AI Chatbots for Websites: What a Chimney Sweep Should Let Them Handle

A chimney sweep's website chatbot should handle booking first, published prices second and safety advice never. It can take booking requests against a diary it can read and read your price list back. It must never judge whether a flue is safe, and any hint of carbon monoxide triggers a fixed emergency message and an alert to a person.

The autumn rush is why the line matters. In the first cold weeks, enquiries arrive in the evening from people who may be messaging several sweeps at once, and a quick reply can decide who gets the job. A bot that answers in seconds helps you there. A bot that answers a worried customer's question about smoke in the living room, however kindly, creates the kind of risk in this trade that cannot be undone.

Follow me on Instagram@sagnikteaches

Booking, prices and safety: three jobs, three different answers

These three kinds of question reach a sweep's website every autumn, and they deserve different treatment because the cost of a wrong answer is so different.

Connect on LinkedInSagnik Bhattacharya

The difference comes down to one line: the bot reads and collects, and the sweep looks and decides. Everything on the first two rows is information that already exists in your diary and on your price list. Everything on the third row depends on seeing, smelling or measuring something in a house the bot cannot see. That is why it gets no discretion at all, however sensible a guess it might make.

Subscribe on YouTube@sagnikteaches

Autumn rush: take booking requests without overbooking the van

Bookings suit a chatbot because the questions are the same every time and the answers sit in your diary. The risk is the diary itself. A bot that offers dates from memory or from an old sheet will promise slots that do not exist. Keep one line per day in a sheet the bot can read ("Mon 3 slots, Tue 0, Wed 2") and update it every evening. Put two rules on it: the bot states the "as of" date whenever it offers a slot, and a sheet more than two days old produces no dates at all, only "I'll take your preferences and reply by tomorrow evening".

A realistic slip shows why. Two visitors write within the same minute, and the bot tells both that Tuesday has one slot left. Both say yes. The sweep rings one of them at 7 am the next day to apologise, and the one who gets the apology reviews the firm badly. The fix is to have the bot take requests instead of confirmations ("I've asked for Tuesday afternoon for you; you'll get a confirmation text by 7 pm"), with confirmations sent in a batch each evening, or to use a booking tool that locks the slot the moment one customer takes it. The limits of a chatbot sitting on a booking page are covered in what a chatbot on a booking page can and can't do.

Now the sum. Say a sweep who gets about 12 booking messages in a quiet week gets 70 in the weeks after the first cold snap, and that each takes about 4 minutes by hand (read it, check the diary, reply, note it down). That is 280 minutes, nearly five hours, usually done at night after a day on roofs. Suppose the bot completes the conversation for 6 in 10 of them, so the sweep only confirms a one-screen summary in about 1.5 minutes, and handles the other 4 in 10 by hand at 4 minutes each. All figures are assumptions to replace with your own.

The 105 minutes saved is useful and modest. The bigger gain is the evening messages. If 1 in 10 of those 25 would otherwise have booked the sweep who replied first, that is two or three jobs a week, and at an assumed $90 a job it is $180 to $270 a week. If your diary is already full for the next three weeks, the better use of effort is evening out the rush: autumn posts and offers written with AI covers filling the quieter weeks instead.

Appliance, fuel and last sweep date: the questions that size the job

The three facts you need before a visit are what the appliance is, what it burns and when it was last swept. They decide how long to book, what to bring and whether the job is one the bot should book at all. Ask them one at a time, with buttons wherever possible, because "stove" can mean four different things and a free-text box gets "the old one". This is an illustrative routing sheet. Time allowances are assumptions, and the sweep should replace every cell.

The last-sweep date goes on the job record for the sweep, and the bot passes no judgement on it. It never says "that's overdue" or "that's fine", because what a gap means depends on the fuel and the appliance, and your own recommended frequency is for you to state. If you want the bot to mention it, give it one approved sentence ("We recommend sweeping [your own guidance by appliance and fuel]") and let it read that back and nothing else. The date also helps the sweep to plan: a flue that has not been swept for years needs a longer slot.

One more question belongs in every booking flow, and it is the one that stops a routine booking from hiding a problem: "Has anything unusual happened since the last sweep, such as smoke or fumes coming into the room, a smell, soot or staining on the wall or ceiling, anything falling down the chimney or a noise from the flue?" An answer of "no" continues the booking. Any other answer ends the routine flow, as in this illustrative exchange.

Visitor: Last sweep was two winters ago. Smoke has been coming back into the room when the wind is up.
Bot: Thank you for telling me. I can't judge that from here, so I have passed it to [name], who will phone you [by a time you really keep]. Please read the safety message below.
[The emergency message appears in full.]

Note what the bot did not say: it did not say that wind-driven smoke is common, it did not suggest what might be blocked, and it did not offer a slot for next week. A person now decides, and the emergency message is on screen in case someone is already unwell. Here is the record a routine booking leaves for the sweep, filled in.

BOOKING REQUEST   Routine sweep
Customer: [first name], prefers text message
Appliance: wood burner, liner "not sure", one flue
Fuel: seasoned wood, used most evenings
Last swept: "about 14 months ago" (customer's words)
Anything unusual: no
Access: two-storey house, ladder from the front
Preferred: Tuesday or Thursday afternoon
Told by the bot: request only; confirmation text by 7 pm;
  no view given on the appliance or the flue

Price questions: read the list back and flag the extras

"How much?" is the question that follows "can you come?" in nearly every chat. The bot can answer it safely if it reads your published list and says what could change the price. It cannot answer "how much for my chimney?" because it has not seen the chimney. Here is the approved wording, with brackets for your own prices and conditions.

PRICES (approved wording)
Sweep, one flue, easy access:        [$ price]
Each additional flue:                [$ price]
Camera inspection (separate visit):  from [$ price]
Certificate:                         [included / $ price]
The price can be higher where the flue has a heavy build-up or
access needs special equipment. We confirm anything extra before
we start work, and we never add it afterwards.

The final two sentences matter as much as the list. They let the bot answer the awkward follow-up ("what if it hasn't been done for years?") with something true, and they stop it improvising a surcharge. An illustrative exchange:

Visitor: How much for a stove sweep? It's been about five years.
Bot: A sweep for one flue with easy access is [$ price]. The price can be higher where there is a heavy build-up or access needs special equipment, and we confirm anything extra before we start. Would you like me to take your details for a booking request?

An ungrounded bot does something different. Asked the same question, it says "a stove sweep usually costs around [a figure it found online]". The customer writes it down, the sweep's own price is higher, and the visit starts with a disagreement about a number that never came from the business. For the same reason the bot never offers discounts, never matches another sweep's price, and never quotes for repairs, liners, pots, cowls or rebuilds. Those need a person who has seen the chimney. If a visitor asks a second time for a price the list does not cover, the bot hands over and says so plainly.

Safety questions: the bot describes, it never diagnoses

Customers will ask safety questions in a chat window because it feels like asking a friend. Write the bot's replies in advance, and treat the table below as a starting point for your own list. The replies are for the sweep's technical lead to approve, ideally with a check against the trade body's guidance and your insurer's requirements.

The pattern is the same in every row: say what the bot cannot do, offer the one thing it can, and hand over. It is tempting to let the bot be helpful, since a customer who asks "is this normal?" wants a reassuring answer. A reassuring answer is exactly what makes the bot dangerous, because a customer who has been told it is probably fine will believe it, and the next visit may be days away. Short, general care text that you wrote and approved, such as which fuel you recommend, can be quoted word for word. Applying it to a particular customer's problem is the line the bot does not cross.

Suspected carbon monoxide: the triggers, the message and the alert

Public-health guidance describes carbon monoxide as colourless and without a smell, so nobody can tell it is present by noticing it. The symptoms it lists include headache, dizziness, feeling or being sick, weakness, tiredness and confusion, chest and muscle pain and shortness of breath, and it notes that symptoms may come and go and can ease when a person leaves the affected area. That last point matters for a chatbot. A customer who types "I get headaches in the evening when the stove is lit, but they go at work" is describing a pattern that should fire the emergency path, and a reply that treats it as a booking enquiry is a serious failure.

Build the trigger list from words customers really use. The sweep should add any others from past call-outs.

Of everything the chatbot says, only this message touches safety, so nobody leaves it to the AI. Write it with a qualified person in your trade, paste it into the bot as text the model is not allowed to rephrase, and reuse the identical words on your phone greeting and in a line pinned beneath the chat window. Draft wording follows. Your technical lead should check every line against current public-health guidance and against what your trade body and insurer advise.

If a carbon monoxide alarm is sounding, or anyone has a headache,
dizziness, sickness, weakness or confusion that could be linked to a
fire, stove or boiler: stop using the appliance if you can, get
everyone outside into fresh air, and call the emergency services or
a medical helpline now. Do not go back in until you have been told
it is safe.
I have also alerted [name] at [business]. [They will phone you by
a time you keep.]

After the message appears, the bot stops answering questions about the problem. It does not diagnose, does not reassure, and does not offer a booking. If the visitor keeps typing, it shows the message again. The message is the safety net, because it works even if nobody reads the alert for an hour. The alert is the follow-up, so it needs to reach a person on a phone, not only a queue read the next morning, and it should carry the fields below. A filled-in example, illustrative:

URGENT HAND-OFF   18:42   Possible carbon monoxide
Customer: [first name], callback number given in chat
Said (verbatim): "My CO alarm went off twice this evening and my
  head hurts. Wood burner was lit."
Appliance: wood burner (from an earlier booking message)
Emergency message shown: yes, 18:41
Anyone unwell mentioned: yes, the customer
Photos: none
Promised: a call by [time you keep]
Transcript: [link]

Four phrases never appear in any answer the bot gives on this path: "that's probably nothing", "it's most likely a false alarm", "open a window and carry on" and "we can look at it on Thursday". Each is a diagnosis or a reassurance. Test the path before launch with the calm version, the panicked version and the version with typos, because the person who needs it most is the person least likely to type neatly.

Three sweep businesses, three different chatbot scopes

How much of this a firm needs depends on what it sells and who answers the alerts. These three illustrative businesses sit at different points on that scale.

For the solo sweep the first question is whether a chatbot is worth having at all. My rule of thumb is that below about 30 enquiries in a peak month, a clear booking page and a short FAQ page do the job, and a bot adds a layer to maintain. Above about 100, it saves real evenings. The software cost is small either way. On the vendors' own pages in November 2026, Botpress listed a free plan with 100 conversations a month and Tidio's free plan included 50 Lyro AI conversations, once. Paid options started at $32.50 a month billed yearly ($39 month to month) for Tidio's Lyro AI Agent plan with 50 conversations a month, or $40 a month for Chatbase's Hobby plan ($32 billed yearly, 700 message credits, no free plan). A peak week of 70 chats can use up the free allowances, so price the paid tier before you trial. Chimney sweep software with reminders and online booking covers the diary and reminder side that a chatbot sits in front of.

The two-crew firm's extra rule is routing. A camera inspection produces a report, and a chatbot should never describe the findings of one. That job stays with the person writing it, and the report-writing side is covered in turning camera findings into plain English. For the retailer, the line is between selling and advising: the bot can say that installation needs a visit, but it never says what a particular flue or room will accept. A related seasonal pattern, in which bookings are easy and the risk sits in triage, appears in an HVAC chatbot that books tune-ups and triages without overpromising.

Whether anybody can answer an alert out of hours is the other factor. If the sweep cannot, the emergency message still does its job on screen, and the bot should say plainly when a person will phone, for example "during working hours, by 9 am". It should never imply a night-time response the business does not provide.

Three questions that decide whether the bot may answer

When you come to add a new answer, and in the first autumn you will, put it through three questions. Is the answer already in your own records: the diary, the price list or text you approved? Could a wrong answer put anyone at risk, or cost the customer more than the visit? Does it depend on seeing, smelling or measuring something? The bot may answer only if the first is yes and the other two are no. Anything else goes to a person, and any hint of fumes or illness gets the emergency message first.

Twenty messages to test before the first cold snap

Start with the greeting, because it sets expectations before the first question: "Hello, this is [business]'s automated booking assistant. It can request a sweep and give you our prices, but it can't say whether a fire, stove or flue is safe." Then write 20 test messages, ten copied from your own inbox with the typos left in and ten taken from the trap list underneath. The bot passes only if all three hold: every mention of fumes, illness or an alarm shows the emergency message and fires the alert, without a single miss; no reply contains a diagnosis, a reassurance or a price missing from your list; and 9 or more routine bookings in 10 leave a complete record.

Re-run all twenty after any change to the price list, the diary link, the emergency wording or the software itself, because a vendor update can change how keywords match. If an update quietly stopped "alarm" from triggering the emergency path, the twenty messages would expose it before a frightened customer did. During the first two weeks of the rush, spend ten minutes each evening on every urgent hand-off and a handful of other chats, noting any question that left the bot stuck. Each note becomes a new price line, a longer trigger list or a clearer "I can't help with that", dated by whoever owns that part.

The same owner handles the emergency wording once a year. Before each autumn, re-read it, check it against the guidance your trade body and insurer currently give, and run the test again. A chatbot is quickly forgotten after launch, and the emergency message is the one piece of it that should never be.

Smaller questions sweeps ask about a website chatbot

Can the chatbot take a deposit or payment?

Only by sending a link to the payment or booking tool you already use, so card details never pass through the chat. A deposit can cut no-shows in the autumn rush, but state the refund terms in your own words on the booking page and let the bot quote them word for word. It should never invent a policy, and it should never promise a refund.

What about customers who write in another language?

Answer in that language only if you have an approved translation of the emergency message and your booking wording, checked by someone fluent. Otherwise have the bot reply in your main language, say a person will follow up, and still show the emergency wording in both. A machine translation of safety wording that nobody has read is a risk you do not need.

Can it send next year's reminder?

The reminder belongs to your booking or diary software, which holds the date of each visit and can send a message later. The chatbot's part is to record the appliance and the customer's last-sweep date accurately at booking, labelled as their own recollection. The date that drives next year's reminder should be the date of your visit, entered by you after the job.

What should it do with photos customers send?

Accept them, attach them to the record and say who will look. It should not describe, interpret or reassure about what a photo shows, whether that is soot staining, a crack or a bird's nest. A confident misreading of a photo is exactly the diagnosis you ruled out, and the sweep seeing the photo before the visit is the real benefit.

Further reads

• AI Chatbot Software for Support: Fire Alarm Servicing Questions Answered — Another life-safety trade deciding what its chatbot may answer.
• AI Chatbot for Plumbers: Answer Questions, Collect Photos, Refuse Jobs — Answer questions, collect photos and refuse jobs on a trades site.
• AI Chatbot Disclosure: What to Tell Customers at the Start of a Chat — Wording for telling visitors they are talking to AI.
• How to Test a Customer Chatbot Before It Goes Live — A general method for testing a bot before customers see it.
• How to Measure Whether Your AI Chatbot Is Actually Working — The numbers that show whether the chatbot is helping.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: public-health guidance on carbon monoxide symptoms and response (read November 2026); Botpress, Tidio and Chatbase pricing pages (read November 2026).