It is 11:40 at night and a seller I know is sitting on the edge of her bed, thumbing through a WhatsApp inbox with 63 unread chats. Around fifty of them say some version of the same thing: "how much?" Eleven ask whether the blue one is still available. Two are real buyers with their card already out, and they wrote four hours ago.
She answers the two. One has already bought somewhere else.
The gap between the moment a customer asks and the moment you reply is where a WhatsApp AI bot either earns its keep or quietly wastes your money. Most of them waste it. I have watched enough seller inboxes to have strong opinions about why, so let me save you the tuition.
Write the job description before you shop
Everyone starts by comparing tools. Start instead by writing down what the thing has to do on an ordinary Tuesday. For a shop selling through WhatsApp, the list is short and unglamorous:
- Answer "how much?" with the actual price of the exact variant the customer named, not a price range.
- Say whether that variant is in stock right now.
- Quote shipping to the customer's city, with a delivery window in days.
- Take the order: item, size, quantity, name, address, phone, payment method.
- Know the moment it should stop talking and hand you the chat.
Five things. A tool that does four of them beautifully and fumbles the fifth is not 80% of an assistant. It is a machine that loses one customer in five, at scale, all night, while you sleep.
Notice what is not on that list: personality, emoji, clever greetings. Buyers do not message a shop hoping for a chat. They want a number, a yes or no, and a way to pay.
Grounded answers, or invented ones
Here is the single technical thing that decides whether your bot is an asset or a liability: where its answers come from.
A bot wired to your product data looks up the black leather crossbody, size medium, and reports 1,890,000 toman, four left, ships tomorrow. A bot running on a pile of general instructions writes something that sounds like a price. Those two behaviours look identical in a demo and nothing alike after a week of real traffic, because the second one will eventually quote a discontinued price to someone who screenshots it.
So when you evaluate anything, feed it your ugliest product. The one with seven colour options, two of them out of stock, one with a different price. Then ask about it the way a customer would: "the mustard one, do you have it, how much with post to Shiraz?" A grounded assistant answers all three parts in one message. A shallow one answers the easy part and ignores the rest.
This is also why the setup work matters more than the tool's feature list. Getting your catalog into the assistant properly, with variants, stock, shipping rules and the return policy you actually follow, is most of the job. I wrote about that in detail in training your AI assistant on your product catalog, and I would read that before buying anything.
Order-taking is where most bots give up
Answering questions is the demo. Taking the order is the product.
Watch what a mediocre bot does after a perfect answer: it says "let me know if you'd like to order!" and stops. The customer now has to restart the conversation, restate what they want, and find their own way to pay. Roughly half of them will not bother.
A serious assistant closes the loop. It confirms the item and quantity back to the buyer, asks for name and full address in one message rather than five, offers the payment options you actually accept, and writes the finished order somewhere you can see it. On a busy campaign day the difference between those two behaviours is the entire campaign.
One detail people underrate: the confirmation message. Repeating the order back in plain text ("1 x mustard crossbody, 1,890,000 toman, post to Shiraz, 2 to 3 days, cash on delivery") kills a huge share of the "but I thought it was the other colour" disputes that eat your afternoon a week later.
The app or the API: decide by volume, not by pride
Every seller eventually asks which version of WhatsApp they should be on. The honest answer is that it depends on how many conversations you have, and nothing else.
The free business app is fine for a shop doing a few dozen chats a day. It runs on one phone, it has labels and quick replies, and you can connect an assistant to it. Its ceiling is real, though: one device, one number, no proper multi-agent inbox, and rate limits that show up exactly when a post goes viral.
The official business platform (the API) is the other road. Multiple team members in one inbox, message templates, delivery and read reporting, no phone sitting in a drawer as a single point of failure. It also costs money per conversation and needs a verified business behind it. If you are doing hundreds of chats a day, that cost is trivially worth it. If you are doing twenty, it is a subscription to complexity. I laid out the crossover point in when to upgrade to the WhatsApp Business API.
My rule of thumb: stay on the app until either two people need the same inbox at once, or you start losing messages during sales. Then move.
The rules, briefly, because breaking them ends the channel
WhatsApp is not Instagram. It is stricter, and the penalty is not lower reach. The penalty is your number being blocked. A few things are worth knowing before you automate anything:
- Customers open the door, not you. Free-form replies are allowed within a 24-hour window after the customer's last message. Outside it, you can only send pre-approved templates.
- Opt-in is not optional. Numbers scraped from comments or bought in a spreadsheet are the fastest route to a ban. Consent has to be real.
- Blocks are the scoreboard. Your number carries a quality rating driven mostly by how often people block or report you. Send less, send better.
- Let people leave. An easy way to stop messages costs you a few contacts and saves the channel.
None of this makes automation risky. Bulk blasting to strangers makes it risky. Answering people who messaged you first is the safest thing you can do on WhatsApp, and it is where the sales are anyway.
Build the exit before you build the bot
An assistant that never escalates will eventually argue with an angry customer about a refund. Decide up front what it must never handle alone: money disputes, damaged deliveries, anything involving a promise you made personally, and any message where the person asks for a human twice.
Good escalation is not a failure state. It is the thing that lets you trust the other 85% of conversations to run without you. I broke down the specific triggers in when AI should hand off to a human.
What to look at after two weeks
Not "messages handled". That number always looks good and means nothing. Look at these instead:
- First-reply time, especially between 9pm and midnight, which is when buyers actually message.
- Share of chats resolved without you, the real automation rate.
- Orders started versus orders completed, which exposes a bot that answers well and closes badly.
- Escalations per hundred chats, and whether the same question keeps causing them. That is a catalog gap, not a bot problem.
- Block rate. If it moves at all, stop and look at what you are sending.
Two weeks of those five numbers will tell you more than any vendor comparison table.
If you want this running without stitching it together yourself, that is the job Vardast does: it connects to your WhatsApp along with your Instagram, Telegram and website, answers from your real catalog and prices in your own voice, takes the order, and passes the chat to you when it should. One assistant, one memory of the customer, every channel.
The seller with 63 unread messages did not have a traffic problem. She had a reply problem. Most WhatsApp shops do, and it is the cheapest one to fix.