A product Reel does something predictable to a comment section. For the first
hour it fills with the same four words: price, size, shipping, link.
Each one is a person holding their phone with some intent to buy, and each
one scrolls on to somebody else's store if the answer takes until tomorrow.
That is the part automation is good at. The rest of this page is about the part it is not, because in a store inbox that second part costs real money.
The questions a store actually gets
Store comment sections are remarkably consistent, and most of what arrives fits one of four answers you would give identically to anybody who asked:
price/rate/kitne ka— the number, in the currency your customers pay in.size/size M available?/fit— the size chart, and where the live stock is shown.shipping/delivery/COD— how long it takes, where you ship, and whether cash on delivery exists.link/where to buy— the product page.
Each becomes one rule. Keywords are any-of, so one match fires it, and
matching ignores case and punctuation, so PRICE?? and price are the
same word. What the matcher will not do is guess: cost is not price unless
you listed both.
A public reply, then the link in private
The shape that works for a product post is two messages. A short public reply under the comment, so everybody reading the thread can see the question was answered. Then a DM with the actual product link, because links in Instagram comments are not clickable and a DM is somewhere the customer can find again later.
The DM also does one quiet piece of work: the person who received it lands in your lead list, with the post they came from and the keyword they typed. Leads export as CSV, which is the honest limit of it. There is no store sync and no marketing automation on the other end; the list is yours to take wherever your customers already are.
What a model must never guess
The case for letting the AI answer in a store inbox is that customers ask questions you did not template. "I am 5'4 and usually wear M, will this work?" is a real question with a real answer, and a model given your size chart can write a good one.
The case against is the same sentence read differently. A model will answer a question whether or not it has any way of knowing the answer, and it will sound friendly and confident either way. A third category of question looks like the fit question and is actually about facts only your systems hold:
- Is this in stock. Nothing connects a reply to your inventory. "Yes, available" is a statement about last Tuesday, sent today.
- Where is my order. No reply can see your courier or your order table. A made-up tracking answer turns a question into a complaint.
- Did my payment go through, can I cancel. These are for a person with access to your admin, every time.
The fix is a rule of writing rather than a feature. Anything with a number that could have changed since you wrote it should be a pointer, not a promise: "the live stock is on the product page" survives forever. Give the model your shipping policy, your size chart and your returns page as knowledge sources, and it answers from those instead of from its imagination.
Most bad AI replies in a store inbox are not the model being wrong. They are the model answering a question that had no answer available to it.
Takeover matters more in commerce
A sizing question becomes a complaint becomes a refund request, sometimes inside the same afternoon. The worst experience a customer can have is your careful personal message followed forty seconds later by a cheerful generated one.
So the moment you reply to a thread by hand, the automation stands down in it. It is one of the ten checks every outbound message passes, evaluated before the messaging window and long before credits, so a thread you have stepped into is never answered twice and never charged for.
What it costs a store
Templated replies are free and unlimited on every plan. A credit is spent only when the model writes something new, one credit per reply. For a store that runs opposite to intuition: the busiest traffic is the cheapest traffic, because the four repeated questions are templates and cost nothing.
Free covers one Instagram account and comes with a hundred AI credits when you sign up. Creator adds 1,500 AI replies a month, the inbox and the lead list, and a second account. Studio is for a store with several brand accounts: five accounts and three team seats. The pricing page has the current numbers and how the two credit buckets are spent.
What to switch on, and when
- Week one, templates only. Four rules, model off. Read Activity every evening, including what was held and why.
- Week two, fill the gaps. Add keywords for the words people actually typed, and a template for any question that arrived more than a few times.
- Week three, the model for the leftovers. Turn on AI replies for questions no template covers, with your policies as knowledge sources, and keep reading what it wrote.
That order is not caution for its own sake. It is the only sequence in which you find out what your inbox actually contains before you pay a model to answer it. The long version of this, with the checks to run in the first week of generated replies, is the post on AI replies for a product inbox.