TL;DR Automating quotes on WhatsApp means having an AI draft the reply and a person approve it before it's sent. In one real case, a home-maintenance company in Milan, the time to reach a quote dropped from about 30 minutes to 10-15, across 95 conversations handled in two months and $30 of total AI spend. This guide covers the mechanism, the numbers, and how to tell if your case is similar enough to be worth it.
Automating quotes on WhatsApp: how it works, what it costs, and how to tell if it's worth it
Automating quotes on WhatsApp means putting an AI inside the conversation with the customer, not instead of the person who answers but ahead of them. The AI drafts the reply, the person approves it: what is called a human-in-the-loop setup. In one real case (a home-maintenance company in Milan), this brought the time to reach a quote from about 30 minutes down to 10-15, across 95 conversations handled in the first two months.
I built that system myself, for a client, and it's the work I do: keep that in mind as you read. In exchange, the numbers below are the real ones, not an illustration. This guide uses that project as a concrete reference: what changes, what it costs, how long it takes to build, and a checklist for whether your case is similar enough to be worth it.
The technical version (architecture, decisions, code) is on the project page.
The problem persists even if you already use AI
If you already paste customer conversations into ChatGPT or Claude to get help writing a reply, you are doing what the owner in the reference case did: work the answer out in a browser tab, then paste it back into WhatsApp. That took him around half an hour per quote. That's his estimate, not a measurement, but the point stands either way: the AI was already there, it was just outside the channel.
The cost of staying outside the channel is the repeated copy-paste, a price list kept in someone's head, and no trace of how many requests come in or how long they actually take.
How it works, in practice
- The customer messages the company's WhatsApp Business number, the same way they would any contact, photos and voice notes included.
- An AI reads the conversation, looks at the photos, checks the company's price list, and drafts a reply.
- The reply shows up as a draft in a dashboard. It isn't sent.
- A person reads it: sends it as is, corrects it, or answers in their own words instead.
- Only then does the message reach the customer.
The price list, the jobs the company doesn't take on, and the AI's instructions are documents the company edits itself, with saved versions, so changing a price doesn't require touching the code.
What changed, in numbers
In the reference case, from 1 July to 1 September 2026:
| Before | After |
|---|---|
| ~30 minutes per quote (estimate) | 10-15 minutes (estimate) |
| Manual copy-paste into a browser AI | AI inside the channel, with review |
| No record of requests | 95 conversations tracked |
| - | 50 approved drafts, 42% hand-corrected |
| - | $30 of AI spend over two months |
The before/after time is the owner's estimate, not a measured figure: I'm including it because it's the number readers feel most, not because it's as precise as the others.
Why no message goes out unchecked
When a reply contains a price, a mistake isn't an annoyance: it's a commitment made to a customer that then has to be walked back. That's why, in this kind of system, every reply goes past a person before it's sent, not just the ones with a price in them.
In the reference case, of 50 approved drafts, 21 (42%) were corrected by hand. That isn't a sign the AI performs poorly. It's proof the check is doing its job: at that rate, skipping it would have sent twenty-one inaccurate quotes to real customers.
Checklist: does this make sense for your business?
Answer yes or no to these four conditions:
Customers reach you through a conversational channel (WhatsApp, Messenger, site chat), not just a form or phone call.
The reply takes judgement but follows rules: a price list, constraints, things you don't do.
Volume is regular enough to justify a tool, not sporadic requests that are all different from one another.
Whoever answers today is the same person running the business, and every half hour spent writing a quote is half an hour taken from everything else.
Zero: it's early. You need a process to automate before you need the tool.
One or two: worth a conversation, but the solution will need adapting.
Three or four yeses: the reference case's pattern likely applies to you.
What it costs and how long it takes
In the reference case, building the system cost fifteen working days, not full-time, from the first line of code to the first real customer. Running it cost $30 of Claude usage over two months, about $0.32 per conversation, plus ordinary web-application hosting.
The number to watch isn't the monthly one but the build number: it depends on how many rules and exceptions your process has, and that's the one worth discussing case by case.