Every Instagram chatbot can send a message. That is the easy half, and it is the half every feature comparison is built around: triggers, keywords, flows, buttons.
The half that decides whether you keep the thing running for six months is what happens when it should not send. A business account has customers, a reputation, and somebody whose job it is to answer the hard messages. A tool that sends unconditionally will eventually send the wrong thing to the wrong person at the wrong hour, and you will find out from a screenshot.
Ten checks, in a fixed order, first failure wins
Before any reply goes out, ten checks run in sequence. The first one that fails stops the send and the reason is recorded on the run.
- kill_switch — everything off, immediately, for the whole account.
- connection — the Instagram connection is live and the token is valid.
- takeover — a human has replied in this thread, so automation stops here.
- window — 24 hours since the person's last DM, 7 days for comments.
- dedupe — this person has not already been answered for this trigger.
- cooldown — enough time has passed since the last automated message to them.
- quiet_hours — it is not the middle of their night.
- rate_budget — the hourly cap has not been spent.
- credits — there is balance for a generated reply, if one is needed.
- content_safety — the text itself passes.
The order is not decorative. takeover runs before window because if you are already handling a thread, whether the window is open is irrelevant. credits runs late because there is no point charging for a message four earlier checks would have blocked.
The three that save businesses most often
takeover. You see a DM, you answer it yourself, and from that moment the bot is out of the thread. Without this, the classic failure is a customer getting your careful personal reply and then an automated one forty seconds later, which reads as though nobody is paying attention.
rate_budget. One Reel travels. Two thousand comments arrive in an evening. Without a cap, two thousand near-identical DMs go out in an hour, people report them, and Meta's spam signals do not care that you used the official API. The cap turns a viral post into a queue instead of an incident.
quiet_hours. A business reply at 3am is technically allowed and practically a bad idea. Most Indian accounts we see settle around 9pm to 8am IST, which also means the morning batch lands when people are actually awake to read it.
The setup is a settings page rather than a build, and what you actually configure is a short list. The order to work through it in:
- 01
Week one: templates only
Write four or five replies by hand for your most common questions. Turn the model off entirely. Templated replies are free and unlimited, so this costs nothing and it tells you what your inbox is really made of.
- 02
Week two: read the blocked runs
Look at what got stopped and by which check. Too many window failures means your triggers are firing on old comments. Too many dedupe failures means two triggers overlap.
- 03
Week three: let the model handle the leftovers
Turn on generated replies for the questions no template fits. Credits are spent only here. Keep the templates as fallback, so a slow or unsure model still produces your words rather than nothing.
- 04
Then stop touching it
The common mistake is continuous tuning. Set the caps, set the hours, and let a month of data accumulate before changing anything.
What it will not do for your business
Worth being blunt, because half of the questions we get are about things outside the product.
It does not do WhatsApp. It does not post content or schedule posts. It cannot message people who have not messaged you first, which rules out follower outreach entirely and is a platform rule rather than a plan limit.
It also will not make a bad offer work. If the answer to "how much?" is a number people balk at, answering faster gets you to the balk sooner.
Speed only helps a message that was going to land. It arrives at the objection quicker, it does not remove it.
The cost model, since it changes the setup
Templated replies are free and unlimited. Credits are spent only when the model writes something new.
For a business this has a specific consequence. Your highest-volume questions are also your most repetitive ones, which means the traffic that would cost the most is exactly the traffic a template handles for free. If price, delivery and availability are 70% of your inbox, 70% of your automation has no marginal cost.
The spend goes to the messages that genuinely differ, which is usually a smaller number than people budget for.
How to tell in a week whether it is working
Not by reply volume. Volume goes up the moment you switch anything on, and it tells you nothing.
Three things worth watching instead: the share of first replies that went out within five minutes, the number of threads where a human took over (that number should be small and should be the hard ones), and the number of blocked sends where you disagreed with the block. That third one is the real quality signal. If you keep finding replies that should have gone out and did not, a setting is wrong. If you never find one, the checks are doing what they were built for and you can stop reading the logs.