How to launch an AI agent in Telegram: a ten-step checklist

Less than people assume. The technical minimum is a Telegram account and a clear description of the job in plain words. Everything else is either handled by the service or not needed at all on the first launch.

• A Telegram account. The bot is created automatically and handed to you; you do not need to wrestle with a BotFather token by hand.
• One job, written in plain words. Not "I want AI" but "triage incoming requests and draft a reply" — everything downstream depends on how precise this is.
• Half an hour on day one. The assistant answers immediately, but its role and boundaries get sharpened by the first conversations.
• A model API key is NOT required: inference is included in the subscription. BYOK services charge less per month but require your own account with a model provider.
• Before paying, check the available payment methods, currency and full amount for the selected period. If your preferred method is unavailable, clarify the terms before paying.

This is the first fork and it determines everything after it. A platform's shared bot (Mira, Manus, Poke) is faster and often free, but you get no @name of your own, no isolation and no way to hand the bot to anyone else. Your own bot costs a subscription — but it is yours.

• A shared bot — if the job is personal, one-off, and you do not mind the conversation running through the platform's own assistant.
• Your own bot — if the assistant will handle your documents and customer data, or if you want a separate name and separate memory per role.
• Your own bot is mandatory if you might ever show it to a team or to customers: a shared bot cannot do that at all.

Order matters: the first three steps affect answer quality more than the choice of model. From step seven onward, the checklist is about the agent still being useful a month later, not just on launch day.

1. Write down one job and its result. "Drafts a reply to a customer in five minutes" is a job. "Helps with the business" is not.
2. Describe the role and the boundaries. What it does on its own, what it only proposes, what it never does. Boundaries are worth more than skills: they protect you from confident nonsense.
3. Take a ready-made setup for your niche if one exists. Agent OS has fourteen, from an auto shop to a tutor; the role, boundaries and first day are already written, and editing beats writing from scratch.
4. Create the bot and receive it in Telegram. With us that step is automatic; on your own it means a BotFather token and hosting.
5. Upload the working context: price list, playbook, reply templates, examples. The assistant reads PDF, DOCX, TXT, Markdown, XLSX and CSV.
6. Test the memory. Say "remember: …", ask again in a fresh conversation, and confirm the fact stuck and is visible in your dashboard.
7. Turn on one or two scheduled skills — a morning digest, for instance. Do not switch on three at once: you will not know which one works.
8. Run ten real requests in a row and mark where you had to rewrite the answer. That list is your role edit list.
9. Set up AI disclosure if the agent will talk to anyone but you: a person must learn they are talking to an AI at the start of the first interaction.
10. Check what happens when the limit runs out and how to cancel. Everyone skips this step and remembers it at an inconvenient moment.

It depends on what you count as a launch. A first bot reply takes minutes almost everywhere. A working agent you would trust with a real job takes from one evening to two months, and the spread comes not from the service but from whether a ready-made setup for your niche exists.

• Agent OS: your part is a couple of minutes describing the job; the build runs itself and the link arrives in Telegram. Tuning the role on the first conversations takes an evening.
• Zapier Agents: under five minutes to a first agent, after which cost per run becomes the constraint.
• Botpress: five minutes to a demo and, per user reports, four to eight weeks to production.
• n8n: two weeks to two months — it is an automation platform, not a ready assistant.
• A freelancer: average Fiverr delivery is sixteen days, and support after delivery is not included.

Five mistakes that repeat more often than any others and cost the most.

• Too broad a role. An agent for everything answers worse than an agent for one job — on any platform and any model.
• No boundaries. Without an explicit "never do this", the model fills the gap with a confident guess.
• No context loaded. An assistant without your documents knows exactly nothing about your business and has to guess.
• Limits unchecked before paying. Ask what happens when they run out: top up, stop, or economy mode — those are three different products.
• No exit plan. Ask whether you can export the assistant and its memory. If you cannot, you are renting a dependency, not a tool.

The moment someone other than you talks to the agent, requirements apply that never came up during personal use.

• AI disclosure. Article 50 of the EU AI Act has applied since 02.08.2026: a person must learn they are interacting with an AI at the start of the first interaction. Penalties reach €15M or 3% of turnover.
• Telegram's rules. The platform requires working support and terms commands on any bot that takes payment, and forbids using conversations to train models.
• Responsibility for answers. A model's answer can be wrong; legal, medical and financial topics need a disclaimer and human review.
• Customer data. What enters the assistant's memory, where it is stored and how it is deleted on request — you need that answer before the first customer, not after.
• The cap on replies to strangers. In Agent OS client mode unlocks on PRO, is off by default and is capped at 20 replies a day; if your volume is higher you need a bot builder, not a personal assistant.

Steps and capabilities are described from the Agent OS product documentation and its capability map. Competitor timelines and legal requirements come from the competitive market review based on primary sources (vendor pages, the European Commission, Telegram documentation).

• 14 ready-made assistant setups: start from the closest to your niche
• Current Agent OS plans and payment terms
• AI assistant in Telegram services, 2026
• What happens to your data: the privacy policy