A costs $1,020 and B costs $450 plus 800 calls over the allowance at $0.30, so $690. At only 100 calls, A costs $340 and B still costs $450. So the cheaper quote flips with volume, and the point is to run every quote at your low, expected and high overflow volumes before choosing.

A fortnight to measure before you buy

1. Pull call logs for two weeks: calls offered, answered, abandoned, longest wait, by half hour.
2. Have reception tally call reasons for three mornings: appointment, prescription, results, forms, admin, clinical question, other.
3. Mark which reasons your hand-off rule would let an assistant handle. Count them as a share of peak calls.
4. Take that share as the maximum AI overflow volume for the quote sheet.

Filled in for an illustrative surgery: 1,100 calls in the two peak hours across the fortnight, of which a tally of 150 shows 31% appointments, 22% prescriptions, 14% forms and admin, and 33% results or clinical questions. The assistant could, at best, take the first three groups: 67% of peak calls. Plan for half of that in practice, since some callers will ask for a person.

Callback and overflow settings that misfire

A realistic one: a practice switches on callbacks but forgets that the outbound callback line shows a different number from the surgery's. Patients ignore calls from an unknown number, callbacks go unanswered, and the queue refills. The fix is to present the main surgery number on outbound callbacks and to send a text beforehand if the system supports it.

Another: AI overflow opens at the wrong time. The threshold is set to "when all lines are busy", which at a cloud practice never happens because capacity is software-defined. Set it by wait time, such as "over three minutes in queue", and check by test calls.

A third: recordings are kept indefinitely because nobody picked a retention period. Choose one with your information governance lead and set it in the system.

Test calls to run before go-live

1. Ring in at the busiest time with three staff phones and watch the queue behave.
2. Press for a person at every stage; confirm each route reaches a human.
3. Say "chest pain" to the assistant, then a looser phrase such as "my dad is gasping". Confirm the transfer.
4. Give a long, rambling reason and see whether the callback note is accurate.
5. Unplug the internet connection at the surgery and see where calls go. Write down the failover number.
6. Check a recording and the transcript for the right date, caller number and notice message.

Run those again after any supplier update. For measuring results afterwards, how to measure an AI receptionist's return in the first 90 days is the base method. The security and retention questions in AI scribe security: audio, retention and BAA vendor questions carry over to any audio product, and AI receptionist or answering service is the comparison if overflow is better covered by people.

One morning, walked through on the new system

Take an illustrative surgery with two receptionists on the phones at 8:00 and a third arriving at 8:30. Under the old set-up, eight lines fill within two minutes and the rest of the callers hear engaged. Under a cloud system with a queue, a callback offer after four minutes and an overflow assistant after six, the morning looks different.

• 8:00 to 8:10. Both receptionists are on calls. New callers hear their queue position and the emergency message.
• 8:10 to 8:30. Callers who have waited four minutes are offered a callback and most take it. Staff work the callback list in order after the live queue.
• Overflow. At six minutes, callers who wanted to ask about a repeat prescription or a form are offered the assistant. The assistant records the request and sends it to the right worklist, and anything with an urgency phrase transfers to a person.
• 8:30. The third receptionist starts on callbacks. The manager looks at the real-time view: longest wait, callbacks pending, transfers from the assistant.
• 10:00. Review the morning: did every transfer reach a person, were any callbacks missed, did any caller ask for a person twice?

Telling patients and staff what has changed

A phone change lands badly when it surprises people. Patients need to know three things before the first AI-answered call: that an automated assistant may answer, that they can always ask for a person, and that calls are recorded. A website banner and a text message to patients who have booked online cover most of it. Illustrative text: "From Monday, if our lines are busy you may be offered a callback or an automated assistant. Say 'person' at any time to speak to a member of staff. Calls are recorded for safety and training."

Staff need something different: what the assistant can and cannot do, where its notes appear, and what to do when a patient is cross about it. A one-page note and a ten-minute walk-through at a team meeting beats a long manual. Ask reception to log every call where a patient says the assistant got something wrong, then review the log weekly in the first month. Treat those entries as the early warning your hazard log was written for.

Which set-up suits a single site, a branch or a group

A single-site surgery with one queue problem: cloud system with queue position and callbacks first; AI overflow only if the fortnight's tally shows a large block of scripted calls. A surgery with a branch: a system that lets staff at either site take calls from one queue, plus one set of reports. A practice in a larger group: look at supplier assurance on the national purchasing framework, group-level reporting and a single DPIA across sites.

If you would rather map this with someone, that is what my AI implementation consultation covers. The AI guide collects the rest of the series.

Questions practice managers ask about phone systems and AI

Do we need a cloud phone system to add AI call handling?

Usually yes, or at least a digital line. AI overflow, callbacks and call data rely on a system that can route calls by rule and expose data to other software. Some AI assistants sit alongside an existing system and take calls via forwarding, so ask each supplier what it needs from your telephony before you compare prices.

Does AI call handling need DCB0160?

Treat it as yes where the tool touches patient access or clinical content. NHS England's guidance on ambient scribing says deploying organisations complete the DCB0160 documentation and a DPIA, and that safety requirements apply to all digital products used in the NHS. Ask your clinical safety officer to confirm scope for a phone assistant.

Can we record all calls?

Many practices do, but patients must be told before the recording starts and the purpose must be defined, such as complaints and training. Set a retention period, restrict who can listen, and check whether an AI assistant keeps its own copy of audio or transcripts.

What should a quote include?

Ask for set-up, monthly cost per line or per user, call minutes, AI per-call or per-minute charges, recording storage, number porting, contract length and exit terms. Compare quotes on a full year of your own call volume, not on the monthly headline.

Further reads

• How Much Does an AI Receptionist Cost for a Small Business? — What an AI receptionist costs across business types.
• AI Receptionist vs Front Desk Hire: A Dental Practice Comparison — Another clinic comparing a machine with a front-desk hire.
• What Does an AI Receptionist Cost a Dental Practice in 2026? — Cost sums for a clinic that books by phone.
• Healthcare AI Call Centre: When a Group Practice Should Centralise Calls — When several sites should share one call queue.
• What to Do When an AI Receptionist Gets a Booking Wrong — What to do when automated booking goes wrong.
• What Is an AI Receptionist and How Does It Handle Bookings? — A plain explanation of how AI receptionists book.
• Medical Practice Phone System: Add AI Answering to Existing Lines — Three ways to put AI call answering on a medical practice's current phones: forwarding, SIP or replacement, plus overflow rules, notices and number porting.
• 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: X-on Health home, Surgery Assist and Voice Agents for General Practice pages; Babblevoice home page; NHS England guidance on AI-enabled ambient scribing products; Openreach digital phone line upgrade page (all read 4 November 2026).