Emergency out of hours

Using AI for emergency call triage: safe guardrails

Safe rules for AI answering emergency plumbing and heating calls: fixed gas and CO scripts, no diagnosis, vulnerable callers and fast human handover.

Ava research teamPublished 6 min read

On this page
  1. Where AI fits on an emergency line
  2. The core principle: collect, read, route
  3. Hard rules for safety-critical calls
  4. Fast human handover
  5. Writing the approved scripts
  6. Questions to ask any supplier
  7. Testing before go-live, and checking after
  8. Train your people alongside the system
  9. Callers, transparency and data
  10. Example: a gas smell at 11pm
  11. Frequently asked questions
  12. References

Key takeaways

  • 81% of UK construction businesses used no AI at all in autumn 2026, against 64.1% of all businesses[1], so most firms are starting from scratch.
  • On emergency lines, AI should collect details, read approved safety wording and route calls; it should never diagnose.
  • Gas smells and carbon monoxide trigger fixed scripts with the emergency number and an immediate urgent flag.
  • Humans check the output: 84% of UK businesses using AI reported some human checking of AI outputs[2].

Where AI fits on an emergency line

Emergency calls arrive when nobody is free to answer: engineers are on jobs, the office has closed, the owner is driving. AI call answering can pick up every call, take the details, and route urgent ones to the engineer on call. Used well, that means fewer lost customers and fewer emergencies stuck in a voicemail box until morning. Used badly, it means a machine giving advice to someone standing in a gas-filled kitchen.

The trade is early in its adoption. In autumn 2026, 81% of UK construction businesses said they did not use any AI technology, compared with 64.1% of all businesses and 65.7% of the smallest firms[1]. Among construction businesses that had not adopted it, lack of AI expertise was the most cited barrier, ahead of cost[1]. Clear rules make adoption safer and easier to explain to your team.

UK businesses not using any AI technology, autumn 2026Late September to early October 2026
UK businesses not using any AI technology, autumn 2026Bar chart. Construction businesses: 81%; Smallest businesses (0 to 9 employees): 65.7%; All businesses: 64.1%.Construction businessesConstruction businesses: 81%81%Smallest businesses (0 to 9 employees)Smallest businesses (0 to 9 employees): 65.7%65.7%All businessesAll businesses: 64.1%64.1%UK businesses not using any AI technology, autumn 2026Bar chart. Construction businesses: 81%; Smallest businesses (0 to 9 employees): 65.7%; All businesses: 64.1%.Construction businessesConstruction businesses: 81%81%Smallest businesses (0 to 9 employees)Smallest businesses (0 to 9 employees): 65.7%65.7%All businessesAll businesses: 64.1%64.1%

Source: Office for National Statistics, 2026[1]

Show the data
ItemValue
Construction businesses81%
Smallest businesses (0 to 9 employees)65.7%
All businesses64.1%

The core principle: collect, read, route

On an emergency line, an AI system should do three things. It collects the details a human needs: name, address, phone number, the problem in the caller's words, who is in the home. It reads wording you have approved, word for word, for safety situations. And it routes the call according to your rota and rules. It should not diagnose faults, estimate repair costs it has not been given, or reassure callers that a situation is safe.

That boundary protects callers and protects you. A human engineer who gives poor advice on the phone can be questioned and retrained. A system that improvises advice at scale can repeat the same mistake to every caller that night.

In practice, write the job description for the system the way you would for a new call handler on their first night: what to ask, what to say word for word, who to call and when, and what never to do. If a rule would be unclear to a new member of staff, it will be unclear to the system too. The same document then doubles as training for your human team.

Hard rules for safety-critical calls

Write these rules into your system configuration and your staff guidance.
TriggerThe system mustThe system must never
Smell of gasRead your approved gas message, including the National Gas Helpline 0800 111 999, mark the call urgent and alert the on-call engineerSuggest looking for the leak, using switches, or that the smell is probably harmless
Carbon monoxide alarm or symptomsRead approved advice to stop using the appliance, go outside and get medical advice, with 999 for serious symptomsSuggest silencing the alarm or that it may be faulty
Water near electricsAdvise keeping away from wet sockets and fittings, then take detailsTalk the caller through electrical isolation
Vulnerable caller or householdMark priority and transfer to a human or alert the on-call engineerLeave the caller with only a callback promise
Distress, confusion or a request for a personTransfer to a human or the rota immediatelyKeep the caller in an automated loop

Base the safety wording on recognised sources. The NHS advises people who suspect carbon monoxide poisoning to stop using the suspect appliance and get medical advice[3], and lists the National Gas Helpline on 0800 111 999 for suspected gas leaks[3]. When a gas escape is reported, the gas conveyor must attend as soon as reasonably practicable and stop the escape within 12 hours[4], which is why the emergency number always comes before your own engineer.

Some products are built around these limits. Ava, for example, reads your approved safety message and the gas emergency number, gives no technical advice and flags the call urgent. Whatever you use, check that it behaves this way in testing, not just in the sales demo.

Fast human handover

  1. Define urgent

    List the situations that always go to a person: gas, carbon monoxide, uncontrolled water, no heat with a vulnerable household, and any caller who asks for a human.

  2. Set the rota

    Engineer on call first, then a back-up, then a named owner or manager. Keep it updated in the system, including holidays.

  3. Set time limits

    If the first person does not answer within a short set time, move to the next. Never let an urgent alert sit unanswered.

  4. Give the caller a clear next step

    If nobody can be reached, the system should say what happens next and when, and repeat any safety advice.

  5. Log everything

    Every alert, attempt and handover should be time-stamped so you can review gaps the next morning.

Writing the approved scripts

  • Keep each safety script short, in plain English, and read it aloud to check it sounds natural.
  • Put the emergency number first in the gas script, then the instruction to leave the property and not use switches or flames.
  • Use the same wording as your human call handlers, so customers hear one consistent message.
  • Date each script and review it at least yearly, and after any incident or near miss.
  • Have a competent gas engineer approve the gas and carbon monoxide wording.

Questions to ask any supplier

  • Can I see and approve the exact words the system uses for gas, carbon monoxide and water emergencies, and can it be stopped from adding its own?
  • What happens if a caller mentions a gas smell halfway through a call about something else?
  • How does it decide a caller is vulnerable or distressed, and what does it do then?
  • How are urgent alerts sent, how quickly, and what happens if nobody responds?
  • Can I listen to recordings or read transcripts of every call, and how long are they kept?
  • Where is call data stored and processed, and who can access it?
  • Which features are live today and which are planned? Ask to see the live ones working on a test call.

Be wary of any supplier that cannot answer these clearly, or that describes its system as able to diagnose boiler faults over the phone. For emergency lines, the safest system is the one with the narrowest, most predictable job.

Testing before go-live, and checking after

Run realistic test calls before the system answers a single customer: a calm caller with a gas smell, a panicking one, a caller who mentions a smell only at the end, a confused elderly caller, a caller with a strong accent, a caller on a bad line. Check that every safety rule fires and every handover reaches a person. Repeat the tests after any change to scripts or settings.

Then keep checking. Most UK businesses already do this: 84% of those using AI reported at least some human checking of AI outputs, and 67% reported significant checking[2]. For emergency lines, review a sample of calls each week, and every call flagged urgent, for the first few months.

Train your people alongside the system

An AI answering service changes the job of everyone on the rota. The on-call engineer now receives alerts rather than raw calls, so agree how quickly they acknowledge them and what they say when they ring back. The office needs to know how to update the rota, change a script and pull a call record. Run a short briefing before go-live and a second one after the first month, using real calls from your own line. Make sure nobody treats the system as a reason to stop answering the phone themselves when they can.

Callers, transparency and data

Many callers are comfortable with automated systems; six in ten UK adults reported using AI chatbots in the past three months in 2024[5]. Others will want a person, especially when frightened. Make it easy to ask for one. Tell callers early that they are speaking to an automated assistant, and say if calls are recorded and why, in your greeting and your privacy notice. Record only what you need for the job, and be careful with health details given during emergency calls.

Example: a gas smell at 11pm

Example: a caller says the boiler stopped and, near the end of the call, that the kitchen smells of gas. The system reads the approved gas message, gives 0800 111 999, tells the caller to leave the house and avoid switches, marks the call urgent and alerts the on-call engineer, who rings back within minutes to confirm the caller has phoned the National Gas Helpline.

Example, continued: the next morning the office reviews the call, sees the smell was mentioned late, and adds a question about gas smells near the start of every no-heat call. The script is re-tested that afternoon.

Frequently asked questions

Can AI answer emergency plumbing calls safely?

Yes, if it only collects details, reads approved safety wording and routes calls to people. It should never diagnose faults or tell a caller a situation is safe.

What should an AI system say if a caller smells gas?

Your approved gas message, starting with the National Gas Helpline on 0800 111 999[3], then flag the call urgent and alert a person.

Should callers be told they are speaking to AI?

It is good practice to say so at the start, offer a person on request and explain any call recording in your greeting and privacy notice.

How often should AI call handling be checked?

Review every urgent call and a weekly sample of others, especially in the first months. Most UK businesses using AI report human checking of its outputs[2].

References

  1. Office for National Statistics (2026). Business insights and impact on the UK economy: wave 165 dataset (AI tables) (opens in a new tab) ↑

  2. Department for Science, Innovation and Technology (GOV.UK) (2025). AI adoption survey report 2025 (opens in a new tab) ↑

  3. NHS (2026). Carbon monoxide poisoning (opens in a new tab) ↑

  4. legislation.gov.uk (The National Archives) (1996). The Gas Safety (Management) Regulations 1996, regulation 7 (opens in a new tab) ↑

  5. Department for Science, Innovation and Technology (GOV.UK) (2024). Public attitudes to data and AI: Tracker survey (Wave 4) report (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