Most conversations with a business chatbot are ordinary: prices, hours, bookings. But every service business also receives the occasional message that is not ordinary. A burst pipe, a gas smell, severe tooth pain, a client in crisis, a customer who says they are going to hurt themselves. What the assistant does in those moments is the true test of whether it is safe to put in front of the public. An assistant that carries on with a booking flow, reassures someone who needs help or promises a technician it cannot send can cause real harm and real liability. This guide explains how to design emergency handling: the principles, how to write the triggers and the wording, how to alert people, what to do out of hours, how to test, and how to review after a real incident. It applies to any business, and links to industry-specific guides where relevant.
This is general information, not legal, medical or safety advice. Emergency handling must be written and approved by people qualified for your industry, and must follow local rules and your insurance requirements.
The principles
- Safety before sales. In an emergency, drop the booking flow and help.
- Route, do not diagnose. The assistant is not a professional and must not assess severity or give treatment or repair instructions.
- Give approved wording only. Write it in advance with qualified people.
- Get a human involved fast. Alert someone who will act, by more than one route.
- Be honest about what you can do. Never promise arrival times or outcomes you cannot guarantee.
- Direct to emergency services when in doubt. It is always better to over-refer than to under-refer.
These principles are simple. The work is in turning them into rules the assistant follows reliably.
Write the triggers
Start by listing the emergencies your business could meet, and the words people use. Ask your staff: what are the frightening messages you have received?
| Business | Examples of urgent messages |
|---|---|
| Plumbing and heating | Flooding, burst pipe, gas smell, no heat with vulnerable people at home |
| Electrical | Burning smell, sparks, smoke, shock, exposed wires |
| Dental and medical practices | Severe pain, swelling, bleeding, trouble breathing, injury |
| Roofing and building | Water pouring in, structural damage, tree through the roof |
| Pest control | Stings with a reaction, nests near entrances with allergic occupants |
| Legal and professional | Deadlines within days, safety risks, distress |
| Any business | A person who says they are in danger or may harm themselves |
Include misspellings, slang and phrases in the languages you serve. 'my kitchen is flooding', 'water everywhere', 'smoke coming from the plug', 'i cant breathe' all need to be caught.
Two tiers of response
| Nivå | När | What the assistant does |
|---|---|---|
| 1: danger to life or safety | Anything suggesting immediate risk | Tells the visitor plainly to contact emergency services, gives short safety instructions you approved, alerts your on-call person, and stops the normal flow |
| 2: urgent but not life-threatening | A situation that needs fast professional help | Gives approved interim guidance, gives your emergency number or route, captures name and phone, alerts the team, and is honest about timing |
Keep the wording short, calm and specific. The person reading it may be frightened.
Wording templates for your team to adapt
| Situation | Mall |
|---|---|
| Tier 1 | This sounds serious. Please call emergency services now. If it is safe, move away from the danger. I have alerted our team, but please do not wait for us. |
| Tier 2, open | I am sorry this is happening. Please call us right now on [number], as we are open and can help. I have alerted the team, and they will look for your call. |
| Tier 2, closed | I am sorry this is happening. We are closed, but our emergency line is [number]. I have alerted our on-call person, and they will reply as soon as they can. If it gets worse, or anyone is at risk, call emergency services. |
| Personal crisis | I am concerned about how you are feeling. You deserve support from a person. If you are in immediate danger, please call emergency services. In [country], you can also contact [crisis line]. I have asked a member of our team to reach out. |
These are drafts. Your industry and region determine the right words, the right numbers and what is appropriate to say.
Alerts that reach a person
- Instant email to a monitored address with the whole conversation.
- A text or push notification to the on-call phone; see webhook automations.
- A second route if the first is not acknowledged within minutes.
- A named person on duty at all hours you promise coverage.
An alert that nobody sees is worse than no alert, because it creates false comfort. Test the route at night, on weekends and when the usual person is away. Crescelo can hand any conversation to your team with notifications and full history; see the human handover feature and live human handover flow design.
Out-of-hours honesty
If you have no out-of-hours cover, say so and give the visitor a real alternative: an emergency service, a utility's emergency line or a local provider. If you have on-call cover, say who and how quickly, and only what you can deliver. Never tell someone 'we will be there in 30 minutes' unless you can. See after-hours appointment booking.
What the assistant must never do
- Diagnose or reassure ('that sounds like nothing').
- Give repair, medical or treatment instructions.
- Continue a sales or booking flow after a danger signal.
- Promise arrival times or outcomes.
- Ask for sensitive details or images in an open chat.
- Argue with a visitor who says they are in danger.
Test the rules
- Write thirty realistic emergency messages for your business, with typos and casual phrasing.
- Add ten borderline messages that should not trigger emergency wording, such as 'do you do emergency callouts?'.
- Run all forty in the Testing Studio.
- Check the response wording, the alert and the stop of the normal flow.
- Test at night that the on-call person receives the alert.
- Repeat after any change to the assistant's instructions, and every quarter.
| Test type | Pass condition |
|---|---|
| Clear emergency | Tier 1 wording and alert every time |
| Urgent but not life-threatening | Tier 2 wording and alert |
| Not an emergency | Normal answer; no false alarm |
| Other language | Understood and handled |
| Repeated messages | Assistant stays on the emergency path |
Review every real incident
After any real emergency conversation, review it within a day: what did the visitor say, what did the assistant do, how fast did a person respond, what happened next and what would you change? Adjust triggers and wording, and record the change. Real cases teach more than any test list.
A sample conversation
| Speaker | Meddelande |
|---|---|
| Visitor | theres a burning smell coming from the fuse box |
| Assistant | This could be dangerous. Please leave the area, do not touch anything electrical and call emergency services now. I have alerted our on-call electrician and they will contact you as soon as possible. What is your name and phone number? |
The assistant did not offer a booking, did not diagnose and did not promise a time. It gave safety guidance, alerted a person and asked for contact details.
Common mistakes
- Emergency rules written by marketing, not by qualified staff.
- Alerts sent to an unwatched inbox.
- No testing at night.
- Triggers that only catch formal wording.
- Wording that promises more than you can deliver.
- Never reviewing real incidents.
Frequently asked questions
Can a chatbot handle emergencies?
It should recognise urgent wording, give approved safety wording, alert a person and never diagnose or reassure. It is a route to help, not the help itself.
Who should write the emergency wording?
Qualified people in your industry, such as your clinical lead, senior technician or safety officer.
What if nobody is on call?
Say so honestly and direct the visitor to emergency services or another appropriate route.
How do I test emergency handling?
Write realistic urgent and borderline messages, run them in the testing area, and check the wording and alerts, including at night.
Should the assistant continue booking after an emergency signal?
No. Stop the normal flow and focus on safety and alerts.
How often should I review the rules?
After every real incident and at least every quarter.
See how the AI assistant hands over rather than guesses, test your rules in the playground and read dental emergency triage and electrician service call scheduling for industry examples.
Further reading
- Google Search Central: creating helpful, people-first content: What Google looks for in useful pages.
SVARA ALLA BESÖKARE. BOKA DEM SOM ÄR REDO.
Lägg in Crescelo på din webbplats på 15 minuter
Träna din assistent på dina prislistor, dokument och tjänstevillkor. Aktivera välgrundade svar dygnet runt och bokning av tider direkt i chatten utan krockar.