WhatsApp assistant: replying to customers in 20 seconds (even on a Saturday at 22:00)

A company from Świdnica that installs air conditioning and heat pumps, 9 people, two installation crews and one service technician. The WhatsApp Business number was on every flyer and on every unit they'd installed, so customers wrote there about everything: "it's dripping", "how much is a service visit", "when are you coming", a photo of the remote with "what does this icon mean". The owner replied, between one site visit and the next, usually in the evening. Average response time: around three hours. In July, in season, up to a day.

We counted over three weeks: 412 messages, 38% of them after 17:00 or before 8:00. A customer who writes at 21:47 that his air conditioner is dripping will have messaged two other companies by morning.

What the assistant does

It reads the message (text, voice note or photo), works out what it's about and replies in around 20 seconds. For the reply it uses three sources: a database of typical faults that we wrote up with the technician in one afternoon, a price list with 24 items, and the crews' calendar. If a question goes beyond those three things, the assistant says plainly that it's passing the matter to a human, and does so with a summary of the conversation.

That Saturday conversation is real; I only changed the address. From the first message to the entry in Bartek's calendar, 19 seconds passed. On Monday at 13:30 Bartek was on Lipowa street.

What went wrong

The price that wasn't on the list

On day four a customer asked about installing air conditioning in a house with three rooms. The price list had the installation of one unit. The assistant "helped": it multiplied the single price by three and took off a discount nobody had given it. The customer got a figure PLN 1,400 lower than the real quote. The owner found out when the customer rang to complain, because "the message said otherwise".

The fix took an hour and is simple: the assistant doesn't quote any number it hasn't found in the price list or the calendar. Every question about a price outside the 24 items ends with "that's an individual quote, I'm passing it to the owner, she'll get back to you by the end of the day" and a task in her calendar. There hasn't been a single made-up amount since. We check it every week by reading 30 random conversations, because after that slip nobody wanted to take anyone's word for it.

Voice notes

Around 15% of messages are voice notes, often recorded in the car. At first the assistant didn't understand them and replied "please write it as text", which a few people ignored and sent another voice note. We added transcription. It works surprisingly well even with background noise, less well with street names; twice "Wałbrzyska" turned into "Wambierzycka". That's why the assistant always repeats the address back and asks for confirmation.

"Is this a bot?"

The first version pretended to be a person and signed off with a name. Two customers asked outright whether they were talking to a bot, and one took offence when he figured it out. Now the assistant introduces itself as "Klima-Term Assistant" in the first message and says you can type "human" at any time for the owner to take over the conversation. That happens in 6 out of 100 conversations. Fewer than I feared.

The customer who writes at 22:00 and expects installation tomorrow

No assistant will solve that one. But it gives the nearest realistic slot from the calendar straight away, so the customer doesn't wait until morning to find out it'll be a week anyway. Some of them would have written to a competitor in the meantime, who'd tell them the same thing in the morning.

What changed after six weeks

Efekt: Time to first reply dropped from around three hours to 20 seconds. The assistant closes 71% of conversations on its own, with a calendar entry or an answer from the fault database. Evenings and weekends now bring in around nine jobs a week; before, those hours produced two or three. The owner stopped replying from her phone after 21:00, which she says matters more than the jobs.

The build took 12 days, three of which went on voice notes and two on writing up the fault database with the technician. He's the one who knows that "it's dripping" is a blocked condensate drain 80% of the time, and "it's not cooling" is usually a dirty filter the customer can clean themselves.

What I need to do this for you

A WhatsApp Business number (Messenger works the same way), a price list in any form, a calendar the team actually keeps, and two hours with the person who knows the typical customer questions. I do the rest myself, in a week or two.