AI Voice Agents for GP Practices: Costs and Safety Checks

Expect a quote, not a price list: GP-specific suppliers rarely publish rates. As a guide, one general voice platform lists $0.07 to $0.31 a minute, so a practice taking 1,500 calls a month could see roughly $200 to $870 in usage, before set-up, integration and safety work. Complete a DPIA and DCB0160 clinical safety work before go-live.

The trap is comparing a platform rate with a supplier's all-in monthly fee. They are different things. A platform bills for the minutes the technology runs; a GP supplier bundles the technology, the integration with your clinical system, the call flows, the safety documentation and support. The bundle may be worth paying for, but only if you can see what is inside it. Price the layers separately and the comparison becomes fair; then finish the safety checks before the first patient is answered by a machine.

Follow me on Instagram@sagnikteaches

I am not a clinician or a lawyer, and nothing here replaces your clinical safety officer or data protection adviser. I have kept to published list prices I checked on 4 November 2026 and said where a figure is mine. The running example is a mid-sized practice of about 9,000 registered patients with 1,500 calls a month and an 8am surge, using the AI receptionist at the 8am rush tutorial for the operational side.

Connect on LinkedInSagnik Bhattacharya

Premise check: what "AI voice agent" means for a practice

Three kinds of product get called a voice agent, and the price depends on which one you are being sold.

Subscribe on YouTube@sagnikteaches

• Call routing with a script. Answers, asks "what is your call about?", routes to the right queue and sometimes texts a link. Simple, low risk, usually the cheapest.
• Task agents. Takes a repeat-prescription request, books or cancels an appointment, collects a message for a named team. It needs integration with the clinical system and has a higher safety burden.
• Clinical-conversation tools. Anything that asks about symptoms and shapes what happens next. This is the high-risk category, and in my view not one for a first purchase. Software that assesses symptoms can also count as a medical device, which brings legal duties for its maker, so ask the supplier for its written view on that.

Ask the supplier to state in writing which of the three it is selling you, and in which category each call type will sit. For the cost comparison below I assume the first two.

How the pricing is built: six layers on a quote

A supplier's monthly fee is the sum of layers you can ask about one by one.

1. Platform usage. The per-minute or per-call charge. This is the only layer with public numbers.
2. Telephony. The number, forwarding, call transfer and the minutes the phone company bills. On Retell's page the call rate was $0.015 a minute for the default country shown, with no charge for SIP trunking or custom telephony, and its own phone numbers cost $2 a month.
3. Set-up and integration. One-off work to connect the agent to the appointment system and build the call flows. Often quoted as a fixed sum or bundled into a minimum term.
4. Safety and compliance work. Data agreements, clinical safety documentation, testing. Sometimes included, sometimes a separate line, sometimes a plan tier. On Vapi's page, the compliance add-on for health-data handling was $2,000 a month, which shows how large this line can be; Retell's documentation, by contrast, says its BAA and DPA are self-signed with no additional fee. The spread is the reason to ask.
5. Support and monitoring. Named support, response times, call review. Vapi's page, for example, offers email support with no response commitment when you buy no success package, a two-business-day response on Core ($29 a month) and one business day on Pro (from $999 a month).
6. Your own time. Staff time to write the scripts, test, review calls and answer patients' complaints. This is the layer most business cases forget.

The published numbers I could check

None of these is a GP supplier's price. They show the shape: per-minute platform pricing is cheap, compliance and support packaging is where costs climb, and per-call packages are expensive at practice volumes. A GP-specific supplier will quote above the platform rates because it is adding integration, safety work and support, which is why the quote is the thing to scrutinise, not the headline.

A worked sum for a practice with 1,500 calls a month

Illustrative assumptions: 1,500 calls a month; the agent fully handles 40 per cent (600 calls) and answers and passes on the rest (900 calls) with a 40-second hand-off; the average fully handled call lasts 3 minutes. That gives 600 x 3 = 1,800 minutes of full calls plus 900 x 0.7 = 630 minutes of hand-offs, so about 2,430 minutes. Add 15 per cent for the busiest weeks and you have roughly 2,800 minutes a month.

My own arithmetic from list prices, so check the live pages. If Retell's $0.015 telephony line applies on top, add about $42 a month (2,800 x $0.015). Two readings follow. First, call length drives the bill on the platforms, so a flow that gets to the point in two minutes instead of three cuts the usage by a third. Second, per-call packages get expensive at this volume; Smith.ai's Enterprise tier lowers the rate (500 calls at $1.60 each, custom pricing from 1,000 calls), but the shape stays the same. None of these figures includes the integration or the safety work, which is where a GP supplier's quote will sit well above these numbers.

Build the all-in monthly picture

Take the platform line as the floor, and add what the supplier quotes for each other layer. If a quote arrives as one figure, ask for it split. Illustratively, a supplier quoting a single fee of $1,800 a month might be: $450 of platform usage, $350 of telephony and call transfer, $500 of integration support, $300 of monitoring and $200 of safety documentation upkeep. Whether that is good value depends on what the practice would otherwise pay in staff time. A reception hour might cost the practice around $15 to $20 in wages, so $1,800 would equal about 90 to 120 reception hours a month, and the agent must clearly take more than that from the phones for the quote to make sense. Use your own wage figures; these are placeholders.

One-off costs to ask about

• Set-up and integration, with a fixed price and what it includes.
• Testing: how many test calls, and who writes the scenarios.
• Training for reception staff on what the agent does and how to take a hand-off.
• Clinical safety documentation supplied by the supplier, and what you must write yourself.
• Any minimum term and early-termination fee. The contract terms to check before signing a voice agent sets out the clauses on minutes, number ownership, export and exit; they apply to a GP supplier as much as a dentist's.

Safety check one: DCB0160 clinical risk management

DCB0160 is the clinical risk management standard for organisations that deploy and use health IT systems; its companion for the companies that make them is DCB0129. The AI and Digital Regulations Service, a guidance site for health and care adopters, describes compliance as legally required for adopters, who must also make sure the developer complies with DCB0129. I found no national playbook written for voice agents. The nearest is the national guidance on AI scribing products in health and care settings, whose quick implementation guide starts with the same steps and points practices without the in-house skills to their commissioner for help. The standard's own page sat behind a bot check when I tried it, so read the current text before you start. In outline, as the deploying organisation you:

1. Appoint a clinical safety officer. A named person with the clinical knowledge, experience and competence to own the clinical risk work for this system. The scribing guidance makes this the first action.
2. Put a clinical risk management system in place. Write down the governance (who is responsible), the activities (how hazards are found, assessed and reviewed) and the training staff receive.
3. Keep a hazard log. Every way the agent could harm a patient, how likely, how severe, what controls reduce it, and who accepted what residual risk.
4. Produce a clinical safety case. The summary argument, with evidence, that the system is acceptably safe for your use, completed before go-live. A product that cannot meet the standard should not be adopted.
5. Set up monitoring. The scribing guidance pairs the safety case and hazard log with a monitoring framework: what you review after go-live, how often, and who acts on what you find.
6. Ask the supplier for its documentation. Its clinical safety case report under DCB0129. The same guidance says the safety requirements apply to digital products used in the health service whether or not they count as medical devices. If the supplier cannot produce the report, stop.

Hazards a voice agent adds, with the control for each

The control column is the practical value of the work: a log entry with no control is a worry, not a plan. Test each control with scripted calls and keep the recordings as evidence in the safety case.

Safety check two: the DPIA

A voice agent processes health information and recordings, so a data protection impact assessment (DPIA) is the safe default before go-live. The questions a DPIA asks are practical: what data is collected, why, where it goes, who can see it, how long it is kept, and what could go wrong for patients. The guide to whether you need a DPIA before using AI tools gives the test and a layout, and the information governance guidance that accompanies the national scribing guidance includes a DPIA template that adapts well to a voice agent. For a voice agent, make sure it covers:

• The call recording. Retention, who can listen, how a patient's deletion request is handled.
• Sub-processors. The speech, language-model and telephony companies the call passes through, and where they process data. Retell's compliance page, for instance, says it does not currently operate services inside the EU, so ask any platform where your patients' call data would be processed and record how that transfer is protected.
• Model training. A written statement that recordings and transcripts are not used to train models.
• Special category data. Health information needs a lawful basis and the right safeguards; record the basis you rely on.
• Your data protection officer. Ask them to review it, and keep their advice with the DPIA.

Ask the supplier for its security documents, and for a data processing agreement signed before go-live. If its page lists compliance as an optional add-on, as the platform pages above do, find out whether your use requires it and what it costs.

Patient communication: what to say and where

Patients should know when they are speaking to a machine, why calls are recorded and how to reach a person. Do this on four channels:

1. The call itself. The first sentence says it is an automated assistant, the call is recorded, and "say receptionist at any time" gets a person.
2. The website. A short page explaining what the agent does and does not do, and how recordings are handled.
3. A text or letter before go-live. Two lines, no jargon, sent to patients before the change.
4. The waiting-room notice. For patients who do not use the website.

Sample opening (illustrative):
"Hello, this is [Practice name]'s automated assistant. This call is recorded.
If you need a person at any time, say 'receptionist'. If this is an
emergency, hang up and call 999. How can I help you today?"

Sample website text (illustrative):
"From [date], an automated assistant answers some calls to the practice. It
can take messages, book routine appointments and pass you to our team. It
cannot give medical advice. You can ask to speak to a person at any time."

Think about who is left out. Patients with hearing problems, those who do not speak English as a first language, and older patients who dislike automated lines may drop off. Track hang-ups in the first month, compare them with the same weeks before go-live, and keep a human route that is genuinely easy to reach. If there is a practice participation group, show them the script before launch.

Go-live checklist and what to watch in the first month

1. Safety case signed by the clinical safety officer, with the hazard log and supplier documents attached.
2. DPIA complete and reviewed by your data protection officer, with a signed data processing agreement.
3. Fifteen scripted test calls passed, covering chest pain, a child with a fever, a request for a repeat prescription, a bereaved caller, a caller who cannot be understood and a nuisance call.
4. Fallback tested. Switch the agent off and confirm calls reach reception.
5. Patients informed on all four channels at least a week ahead.
6. Reception briefed on hand-offs: what the agent passes on, in what format.
7. A review calendar. Listen to a sample of 30 recordings in week one, then ten a week, and log every error.
8. Measures set: calls answered, calls transferred, average wait, hang-ups, and complaints about the phone line, each against the previous month.

The measures are the same as in how to measure an AI receptionist's return in the first 90 days; in a practice, add "safety events logged" as a first-line number. For the question of whether a human answering service might suit you better, AI receptionist or answering service compares the two.

Questions to put to a supplier, in writing

Send these before a demo. Written answers let the clinical safety officer and the data protection officer read the same document you did.

Subject: Questions before we consider [product] for our practice

Hello [first name],

Please answer in writing:
1. Which of these does the product do: route calls, take tasks, or
   discuss symptoms? Please list each call type and its category.
2. A quote split into: platform usage, telephony, set-up and
   integration, safety documentation, support, and any minimum term.
3. Included minutes or calls, how a minute is counted, overage rate and
   whether we can set a spend cap.
4. Your clinical safety documentation (hazard log and safety case) and
   the name of your clinical safety officer.
5. A draft data processing agreement, your sub-processors, where they
   process data, and a statement that recordings are not used for model
   training.
6. How the product behaves on a red-flag phrase, and a recording of
   your own testing of it.
7. What happens if your service is down: how quickly do calls reach a
   person?
8. Our exit rights: data export, number ownership and notice period.

Kind regards,
[your name], [practice]

How to read the answers

A good supplier answers questions 4 and 6 with documents, not reassurance. If the reply to question 4 is "we are working on it", the product is not ready for your use. If question 3 is answered with a per-call price that you cannot reconcile with your call volumes, ask for the working. A reply that includes a recording of its own red-flag testing is rare and worth a lot.

A realistic mistake and how it showed up

Illustratively, a practice launches with a script that handles "repeat prescription" calls. A patient says "I'm out of my tablets and I feel dizzy", and the agent, which has been told to focus on the repeat request, records the request and ends the call politely. The dizziness never reaches a human. Nothing is wrong with the product's price; the hazard log had no entry for a repeat request that arrives with symptoms. The fix was a rule that any symptom word during any call triggers the transfer path, and a monthly review of 30 recordings looking for exactly that pattern. The point of the log is to find the gap before a patient does.

Deciding: a rule of thumb for the partners

Go ahead if three things are true: the quote splits every layer so you can see what you pay for; the supplier hands over its safety documentation and a signed data agreement before you commit; and the agent's first jobs are narrow (routing, messages, simple booking) with an easy route to a human. Wait if the quote is one number with no split, if clinical safety documents are "available later", or if the sales pitch includes triage. Revisit in six months, because the products and the standards around them are moving quickly.

If it would help to size your call volumes and draft the questions for a supplier, that is what my AI implementation consultation covers. The steps above stand without it: get the quote split, finish the safety work first, and let the recordings, not the demo, decide.

Questions practice managers ask about cost and safety

Why do GP voice-agent suppliers not publish prices?

Most price by quote because cost depends on call volume, integrations with the clinical system, hours covered and the safety work they do for you. Published platform rates only cover the technology layer. Ask for a written quote that lists set-up, monthly fee, included minutes, overage and every optional extra.

Does a voice agent need clinical safety sign-off?

Software that changes how patients reach care is generally treated as a health IT system, so your practice should work through the deployment clinical risk standard, DCB0160, with a named clinical safety officer. Check the current standard text and your commissioner's expectations, and ask the supplier for its own safety documentation.

Can a voice agent triage symptoms?

I would not let one. The safe pattern is to collect the reason for calling, recognise red-flag phrases, give emergency instructions and pass the call to a human. Any supplier claiming clinical triage should show you its clinical safety case, its evidence and its registration status before you consider it.

What should we tell patients?

Say it plainly on the phone message, the website and a letter or text: an automated assistant answers first, calls may be recorded, and a person is one request away. Offer a clear route for patients who cannot or will not use it, and review who is hanging up.

Further reads

• What Does an AI Receptionist Cost a Dental Practice in 2026? — A comparable clinical-practice cost breakdown.
• What to Check in an AI Tool's Privacy Policy and Terms — What to read in the privacy terms behind a supplier.
• How to Evaluate an AI Implementation Proposal or Quote — Judge a supplier's quote line by line.
• Telling Patients You Use AI Note-Taking: Consent Wording That Works — Wording for telling patients when AI is involved.
• What to Do When an AI Receptionist Gets a Booking Wrong — What to do when an agent books wrongly.
• How to Prepare Your Small Business for AI Agents — Groundwork before any agent goes live.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Retell AI pricing page and compliance documentation, Vapi pricing page and Smith.ai AI receptionist pricing page; the national guidance on AI-enabled ambient scribing products in health and care settings; the AI and Digital Regulations Service page on clinical risk management standards (all read 4 November 2026; the DCB0160 standard page itself was behind a bot check). All prices are USD list prices for technology platforms, not quotes from any GP-specific supplier. Not legal or clinical advice.