It is 11:40 at night. A message lands in your Instagram inbox: "hey, is the beige one still in stock?" Your auto-reply fires back inside a second: "Thanks for reaching out! We'll get back to you soon."
Technically, you replied. You also just handed that person eight hours to go find the beige one somewhere else.
That gap between answering and selling is the whole argument here. Most tools marketed as an AI sales assistant are support bots in a sales jacket. They resolve questions beautifully. They close nothing. And the difference has almost nothing to do with how smart the model is. It comes down to what the assistant does with the last line of every single message.
A support bot ends the turn. A sales assistant ends with a hook.
Look at the shape of two replies to the same question.
Support bot: "Yes, the beige one is in stock." Full stop. Conversation over, ball back in the customer's court, at 11:41 at night, with nine other tabs open.
Sales assistant: "Yes, beige is in stock, sizes 38 and 40 left. Which one do you take?"
Same fact. Same speed. But one of them has a buyer who is now typing a number instead of closing the app.
I keep watching sellers benchmark assistants on accuracy alone. Did it get the price right? Did it know the fabric? Accuracy is the floor, not the product. The number worth staring at is how often a reply moved the conversation one step forward. One step forward looks like:
- a size, a colour, a quantity
- a delivery city or an address
- a payment link opened
- a "yes, send it"
Everything else is customer service. Useful, cheap to run, and not why you bought the thing.
Qualify in two questions, not seven
Long qualifying scripts have killed more DM sales than high prices ever did. The seller (or the bot) asks what the customer is looking for, then their budget, then the occasion, then their size, then their city, and by question five the person has decided this feels like filling out a form at a bank.
Two questions. That's the budget. And they should be the two whose answers actually change what you recommend.
For a clothing page, that's usually size and what they'll wear it for. For skincare, skin type and what they've already tried. For furniture, room dimensions and delivery city. Notice what's missing: "What's your budget?" Asking a stranger their budget in the first minute reads as a test they might fail, and most people answer it defensively or not at all. Let the recommendation carry the price instead.
The other reason to keep it short: a big share of first-contact DMs are not buyers at all. They're browsers, price-checkers, people who tapped the wrong story. I've written before about what first-contact DMs actually look like when you sort them by intent, and the practical lesson is to spend very little effort on message one and a lot on message three.
The recommendation is where the money is made or lost
An assistant that lists nine products is a search box with manners. An assistant that names one, with a reason, is doing sales.
Say someone asks for a gift under $60. Here's the weak version: "Here are our products under $60: [nine items]". And the version that sells: "For a gift, the ceramic set at $54 is the one people come back for, and it ships in a box you don't need to wrap. If you want something smaller, the candle at $29 is the safe pick."
Two options, one clearly recommended, a reason attached to each. That's how a good shop assistant talks in a physical store, and it works for exactly the same reason: you removed the decision fatigue without removing the choice.
This only works if the assistant is reading your real catalogue. Prices, variants, live stock. An assistant guessing from a vague product description will eventually promise a colour you stopped carrying in spring, and you'll find out when the customer is angry. Feed it the catalogue, mark what's out of stock, and give it explicit permission to say "we don't have that, here's the closest thing."
Honestly, "we don't have it, but this is close" is one of the highest-converting sentences in any DM inbox. Bots almost never say it.
"Too expensive" is not the end of the conversation
Watch what most assistants do when a customer pushes back on price. Option one: they repeat the price, slightly more politely. Option two, and this one is worse, they instantly produce a discount code.
Drop a code the second anyone hesitates and you've taught your entire audience that hesitation is a coupon machine. Within a month, everyone hesitates.
The third path is to answer the question hiding under the objection, which is almost never "lower the number" and almost always "convince me this is worth it." Cost per use. What's included that competitors charge extra for. Warranty. Free returns. Whether shipping is already in the price.
Then give the assistant exactly one lever and let it pull that lever once: free shipping over a threshold, a bundle price, a first-order discount. One authorised lever, used deliberately, beats a discount reflex. If you want the actual scripts for this moment, I went deep on what to say after a customer tells you it's too expensive.
The rule that matters: an assistant must never invent a discount you didn't give it. Ever. That's not a sales philosophy, that's your margin.
Ask for the order
Here's the part that separates the tools. Most assistants, even the good ones, will happily answer questions until the sun comes up and never once ask for the sale.
The moment of yes has a short half-life. Someone who says "okay I'll take the beige, size 38" at midnight is a customer for about four minutes. If your reply is "great! please visit our website and search for the product", you've converted a buyer into a task on their to-do list. Tasks don't get done at midnight.
A closing message should carry all of it in one place:
- the exact item and variant, spelled out, so there's no ambiguity later
- the total including shipping, not the sticker price
- the payment link, right there in the chat
- one line about what happens after they pay, and when it ships
Three taps from yes to paid. Every extra tap is a place where the sale falls out.
The follow-up is the cheapest revenue you own
Most conversations don't end in a no. They end in silence. The customer got interrupted, the baby woke up, they meant to come back.
One follow-up, roughly a day later, referencing the actual item, brings a chunk of those back. "Still want me to hold the beige 38?" works. "Just checking in!" does not, because it puts the work of remembering back on the customer.
Same logic for the out-of-stock conversation. If someone wanted something you didn't have, that's a name on a list, and a message the day it lands. Sellers sit on those lists for months and never send the message.
Where average order value actually comes from
This is the part vendors oversell, so let me be boring and specific.
An assistant lifts your basket in one narrow window: after the customer has committed, before they pay. Not before. Suggest a complement while they're still deciding on the main item and you've added a second decision to a person who hasn't finished the first one. Suggest it after "yes, I'll take it" and you're doing what every good cashier does.
The size of the lift is modest and real. If your average order is $70 and one buyer in five accepts a $14 complement, your average order becomes $72.80. Call it 4%, on revenue you were already getting, with no extra ad spend. That compounds quietly across a year. It does not double your business, and anyone telling you it will is selling something.
The attachment has to be genuine, too. The strap that fits that watch, the refill for that bottle, the second pair at a bundle price. Random cross-sells read as desperate. There's more on getting the timing right in how upsell and cross-sell actually work inside a chat.
A five-minute test for whatever you're running now
Open your own page from a second account and send these five messages. Don't be gentle.
- "Is this available?" Does the reply end with a question, or a period?
- "How much?" Does it give the price plus what's included, or a bare number?
- "That's too expensive." Does it discount immediately? Fail.
- "Okay, I'll take it." Count the taps between that message and a payment page.
- Then go silent. Does anything at all happen tomorrow?
Anything failing three or more of these is a support bot, no matter what the pricing page called it. That's fine, as long as you know what you bought and stop expecting it to move your revenue line.
If you'd rather have one assistant that reads your real catalogue, answers on Instagram, WhatsApp, Telegram and your site in your own voice, asks for the order, and hands you the conversation the moment it gets out of its depth, that's the thing Vardast was built to be. Connect a channel, point it at your products, and watch a week of DMs before you decide.
Back to 11:40 at night. The beige one, size 38. The only question that matters is whether your assistant asked which size, or thanked her for reaching out.