It's 11:40 at night. A customer opens your DMs and types: "hii do u still have the blak one in 38?? the one from ur story yesterday"
Your bot reads it, looks for a keyword it knows, finds none, and sends the message it sends to everyone it doesn't understand:
"Sorry, I didn't get that. Please choose: 1) Prices 2) Shipping 3) Talk to an admin."
She picks 3. The admin is asleep. By morning she's bought the same boots from a page that answered.
I've watched this exact scene play out in more seller inboxes than I can count, and here's the uncomfortable part: the bot didn't malfunction. It did precisely what it was built to do. That's the whole problem with rule bots, and it's why the rule bot vs AI assistant question isn't really about technology. It's about who has to do the thinking when a customer says something you didn't predict.
Before the end I'll tell you when a simple rule bot is honestly all you need, because sometimes it is. First, let's look at where it breaks.
What a rule bot actually is
Strip away the marketing and a rule bot is a list of "if this, then that." If the message contains price, send the price list. If it contains shipping, send the shipping text. If the customer taps button 2, jump to branch 2 of the flowchart.
Keyword bots and flowchart bots are cousins. One listens for words, the other makes the customer click through a menu. Both rest on the same bet: that you can predict what people will say before they say it.
For a narrow slice of messages, that bet pays. For a real shop inbox, it loses far more often than people admit.
Where rigid bots break
There are four usual suspects. You've probably met all of them.
1. Typos and slang
Customers don't write like your keyword list. "Price" arrives as "prise", "how much", "hm", "cost?", or a screenshot with a circled item. You can keep adding synonyms, and plenty of sellers do, until the list is 200 lines long and still misses the 201st way of asking.
2. Two questions in one message
"Do you have it in blue and how long to Toronto?" A keyword bot catches one of those, usually whichever rule sits higher in the list. The customer gets half an answer and has to ask again. Most don't bother.
3. Context
This is the big one. A customer asks about the grey hoodie. Two messages later they write "ok and in L?" A rule bot has no idea what "in L" refers to, because every message is a fresh start. It doesn't remember the hoodie. It barely remembers the customer.
4. The question nobody planned for
"Is this fine for someone allergic to wool?" "Can I pay half now?" "My sister bought this last month, is the new batch the same size?" Every shop has a long tail of questions nobody wrote a rule for. With a rule bot, that tail goes straight to the fallback message, and the fallback message is where sales go to die.
I wrote more about this in what an Instagram auto reply bot catches and what it drops, but the short version is this: the questions a bot drops tend to come from the people closest to buying. Someone asking about allergies or split payments has already decided they want the thing.
What an AI assistant does differently
An AI assistant doesn't match words against a list. It reads the message the way a person would, works out what's being asked, and answers from what it knows about your shop: your products, sizes, prices, shipping rules, return policy.
Take the 11:40 pm message again. A decent AI assistant reads "blak one in 38" plus "the one from ur story yesterday," checks the catalogue, and replies something like:
"Hi! Yes, the black ankle boot from yesterday's story is still available in 38. It's $89 and ships in 2 to 3 days. Want me to hold a pair for you?"
Same customer. Same hour. Nobody woke up. The assistant handled the typo, understood which product she meant, answered the actual question, and asked for the sale.
Here's what that looks like day to day:
- Typos stop mattering. "Blak", "black" and "the dark one" all land in the same place.
- Multi-part questions get full answers. Colour and shipping, in one reply.
- The conversation has memory. "Ok and in L?" means the hoodie, because the hoodie came up two messages ago.
- The long tail gets handled, or handed off honestly. If the answer isn't in your info, a good assistant says so and brings a human in instead of pretending.
- It sounds like your shop, not like a phone menu from 2004.
"But AI makes things up." Fair. Here's the real risk.
I'm not going to pretend AI assistants are magic. The fear sellers bring up most is that the assistant will invent a discount, promise a colour you don't stock, or quote last season's price.
That fear is legitimate. It's also mostly a setup problem, not a technology problem. An assistant that answers from a real product catalogue, with clear instructions about what it may and may not promise, stays grounded. An assistant dropped into your DMs with a two-line description of your shop will guess, and guessing is how you end up honouring a discount you never offered.
The fix is boring: give it your real catalogue, prices and policies, and decide the moments it should step back. I've got a whole piece on when AI should hand off to a human, because that handoff is where trust is won or lost.
A rule bot never makes things up, true. It also never answers most of what your customers actually ask. Pick your failure.
Side by side, in one small shop
Say you run a small clothing page and get about 120 DMs a day. In a typical inbox, the split looks roughly like this:
- about 40 are simple: price, sizes, "do you ship to X"
- about 50 are the same questions asked messily: typos, two questions at once, follow-ups like "and in blue?"
- about 30 are the odd ones: fabric questions, gift wrapping, "can I swap after buying"
A well-built rule bot handles most of the first 40. On a good day it handles some of the middle 50. It handles almost none of the last 30. Call it 55 to 65 out of 120 answered properly. The rest hit the fallback message or wait for you at midnight.
An AI assistant trained on your catalogue handles the first 40, most of the 50, and a good share of the 30, passing the genuinely tricky ones to you with the conversation attached. That's the gap between your inbox being a chore and your inbox being a sales channel.
Those numbers are illustrative, not a promise. But the shape is what I see again and again: rule bots do fine on the easy third and fall apart on the messy middle, and the messy middle is where most real buyers live.
When a simple rule bot is still enough
I said I'd be honest about this. There are real cases where a rule bot is the right tool, and paying for more would be wasted money:
- Comment-to-DM giveaways. "Comment WIN and we'll DM you the link." One trigger, one message. A keyword rule is perfect for it.
- Single-purpose flows. Collecting an email for a waitlist, sending a PDF, confirming a booking from three fixed time slots.
- Very low volume with a tiny catalogue. If you sell one product and get ten DMs a day, you can answer the edge cases yourself over coffee.
- Fixed-text replies where the wording must never change, like a legal notice.
Notice what those have in common. The customer isn't really talking. They're pressing a button. The moment people start asking questions in their own words, the rule bot's bet starts losing.
The mistake I see most isn't choosing a rule bot. It's choosing one and then asking it to do a salesperson's job. That's where the DM automation mistakes that quietly cost you sales pile up: a menu sitting where a conversation should be.
A five-minute test for your own inbox
Open your DMs from the last week and scroll. Don't count yet, just read. Then ask yourself three things:
- How many messages have a typo, slang, or two questions at once?
- How many point back to something said earlier ("that one", "same but bigger")?
- How many would a new employee need to check with you before answering?
If the answer to all three is "almost none," a rule bot will do the job. If it's "most of them," and in my experience it usually is, you've outgrown keywords whether or not you've admitted it yet.
You don't have to pick a side forever, either. Plenty of sellers keep a keyword trigger for giveaways and let an AI assistant run everything conversational. The two tools don't have to fight.
Where Vardast fits
This is the gap Vardast was built for. It connects to your Instagram, WhatsApp, Telegram and website chat, learns your products and policies, and answers customers in your brand's voice, typos and follow-ups included. When something needs you, it hands the conversation over instead of guessing. If you're tired of your bot saying "Sorry, I didn't get that" to people who are holding their card, see how Vardast handles your DMs.
Back to the woman at 11:40 with the "blak one in 38." She was never a difficult customer. She knew what she wanted and she was ready to pay. The only thing between her and your checkout was a bot that needed her to speak its language. Flip that around, and most of your midnight DMs stop being lost sales.