At 11:40 on a Tuesday night, someone sends a shop a screenshot of a story from six days ago and types four words: "is this still available?"
That's the whole message. No product name. No size. No colour.
A saved reply can't answer that. A keyword rule definitely can't. A human could, in about ten seconds, because a human would look at the screenshot, recognise the green ceramic mug, remember there are two left on the shelf, and type "yes, two left, want me to hold one?"
But the human is asleep. By 8am that customer has bought a mug from somebody else.
That gap is the entire reason sellers want to reply to Instagram DMs with AI. It's also the thing most so-called AI automation quietly fails at, so let's be specific about what separates the two.
Three ways to answer a DM, and only one of them reads
Every seller I've watched climbs the same ladder, usually in this order:
- Saved replies. You paste the price list yourself. Fast for you, still manual, still capped at your waking hours.
- Keyword rules. If the message contains "price", send the price list. Fires automatically. Also fires on "what's the price difference between the two blues?" and sends a nine-product image that answers neither half of the question.
- An AI assistant. It reads the message as a sentence, works out what's being asked, pulls the answer from your actual catalogue, and writes a reply in words.
Rungs one and two aren't useless. A rule that sends your shipping policy when someone types "shipping" will be right most of the time, and it costs nothing. Trouble starts the moment a message carries more than one idea, which is most messages people actually send. I wrote a longer breakdown of what a rule-based auto reply bot catches and what it drops, and the short version is this: rules handle the questions you predicted, and your customers have not read your list of predictions.
"Understands the question" is not a marketing phrase
Take a real-shaped message: "hi, do you have the beige one in large, and can it get here by Friday?"
That's three questions wearing one trench coat. Is the beige in stock? Does it come in large? Is Friday possible, to wherever this person lives?
A rule sees "beige" and sends a beige product photo. Job technically done, customer still stuck.
An assistant that reads does something different: confirms the beige, checks large against your stock, says large is gone but the medium runs big, then asks which city they're in so it can give a real delivery date. One reply. Three answers. One question back.
That last move matters more than people credit. A reply that ends in a full stop closes the door politely. A reply that ends in a question keeps someone typing, and people who keep typing buy.
Written out, the difference is embarrassingly plain. The rule-based shop sends a photo and the word "beige ✨". The reading assistant sends something closer to: "Yes, the beige is in stock. Large sold out yesterday, but the medium fits generously and most people size down in it. Which city am I sending to? If you're in a next-day area I can still make Friday."
Same three seconds of customer attention. One of those replies has answered everything and asked for the one detail needed to take the order. The other has sent a picture of a jumper.
Where the conversion difference actually comes from
People assume AI replies convert better because the AI is persuasive. It isn't, particularly. The lift comes from three unglamorous places.
Speed, but only the first reply. Nine seconds and four minutes both feel instant to a customer. Four hours doesn't. The value isn't beating your competitor by a second, it's never being the shop that answered tomorrow.
Completeness. Most lost sales I see aren't refusals. They're conversations where someone asked three things, got one answer, and couldn't be bothered to chase the other two.
Sorting. Only about 15% of first-contact DMs come from real buyers — I went through that anatomy in this look at who is actually sitting in your inbox. The rest are collaboration pitches, wrong taps on a story, and browsers. Your assistant isn't selling to all of them. It's clearing them in seconds so the buyers stop queueing behind them.
Together that shows up as a bigger share of conversations reaching the "how do I pay" moment. It does not show up as some dramatic overnight doubling, and any tool promising you that is selling a story rather than software.
Feed it the boring facts before you worry about the voice
Everyone wants to configure the personality first. Resist. An assistant with a lovely tone and no stock data is a very polite liar.
What it needs, roughly in the order customers ask for it:
- Product names as customers say them, not as your inventory system spells them. Nobody types "SKU-4471 Ceramic Mug, Sage".
- Current prices, plus what moves them: bundles, minimum order, whether tax is included.
- Stock, honestly. "Out until next week" beats silence, and it beats a guess by a mile.
- Shipping: courier, cost by region, and the cut-off time for same-day dispatch.
- Returns and exchanges, written in the words you'd really use with a customer.
- The five questions you personally answer every single day. You know them by heart. Write them down anyway.
That last bullet is the highest-value hour you'll spend all month. Most sellers can name their five questions in two minutes and have never once written them anywhere. For the full method of getting a catalogue into an assistant's head, I laid it out in this guide to training an AI assistant on your product catalogue.
Decide what it's never allowed to do
This is the step people skip. Then they get burned once and switch the whole thing off.
Write a short list of things your assistant may never do alone. If I ran a shop, mine would read: never invent a discount, never promise a delivery date the courier hasn't confirmed, never approve a refund, never take on a complaint that already arrived angry.
Then decide what happens instead. Not silence — that's how you lose the customer twice. It should say something true and human, flag the chat, and let you pick it up with the history right there. Those handover rules are worth getting right the first time, and I went through the ones that matter in this piece on when AI should pass a conversation to a person.
Your first week is a reading week
Switch it on, then read what it says. Not forever. Just the first week.
Thirty conversations a day, skimmed in ten minutes with your coffee. You're hunting exactly two things: replies that were wrong, and replies that were technically right but nothing like how you'd say it. Fix the first by adding the missing fact. Fix the second by pasting one of your own real replies in as an example of your voice.
Sellers who do this end up with an assistant that sounds like the shop. Sellers who skip it end up with one that sounds like a website, then complain that customers can tell. Customers can always tell. They mind far less than you think, as long as the answer is right and it arrived while they still cared.
The honest limit
AI will not close your hardest sales. The customer who needs convincing, the one comparing three shops, the one who wants to haggle — that one is yours. What it will do is keep that person in the conversation until you wake up, instead of letting them drift to a page that answered at midnight.
That's the real trade. You stop being the bottleneck on every small question, and you get your attention back for the conversations where you're worth more than any script.
If you'd rather not assemble all of this yourself, that's roughly what Vardast does: connect your Instagram, hand it your products and your rules, and it answers DMs in your voice around the clock, passing anything it shouldn't touch straight back to you.
The mug, by the way, was still in stock. The shop just found out nine hours too late.