Playbooks · 5 min read
Customer Service on Instagram: What a Bot Can Truthfully Answer
Support is the one use of Instagram automation where being wrong costs money. A missed lead is a lead you never had and never miss. A confidently wrong answer about a delivery date is a customer who now has a screenshot.
So the useful question is not whether a bot can handle support. It obviously can handle some of it. The question is where the line sits, and most people draw it in the wrong place.
The wrong line: sorting by topic
The instinct is to sort by subject. Shipping questions, yes. Returns, no. Sizing, yes. Complaints, no.
It falls apart inside a week, because the same topic arrives in two completely different shapes. "What's your return window?" and "I want to return this, it arrived torn" are both returns. One is a published policy. The other is a person who is already annoyed.
The line that holds: does the answer change per person?
Sort by that instead and the sorting survives contact with a real inbox.
Stable — the same for everyone
Per-person — needs their record
Everything on the left is a fact about your business. You can write it down once, it is correct for everybody, and it stays correct until you change the policy.
Everything on the right requires reading a record about that specific person. And here is the part most tool marketing skips: on Instagram, through the official Graph API, an automation does not have that record. There is no order lookup, no ticketing system, no CRM sync, no shipping API on the other end. Leads export as a CSV and that is the whole of it. A bot that answers "where is my order" is not looking anything up. It is guessing in a friendly voice.
A bot that answers a per-person question has not looked anything up. It has guessed, warmly.
The good news is the left column is bigger than it feels. Pull a month of your own DMs and count. For most product accounts the same handful of stable questions covers the bulk of everything that arrives, and it is arriving over and over precisely because the answer never changes.
Three things it must never answer, regardless of topic
Anyone who is angry or upset. If the message carries heat — an order that went wrong, a promise that was broken, someone who has already messaged twice — automation should stand down and a person should read the thread before anything goes out. An automated reply to an angry customer does not calm anybody. It confirms that nobody is listening.
Money in motion. Refunds, failed payments, disputed charges. Even when the policy answer is stable, the person asking is usually not asking about the policy.
Health, safety, legal, or anything a person might act on. If a wrong answer could hurt somebody, it does not get automated. There is no clever prompt that makes this safe.
What the checks are actually doing for a support account
Ten checks run before any reply, in a fixed order, and the first failure stops the send with its reason recorded. Three of them matter disproportionately when the inbox is support rather than marketing.
takeover — the moment you reply in a thread by hand, automation is out of that thread for good. This is the one that prevents the worst support failure there is: your careful apology, followed a minute later by a cheerful automated greeting.
dedupe — one person, one answer per trigger. A customer chasing a problem messages repeatedly. Without dedupe they receive the same template three times, which reads as being ignored by a machine.
window — 24 hours from their last DM, 7 days on comments. Support conversations go quiet and restart days later. If they message again, the clock restarts from that message; if they do not, nothing you write can reach them.
Building it, in the order that goes wrong least
- Read a month of DMs and tally them. Not to design a flow — to find out which stable questions actually repeat. Most accounts are surprised by which one is top, and it is rarely price.
- Write four or five templates by hand. In your register, not a formal one. Templated replies are free and unlimited, so the highest-volume half of your support costs nothing to automate.
- Set negative keywords before you go live. Words that mean do not reply:
refund,broken,complaint,still waiting,cancel. This is the cheapest safety you will ever configure. - Test on yourself, then go live on one thing. Read the reply as a customer would, not as its author.
- Only then let the model write. Turn on generated replies for the questions no template fits. A credit is spent only there — one credit, one reply. The free tier is 100 credits with no card, and credit packs start at ₹499 when you outgrow it.
The honest boundary
A support bot on Instagram answers the questions you have already answered a hundred times, instantly, in your words, to people who would otherwise wait until morning. That is a real improvement and it is also the entire claim.
It does not resolve anything. It does not know what happened to an order. It cannot start a conversation with someone who has not messaged you first, and it will not carry a hard conversation for you.
If you want the sorting method in more detail, the triage version is here. If you want to see where each of these replies sits in a real conversation, the stages a DM moves through is the map. And if the buying question underneath this is which tool refuses correctly, that is judged by what it blocks.
