What the AI auto-reply does
The AI dialogues module answers incoming direct messages on behalf of your Telegram accounts. Someone writes to an account — the model reads the conversation, your instructions and excerpts from your knowledge base, and replies the way your manager would.
It runs on the server: once started, you can close the tab and the accounts keep replying. All conversations from all accounts sit in one inbox — you see what the AI said and can step in from there.
How it differs from Telegram's auto-reply and bots
On Telegram, "auto-reply" means three different things:
| Telegram Business away message | Bot | AI dialogues | |
|---|---|---|---|
| Where it works | your personal account | a separate bot | your accounts, however many |
| What it needs | Telegram Premium | the person must message the bot | connected accounts |
| What it replies | one prepared text | buttons and a script | based on the conversation |
| Remembers the chat | no | as far as the script goes | yes, recent messages |
If you only need "I'm on vacation", Telegram Business's built-in away message is enough. AI dialogues are for when questions are many and varied and need a real answer — across several accounts at once.
How it replies
- A direct message arrives on an account.
- The module checks the rules: is the reply limit for this person used up, are they on the blocklist, is it within working hours.
- The model gets your instructions, the latest messages of the conversation (10 by default) and relevant excerpts from your files.
- Before sending — a pause of 8 to 45 seconds and a "typing" status, like a person reading and replying.
- The reply goes out, and the conversation continues.
Instead of the model you can switch on one fixed text for everyone — for example, when you only need a first reply with a link.
Knowledge base and instructions
Instructions are a text about who is replying, what can be discussed and what can't. They include the topic and the product — that's what the model talks about.
The knowledge base is your files in PDF, DOCX, TXT and MD: a price list, service descriptions, FAQs. The model picks the most relevant fragments for each question. The more precise the files, the less the model fills in — if a price isn't in the base, say in the instructions that it mustn't name one.
You can keep several settings sets — for example, "Sales" for some accounts and "Support" for others.
When the AI stops
Each conversation has limits so the model doesn't run it forever:
- reply limit per person — 5 by default, then the conversation waits for you;
- stop after a link — once it sends a payment or booking link, the AI goes quiet;
- working hours — in the account's time zone;
- blocklist — people it never replies to;
- run mode — unlimited, by time, or by number of replies.
Replies count toward the account's shared action budget along with other modules: if the same account also comments, the auto-reply won't overspend it.
Handing a conversation to a person
Any conversation can be taken over from the AI with one button and continued by hand. Handing it back is one button too, and the reply counter resets. Conversations can be exported.
If you use GramClaude's direct messages module, people who replied can be passed straight to AI dialogues — the conversation continues without a gap.
Which AI model to use
There are two options:
- built-in model — included in plans and the trial, no key needed;
- your own key — Anthropic or any OpenAI-compatible provider.
If you buy the module on its own, without a plan, you'll need your own key. The conversations the model replies to are sent to the provider you choose — keep that in mind if chats contain personal data.
What the module doesn't do
- It doesn't write first. AI dialogues only reply to incoming messages. You can start a conversation with someone new by hand from the same inbox.
- It doesn't promise the model never errs. Test replies on your own questions before running it on all accounts.
- It doesn't override Telegram's rules. If people report an account, it gets a spam block like any other — see the guide on spam blocks.
