An Instagram AI chatbot that answers DMs from your own documents

You can answer Instagram DMs with AI on a Business or Creator account, and the version worth using reads your own documents rather than the internet. You upload a PDF, paste your FAQ or import a page from your site. When somebody asks about prices, delivery or opening hours, the question is matched against what you uploaded and answered from those passages only. If your documents do not cover the question, it says so and hands the conversation to you instead of guessing. That last part is the whole difference between a useful assistant and a liability.

The only safe way to let AI answer your DMs is to ground it in your own documents and let it refuse. A bot that invents a refund policy costs more than the messages it saved you.

What Instagram already gives you, free

Do these first. They are built into the professional inbox, they cost nothing, and between them they cover the messages you get most often. A tool that ignores them is selling you something you may not need.

• Saved replies. Write an answer once, give it a shortcut, and insert it in two taps. Perfect for the question you answer forty times a week in the same words.
• FAQs. A professional account can set a handful of tappable questions that appear at the start of a new chat, each with a stock answer. People tap rather than type, and you never see the message.
• Away messages and instant replies, through Meta Business Suite. One automatic reply when somebody first writes, or when you are closed.
• Hidden Words. The same filter that hides spam comments also filters message requests, which is the quickest way to stop scam DMs reaching your inbox at all.

Between them these handle the predictable half of an inbox. Where they run out is anything that is not exactly one of your prepared questions. Somebody asks whether the course covers a specific topic, or whether you deliver to their city, or what the difference is between two of your packages. A saved reply cannot answer that. A person has to read it, find the answer in a document, and type.

What a knowledge base is, in practice

It is a named collection of documents, plus the settings for how it answers. Three ways to put something in it, and you can mix them in one base.

1. 1Upload a PDF. Your price list, course outline, policy document or brochure. Up to 10 MB.
2. 2Paste text. The fastest option, and often the best one: your actual FAQ, written the way you would answer it.
3. 3Import a web page. Paste the address of your FAQ, pricing or policy page, and the readable text is pulled out. Navigation, headers and footers are stripped, because a nav bar repeated on every page pollutes every answer.

Each document is split into passages of roughly 1,200 characters, preferring paragraph then sentence boundaries so a passage rarely starts mid-thought, with 150 characters carried over from one to the next in case an answer straddles the join. Each passage is then turned into a vector so a question can be matched against meaning rather than keywords.

The embedding model is multilingual on purpose, not as a nice-to-have. An English-only model scores a Hindi question against a Hindi document at close to noise, which would make the feature work for part of the audience and quietly fail for the rest.

How one DM gets answered

1. 1The message arrives. Campaigns and DM flows are checked first, so an existing keyword campaign always wins and nothing you already built changes behaviour.
2. 2The question is turned into a vector and compared against every passage in the base.
3. 3The four closest passages clearing the similarity floor are taken. If nothing clears it, the two closest above a much lower floor are used instead, and the model's own refusal rule decides whether they really answer the question.
4. 4Those passages, the last few exchanges of the conversation and the language to reply in go to the model. Nothing else does.
5. 5The model returns an answer and a flag saying whether the passages actually answered it. A false flag means your fallback message goes out instead, never a guess.

The second, lower floor is there for a specific failure. A question phrased differently from the document used to get nothing at all, which reads to the customer as being ignored. Letting the model refuse instead means the worst case is the same refusal, and the common case is an answer to a question that was only worded differently.

The rules that stop it inventing things

These are instructions the model is given on every single answer, not a setting you can weaken, and they are the reason the feature is safe to point at customers.

• Use only the passages. No general knowledge, no guessing. Saying it does not know is always better than inventing a price, a policy, a date or a promise.
• Never invent a link, a phone number or a discount code that is not in your documents.
• Never explain what it could not find. Phrases like “not specified” or “the document doesn't say” are banned, because they read as a machine reporting a failure rather than a person helping.
• Never mention documents, passages, or being an AI.
• Speak about you in the third person, by name. “Glow Studio opens at 10am”, not “we open at 10am”, because an assistant writing as though it were you is a different and worse promise.
• Plain sentences, under 850 characters, no markdown or bullet symbols. A DM is read on a phone, and asterisks in a DM look broken.
• Answer fully. If the passages hold the price, the timing and the conditions, give all three rather than the first one, because a half answer only makes the person ask again.

When the documents cannot answer, the fallback message is yours to write. Leave it blank and a sensible default goes out: “I don't have that one to hand. Someone from the team will get back to you here.” Silence would be worse than either, because the person already knows somebody is listening.

Replies come back in the language of the question

This is worth more than it sounds if your audience is not all English speakers. The language of the incoming message is worked out before the model is called and handed to it as a direct instruction at the end of the prompt, where it is hardest to ignore. Your documents can be in English and the reply still comes back in Telugu, French or German.

Hindi typed in English letters gets its own rule, because models otherwise read it as badly spelled English and answer in English. Somebody writing “price kitna hai” gets an answer in the same style they wrote in. That behaviour exists because the owner's own test chat kept getting English back for every Hinglish message, which is exactly what a creator's audience notices first.

Two ways it starts answering

Pick deliberately. The first is cautious and the second is not, and the difference is who gets an automatic answer.

A keyword conversation stays open without the word being repeated, which is the point. Requiring it on every message would make the feature a search box people have to operate rather than something that reads like talking to the business. It closes when the person goes quiet for however long you set, when they type one of your stop words, or after 40 exchanges, which is a ceiling against a loop between two bots rather than a limit anybody will meet.

It gets out of the way of people, campaigns and flows

The failure that would actually lose you a customer is the AI talking over you mid-conversation, or answering while a campaign is waiting for somebody to tap a button. Five separate quiet periods exist for that, and none of them needs setting up.

There is also a writing style field, up to 300 characters, which is the setting most people underuse. Describe how your replies should read, for example that you are a small studio and write casually, and it applies to tone, length and wording within every rule above. It costs nothing and it is the difference between an answer that sounds like your account and one that sounds like a helpdesk.

The tap suggestions after each answer

Each answer can carry up to five follow-up questions as tap buttons, drawn only from what your documents can actually answer. They exist because most people do not know what else to ask, and a button is easier than typing.

Instagram allows 20 characters on a button, which is not enough for a real question, so each one carries two things: the short label people see, and the full question hidden in its payload. Tapping “Bridal makeup price” asks the whole question, and the base that offered the button answers it, even if no conversation was open.

What it costs

An answered DM costs two AI credits: one for matching the question against your passages and one for writing the reply, which carries several retrieved passages with it. A message that nothing answers costs nothing.

At two credits an answer, Growth's 2,000 credits are about a thousand answered DMs a month. That is a useful number to hold against your own inbox before deciding whether any of this is worth it.

Behind the scenes the model call walks a fixed list of providers, so one being rate limited or down hands over to the next rather than failing. When the whole chain fails, nothing is sent and the message waits for you, which is the right direction for an error in a customer conversation.

The limits worth knowing before you set it up

One limit is Instagram's rather than ours, and it governs everything here: you can only reply to somebody for 24 hours after their last message. An answer is instant, so this rarely bites, but it is why an assistant cannot pick up a conversation from last week. The /blog/instagram-24-hour-messaging-window post covers what that window does and does not allow.

What it cannot do

• Answer something your documents do not contain. This is the design, not a gap. If customers keep asking about something, the fix is to add it to a document.
• Work on a personal account. Reading and sending DMs needs an Instagram Business or Creator account connected through Meta's own authorisation screen.
• Start a conversation. Instagram only allows a reply inside the 24 hour window after somebody writes to you, so nothing here can message a stranger first.
• Take a payment or place an order. It can send a link to your checkout, and that is the boundary.
• Read a voice note, an image or a story mention. Text messages only, and an emoji on its own is treated as a reaction rather than a question worth two credits.

A sensible way to start

1. 1Set up Instagram's own saved replies and FAQs first. Free, and they cover the questions you already know the wording of.
2. 2Write your real FAQ as plain text and paste it in rather than uploading a brochure. Documents written for customers retrieve far better than ones written for print.
3. 3Test it in the dashboard. Ask the ten questions you actually get, including two you know are not covered, and check that the uncovered ones refuse rather than improvise.
4. 4Give it a keyword and leave “answer every DM” off for the first week. Read what it sends before widening it.
5. 5Write the fallback message and the writing style in your own words, then switch on “answer every DM” once you trust the answers you have read.

If you want the feature detail rather than the setup, /ai-features/knowledge describes it without the walkthrough.