In this article (21 sections)
In short: an AI agent can take orders over WhatsApp. It understands what the customer writes or says ("three cases of the 500 g pack and one of the large one"), matches it against your catalog and prices, confirms the order and enters it in your system, without anyone copying it by hand. It works well when the catalog is clear and the customer gives a final confirmation. It fails when the catalog is ambiguous or the agent can invent products. The key isn't the AI. It's that the order arrives clean at the place where it gets prepared.
This article explains how it's built, what data it needs, where the typical errors happen and what it costs to run.
The problem it solves
If you take orders over WhatsApp today, you probably recognize some of this:
- Someone on the team answers, writes the order in a notebook or spreadsheet and then enters it into the system.
- Orders arrive after hours and get answered the next morning, with the customer already impatient.
- The same customer sends the order in four messages, with photos and voice notes.
- Transcription errors are discovered after the order has already shipped.
The agent doesn't change the channel, which is the one your customers already use. It changes what happens on your side.
What the conversation looks like
- The customer writes as usual: text, a voice note or a photo of a list.
- The agent identifies who it is by their number and pulls up their history, price list and payment terms.
- It interprets the order and matches it against your catalog. If something is ambiguous, it asks: "The 500 g or the 1 kg one?"
- It checks stock and credit limit before confirming.
- It summarizes and asks for confirmation: products, quantities, total price, delivery date.
- It enters the order in your system and tells the customer it's been recorded.
Step 5 isn't optional. An order the customer didn't confirm in writing is the source of most later disputes.
What has to exist first
Requirement | Why it matters |
|---|---|
A clean catalog with clear names and codes | The agent can't guess which product you mean if three have similar names |
Prices per customer or per price list | If each customer has their own price, the agent has to look it up, not make it up |
Available stock in a system | So it doesn't confirm what you don't have |
A way to enter the order | A connection to your management system or, at minimum, a shared spreadsheet |
A WhatsApp Business number | Required to automate customer service |
If your catalog lives in your team's heads, that's the first job, before you think about AI. Something similar happens with distribution businesses, which we cover in Management system for distributors and wholesalers.
The channel: the WhatsApp Business account
To have an agent reply automatically, you use WhatsApp's business platform, not the regular app. Its interface lets you send messages (text, media and interactive messages such as order confirmations) and receive customer messages and delivery statuses on your server (WhatsApp Cloud API, Meta documentation). Third-party platforms are built on this same foundation.
What each message costs
According to Meta's documentation, you're charged per delivered message, with rates that depend on the message category and the customer's country; messages the customer sends to the business are free. Replies within the 24-hour customer service window, which renews with each customer message, aren't templates (WhatsApp Business platform pricing). What's worth looking at is how much of your traffic is replies and how much is notices you start yourself (for example, "your order is on its way"), because those use approved templates and are charged. The same page indicates that utility messages within the 24-hour window were free until October 1, 2026 and are charged from that date, so check the current rates for your country before budgeting.
Where it fails and how to avoid it
The agent invents products or prices
Avoid it by limiting the agent: it can only offer what's in the catalog and can only quote prices it looks up in your system. If a product doesn't exist, it says it can't find it and offers alternatives.
Ambiguous orders
"Send me the usual" works if the agent has the customer's history and asks for confirmation. Without history, it has to ask.
Voice notes and photos
Voice notes are transcribed and photos of lists are read, but with a higher error margin than text. That's why the summary before confirmation is mandatory.
Changes and cancellations
Define how long an order can be changed. Past that point, the agent hands the case to a person.
New customers
A customer who isn't registered needs minimal details (name, address, payment method). The agent collects them, but credit approval is usually done by a person.
When a person should step in
- Orders above a certain amount.
- Customers with overdue balances.
- Complaints, returns or shortages.
- Products outside the catalog or with special terms.
- Any moment the customer asks to speak with someone.
A good design shows the pending order on one screen, with the conversation next to it, so whoever reviews it can decide in seconds.
What changes in your operation
Before | With the agent |
|---|---|
Someone answers and writes it down | The order comes in already entered |
After-hours orders wait until morning | They're taken at any hour |
Transcription errors | Summary with customer confirmation |
Stock is checked afterward | Stock is checked when the order is taken |
No clean record | Each conversation is linked to its order |
An agent like this fits within what we describe in AI agents for business: what they actually automate.
Industries where it's used most
- Distributors and wholesalers, with repeat orders from regular customers.
- Restaurants and prepared food, with changing menus.
- Supplies and hardware, with many similar codes.
- Pharmacies and beauty distributors, with availability questions.
- Farm producers and suppliers, with seasonal orders.
In every one of them, what determines the result is the quality of the catalog and of the connection to the system, not the AI model.
What it costs
An agent that takes WhatsApp orders and enters them in your system is built custom. Atomix builds AI agents for business starting at USD 3,000 for the agent, plus monthly maintenance from USD 300, with a free working demo in 24 to 48 hours before we quote, and most projects reach production in under a month. Monthly WhatsApp and model usage is paid to the providers based on order volume.
Frequently asked questions
Can I use my current WhatsApp number?
It depends on how it's registered today. Often the number is migrated to the business platform, or a new one is used. Decide at the start and confirm the terms with your provider so you don't lose your contact history.
Does the customer know they're talking to an AI?
We recommend saying so in the first message. Customers are generally fine with an assistant that resolves things quickly and offers to connect them with a person when they want one.
What happens to orders the agent doesn't understand?
It hands them to a person with the full conversation and the partly built order, so nobody has to start from scratch.
Does it work if I have many products?
Yes, as long as the catalog is well named and coded. With thousands of products, the agent searches the catalog instead of memorizing it, and it asks when there are several matches.
How long until it's running?
It depends on the connection to your system and the state of your catalog. If your system can already receive orders from outside, it's fast; if that path has to be built first, that's the part that takes the longest.
If you want to see how it would take your customers' orders, book a call and we'll build a working demo at no cost.