Telegram bot projects are usually scoped against what the platform appears to allow, and then redesigned once they meet what it actually allows. Six constraints do nearly all of that damage. The two biggest: a bot cannot message anyone who has not started it first, which invalidates most "notify our customers" specifications outright, and a bot in a group sees almost nothing unless it is made an administrator. The rest are rate and size ceilings that turn broadcasting into an architecture rather than a loop.
Telegram is an unusually good platform to build on, which is precisely why this article exists. The API is clean, the documentation is honest, and a capable team can ship something useful quickly. The trouble is that the constraints are not where you expect them, so they tend to be discovered after the specification has been agreed with someone who will be annoyed.
These six come up in nearly every scoping conversation we have.
1. It cannot message anyone first
This is the one that ends the most ambitious plans, and it is absolute: a person must start a conversation with your bot before it can send them anything.
So the extremely common brief — "we have forty thousand customers, we want to reach them on Telegram" — is not a bot project. It is an acquisition project with a bot at the end of it. Every one of those forty thousand people has to find the bot and press start, and your conversion on that step determines the entire value of the build.
That is not a reason to avoid Telegram. It is a reason to put the acquisition route in the specification: the deep link in your existing emails, the QR code on the invoice, the button in your app. A bot nobody has started is a well-engineered thing that nothing reaches.
2. In a group, it is nearly blind by default
Privacy mode is enabled unless someone changes it, and a bot in that state receives only messages that are commands directed at it, replies to its own messages, and service events. It does not see ordinary conversation.
Which means every feature of the form "the bot watches the group and does something clever" requires either disabling privacy mode or making the bot an administrator. Both are decisions for whoever owns that group, and both change what you are asking of them: the bot now receives everything said in it.
Raise that in scoping, not in testing. For an internal team group it is usually an easy yes. For a customer community it is a conversation about what the bot stores and why, and it belongs in the privacy notice before it belongs in a sprint.
3. Broadcasting is a queue, not a loop
The figures circulated for Telegram's rate limits — about one message per second to an individual chat, roughly twenty per minute in a group, and around thirty per second across all chats — come from Telegram's own FAQ and are widely repeated. Worth knowing that Telegram does not publish them as formal guarantees, so treat them as the shape of the ceiling rather than numbers to engineer to the decimal.
The shape is enough to decide the architecture. At roughly thirty per second, a hundred thousand recipients is about an hour of continuous sending. That is not a for-loop in a request handler; it is a queue, a worker, and a set of decisions you have to make deliberately:
- What happens when a send returns 429 — the retry delay is in the response, and it should be honoured rather than guessed at.
- What happens to users who blocked the bot since last time, which is a normal and permanent outcome rather than an error.
- How you know where the run had reached if it stops halfway, because an hour is long enough for a deployment to land in the middle of it.
- Whether a message sent twice is acceptable, and if not, how you prevent it.
Telegram does offer a paid broadcast tier with a far higher ceiling for bots that meet its eligibility requirements. It is worth asking about if volume is central to the product, and it does not remove the need for the queue — it only raises the rate the queue drains at.
4. Files have ceilings, and one of them is low
On the standard cloud Bot API the commonly cited limits are 50 MB for what a bot can send and 20 MB for what it can download. The download figure is the one that bites, because it is below the size of plenty of ordinary business documents — a scanned contract, a short video from a phone.
Running a local Bot API server raises the upload ceiling substantially and removes the download restriction, which turns an impossible requirement into an infrastructure line item. It does not touch the rate limits. If your product is document-centric, establish which side of that line you are on during scoping, because it is the difference between a hosted bot and a bot with servers behind it.
5. It cannot see what happened before it arrived
A bot added to a group today has no access to that group's history. For an archive, a migration, or any analysis of what has already been said, the Bot API is the wrong instrument entirely.
6. Some things need a user, not a bot
Which leads to the boundary that decides the shape of a project. The Bot API gives you a bot: a distinct kind of account with defined permissions. Anything requiring a user account — reading history, acting as a person, operating where bots are not permitted — means MTProto and a user session.
That is a materially different build with different operational risk, including risk to the account being used. The service page sets out where that line sits in more detail. The point for scoping is simply that it is a fork in the road, and it is much cheaper to take knowingly at the start than to discover in week six.
What to settle before the estimate
- How will users start the bot? Name the channel and the link. This is the whole project's ceiling.
- Does any feature need group messages? If so, get agreement on admin rights before design.
- What is the largest audience a single send reaches, and is the resulting duration acceptable?
- What is the biggest file in each direction? It decides whether you host a Bot API server.
- Does anything need history or a user account? If yes, that is a different project, scoped separately.
Five questions, usually twenty minutes, and they routinely change the estimate by more than anything discussed in the rest of the meeting.
Questions we get asked
Can a Telegram bot message a user first?
No. A user has to start a conversation with the bot before it can send them anything. This single rule invalidates more specifications than any other, because the common request — notify our existing customers through Telegram — requires every one of those customers to have found and started the bot first. Acquisition becomes the project, not the bot.
Can a Telegram bot read all messages in a group?
Not by default. Privacy mode is on unless changed, and a bot in that state receives only commands directed at it, replies to its own messages, and service events. To see everything a group says, the bot must be made an administrator or have privacy mode disabled — which is a decision for whoever owns that group, and worth raising before you design a feature that depends on it.
How fast can a Telegram bot send messages?
The figures widely cited from Telegram's own FAQ are roughly one message per second to a single chat, around twenty per minute in a group, and about thirty per second across all chats. Telegram does not publish these as formal guarantees, so treat them as the shape of the constraint rather than exact numbers. Exceeding them returns a 429 with a retry delay.
How do you broadcast to a large Telegram audience?
With a queue and a worker that respects the rate ceiling, not a loop. At roughly thirty messages per second a hundred thousand recipients takes around an hour, so the design has to handle partial completion, retries after 429 responses, users who have blocked the bot, and knowing where it had reached if it stops. Telegram also offers a paid broadcast tier at a much higher rate for bots that meet its eligibility requirements.
What file size can a Telegram bot send and receive?
On the standard cloud Bot API the commonly cited ceilings are 50 MB for uploads and 20 MB for files the bot downloads. Running a local Bot API server raises the upload ceiling substantially and removes the download limit, at the cost of hosting that server yourself. It does not change the rate limits, which are enforced regardless.
When is the Bot API not enough?
When the requirement needs a user account rather than a bot — reading history the bot was not present for, acting in channels as a person, or anything on the client protocol rather than the bot interface. That means MTProto and a user session, which is a materially different build with different operational and account risks, and it is worth establishing in scoping rather than discovering mid-project.
If you have a Telegram brief and want the five questions above answered before anyone estimates it, that conversation is usually short and occasionally saves the whole budget. It is how we open every Telegram bot project.