Phones call handling

Safety rules for AI handling patient calls

Red-flag escalation, hand-off to staff and monitoring rules for AI systems that answer calls in GP surgeries, dental practices and private clinics.

Ava research teamPublished 6 min read

On this page
  1. Why safety rules come first
  2. Five non-negotiable rules
  3. The escalation table
  4. Wording the urgent instruction
  5. Building the safety case
  6. Example: a test-call script set
  7. Carers, children and third-party callers
  8. Monitoring after go-live
  9. Hand-offs that actually work
  10. Frequently asked questions
  11. References

Key takeaways

  • An AI call system must never assess, reassure or delay: on any red flag it gives the agreed urgent instruction at once.
  • Write the rules with your clinical lead, as a short table of triggers, required words and hand-offs.
  • When the system is unsure, the default is a person, not another question.
  • Test with scripted calls before go-live and after every change, and review every escalation.

Why safety rules come first

Calls to a practice are not all routine. Some callers describe symptoms that need emergency help, some are in crisis and some are calling about a child or a vulnerable adult. In NHS 111's triaged calls in July 2026, 45.6% of callers spoke to a clinician or clinical adviser and 14.5% were advised self-care[1]. A practice line carries a similar mix of urgency, and an AI system answering it has to handle the urgent minority safely every time.

Outcomes of triaged NHS 111 calls, July 2026Share of triaged calls, England
Outcomes of triaged NHS 111 calls, July 2026Bar chart. Spoke to a clinician or clinical adviser: 45.6%; Advised self-care: 14.5%.Spoke to a clinician or clinical adviserSpoke to a clinician or clinical adviser: 45.6%45.6%Advised self-careAdvised self-care: 14.5%14.5%Outcomes of triaged NHS 111 calls, July 2026Bar chart. Spoke to a clinician or clinical adviser: 45.6%; Advised self-care: 14.5%.Spoke to a clinician or clinicaladviserSpoke to a clinician or clinical adviser: 45.6%45.6%Advised self-careAdvised self-care: 14.5%14.5%

Source: NHS England, 2026[1]

Show the data
ItemValue
Spoke to a clinician or clinical adviser45.6%
Advised self-care14.5%

Public expectations point the same way. Healthcare is the area where UK adults most often say AI needs careful management to prevent negative outcomes[2]. Among GPs not yet using AI, patient privacy and data security worried 82%[3]. Clear escalation rules are the main way a practice shows that care has been taken.

Five non-negotiable rules

  1. Never delay. If a red flag is heard, the urgent instruction comes immediately, before any queue, menu or further question.
  2. Never assess. The system does not judge how serious a symptom is; it recognises trigger words and phrases and acts on them.
  3. Never reassure. Phrases such as 'that doesn't sound serious' or 'you can probably wait' are banned from the system's vocabulary.
  4. When unsure, hand to a person. If the system cannot understand the caller, the caller is distressed or the request does not fit an approved route, it passes the call to staff or gives the urgent route when the practice is closed.
  5. Log everything. Every escalation, hand-off and failed understanding is recorded and reviewed.

These rules also keep the system on the right side of regulation. The MHRA says many software and AI products used in health and social care are regulated as medical devices[4]. A system that triages symptoms is likely to fall into that territory; one that recognises red flags and passes them on with fixed wording is doing a narrower, safer job.

The escalation table

Write the rules as a table that your clinical lead approves and signs. Keep the wording of each response fixed, so it can be tested and audited.

Example escalation table to be agreed and signed by the clinical lead
Trigger categoryExamples the system listens forRequired responseHand-off
Possible life-threatening emergencyChest pain, can't breathe, collapsed, unconscious, stroke signs, severe bleeding, seizure, severe allergic reaction'This could be an emergency. Please hang up and call 999 now.'Log; alert duty clinician if the caller is a registered patient
Unsure if it is an emergency'I don't know if I should go to A&E', worsening symptomsGive 999 for emergencies and NHS 111 for urgent advice, in agreed wordsTransfer to staff if open; otherwise urgent routes
Mental health crisisThoughts of suicide or self-harm, 'I can't keep myself safe'Agreed crisis wording, including 999 if in immediate dangerImmediate transfer to staff if open; log and alert
SafeguardingA child or adult at risk, domestic abuse, someone being hurt999 if in immediate danger; otherwise transferStaff and safeguarding lead
Dental urgentFacial swelling, trauma, bleeding after extractionAgreed urgent dental route and 999 for breathing or swallowing difficultyStaff if open; log
Not understoodRepeated misrecognition, distress, other language'I'll pass you to a member of the team.'Staff if open; urgent routes if closed

The NHS website's own guidance is the reference point for public wording: 999 is for life-threatening emergencies such as serious road traffic accidents, strokes and heart attacks, and NHS 111 can help people who are not sure whether they need to call 999[5]. Use the same plain language your patients will see elsewhere.

Wording the urgent instruction

The exact words matter more than any other part of the configuration. A good urgent instruction is short, direct and impossible to misread as optional. It tells the caller what to do now, not what the practice can offer later.

  • Lead with the action: 'Please hang up and call 999 now', not 'You may wish to consider calling 999'.
  • Keep it to one or two sentences and repeat it once.
  • Do not add a menu, a survey or a request for details after an emergency instruction.
  • Use the same words in the greeting, the hold loop and the AI system, so patients hear one consistent message.
  • For practices in Wales, agree the instruction in Welsh and English and test both.

Example wording for the clinical lead to approve: 'What you have described could be an emergency. Please hang up now and call 999. If you are not sure, call NHS 111.'

Building the safety case

  1. Name a clinical safety lead

    A clinician who owns the escalation table, signs off wording and reviews incidents. Under DCB0160, health organisations deploying health IT need a clinical safety officer; your ICB may help.

  2. Get the supplier's safety documents

    Ask for the manufacturer's clinical safety case and hazard log under DCB0129, and check that the system's behaviour matches what is described.

  3. Write your hazard log

    List what could go wrong in your setting: missed red flag, wrong urgent wording, caller unable to reach a person, system outage. For each, record controls and residual risk.

  4. Agree the escalation table

    As above, signed by the clinical lead, with fixed wording.

  5. Test before go-live

    Run a scripted test set covering every trigger category, in different phrasings, accents and with background noise. Record pass or fail for each.

  6. Go live narrowly

    Start with limited hours or call types, with closer monitoring for the first weeks.

  7. Review and retest

    Review escalations weekly and retest the full script set after any change to wording, models or configuration.

Our guide to DCB0160 and DPIAs for AI tools covers the documentation in more depth.

Example: a test-call script set

Example: a fictional practice, Riverside Medical Practice, wrote forty test calls before switching on AI overflow answering. They included plain triggers ('my dad has collapsed'), indirect ones ('my chest feels tight and I feel sick'), triggers buried in admin calls ('I need to cancel my appointment, I've been having chest pains'), callers speaking quickly, background noise and a caller who switched between English and Polish.

The first run failed three calls: the system missed a trigger buried at the end of a long admin request, and twice it asked a follow-up question before giving the urgent instruction. The supplier changed the configuration and the clinical lead tightened the wording. The second run passed all forty. The practice now reruns the full set after every change and adds a new test whenever a real call reveals a gap.

Carers, children and third-party callers

Many urgent calls come from someone other than the patient: a parent about a baby, a carer about an older relative, a care home nurse about a resident. Triggers must work whoever is speaking, so test phrases such as 'my son', 'my mum' and 'one of our residents' alongside first-person ones.

Agree a lower threshold for hand-off when the caller mentions a baby or young child, a pregnant woman, or a patient the caller describes as confused, very drowsy or not themselves. In these cases the system should not try to complete an admin task first; it should give the urgent route or pass the call straight to a person. Care homes and other professional callers should have a bypass number so they never meet the AI system at all.

Monitoring after go-live

  • Every red-flag escalation reviewed by a clinician within one working day.
  • A weekly random sample of non-escalated calls, checked for missed triggers.
  • Calls where the caller asked for a person, and how long the hand-off took.
  • Calls the system failed to understand, by reason.
  • Staff and patient reports of problems, logged and investigated.
  • Patient safety incidents reported through your usual route, such as the NHS Learn from Patient Safety Events service in England.

Set a stop rule in advance: for example, any missed life-threatening trigger pauses the service until the cause is found and fixed. Agreeing that before go-live avoids difficult decisions under pressure.

Hand-offs that actually work

An escalation rule is only as good as the hand-off behind it. If the system transfers a distressed caller to a queue of twenty, the rule has failed. Give escalated transfers priority in the queue or a direct route to a named role. When the practice is closed, the system should give the urgent routes clearly and end the call, not take a message that nobody will read until morning.

Take care with data on hand-off. Pass staff only what they need: the caller's number, the trigger category and a short transcript excerpt. The Caldicott Principles' emphasis on using the minimum necessary information applies here as everywhere else. For the wider picture, see AI phone answering for GP surgeries, AI receptionists in dental practices and bilingual messages in Wales.

Frequently asked questions

Can an AI phone system decide whether a patient needs 999?

It should not assess severity. It should recognise agreed trigger words and phrases and give the fixed urgent instruction immediately, leaving clinical judgement to people.

Who should sign off the escalation rules?

A named clinical safety lead, usually a GP or senior clinician, working with the practice manager and the supplier's clinical safety officer.

How often should we test the escalation rules?

Before go-live, after any change to wording or configuration, and whenever a real call reveals a gap. Many practices also rerun the full set monthly.

References

  1. NHS England (2026). Statistical Note: Integrated Urgent Care Aggregate Data Collection (IUC ADC) July 2026 (opens in a new tab) ↑

  2. GOV.UK (Department for Science, Innovation and Technology) (2024). Public attitudes to data and AI: Tracker survey (Wave 4) report (opens in a new tab) ↑

  3. Nuffield Trust (2025). How are GPs using AI? Insights from the front line (opens in a new tab) ↑

  4. GOV.UK (Medicines and Healthcare products Regulatory Agency) (2025). Software and artificial intelligence (AI) as a medical device (opens in a new tab) ↑

  5. NHS (nhs.uk) (2026). When to call 999 (opens in a new tab) ↑

Book a demo

See Ava handle your calls.

Book a short demo. We’ll walk through your calls and workflow, and agree what Ava should take care of.

Book a demo