AI agent for business in your team group: how to add it and what to hand it

The short answer

Add your agent to the team group like any other member and leave it on /mention on. From then on it stays quiet while the team talks and answers only when someone writes its @name or replies to its message. Turn on forum topics, and every topic becomes its own session with its own memory: Returns, Shifts and Hiring never mix.

Then hand it the jobs a good team assistant does inside the thread. Sum up what was agreed and who took what. Keep the list of assignments for a topic and say what is still open. Remind you to check a deadline. Draft a reply a teammate can send to a customer. Write the rules of the group so people know how to call it.

That is the practical side of an AI agent for business: it does not live in a separate app, it sits where your department already talks. The decisions, the priorities and the people stay with you.

Add the agent to your group: the steps

We started where every department head starts and asked our test agent how to set this up. The persona is made up: Nora, head of customer support at Stride, an online shoe store, seven people on the team. This is the real chat with our test agent on 30 October 2026.

Two answers in one. The first is the quiet mode: in a group the agent answers only on @mention or a reply, and that is already the default. The second is access for the team: you open a topic and send /allow with the colleague's username, and from then on that person can call the agent there. The agent also suggested /team on for General, so your own messages to colleagues do not get answered unless you tag it.

If you want topics, the Guide lists five steps, and the agent walks you through them in the chat:

1. Create a new group.
2. Add your agent's bot to it.
3. Set chat history to visible for new members.
4. Turn on Topics.
5. Make the agent an admin with the right to manage topics.

When the agent joins, it greets the group and says it will answer only when mentioned, with a button to switch to replies on every message. Here are the commands for groups in one place:

• /mention on or off: answer only on a mention or a reply, or on every message from the owner;
• /allow @username: let a colleague talk to the agent in the current topic (marked experimental in the Guide);
• /allowed and /revoke @username: see who has access and take it away;
• /team on: skip your messages that are addressed to someone else.

The test: a Friday in the Returns topic

Stride ran a promo on boots, and a week later the Returns topic is busy: 64 open tickets against the usual 25. Sam leads the shift, Ana runs the help center, Tom talks to the courier, Leo owns the reply templates, Mia is on the team too. All of them, the store and the tickets are invented for the test.

We ran the tasks in a chat with our test agent. The topic messages went in the way they look in the group, and every request to the agent started with its @name, the way you would write it in the topic. The agent's replies are quoted as they came, shortened in a couple of places and never added to. The agent's bot name in the replies is shown as @stride_desk_bot, the name the team uses in the story.

Task 1. “What did we agree here?” on @mention

Nine in the morning, six people, seven messages, and the thread already has three problems in it. Nora tags the agent at the end.

Fifteen seconds. The opening line is the useful one: nothing was actually agreed. Five updates and two offers read like progress in a busy topic, and the agent separates what someone took from what was only said out loud.

Then it does the part that is easy to miss when you skim. Ana's form text needs an approver, and nobody answered her. Leo asked who writes to the 20 customers with the old promise, and nobody did. Sam needs someone for Sunday. Those three go under “Nobody took”, with the time each was raised. The agent also ranks them: the 20 customers are the most urgent, because they are waiting for a pickup that is not coming.

It ends with two drafts it can write right away, and waits for “go”. That is the pattern you want from an agent in a group: it brings the work to the edge of a decision and leaves the decision to you.

Task 2. Assignments on record, and a reminder

Nora decides who does what and writes it in the topic, so the team sees it and the agent keeps it.

The list comes back in the shape it will be checked in: who, what, by when, and one item still open. The reminder is set for Monday 2 November at 12:30, half an hour after Ana's deadline.

The best line is the one Nora did not ask for. Leo has to write to the customers by Monday, and Tom gets the courier's answer only by Tuesday. If exchanges can still have 24-hour pickups, the message to those customers changes. The agent noticed that the two deadlines come in the reverse order and offered two ways out. A team lead usually sees this when the customers write back.

A one-off reminder like this works on Free. Pro adds recurring reminders and scheduled tasks, so “every Friday at 16:00, post the open assignments in Returns” becomes a standing routine.

Task 3. “Where are we?” the same afternoon

By four o'clock the topic has moved on. Nora asks for the state of the morning list, short.

Four lines and a warning. Mia took Sunday, so nothing is unassigned. Leo, Ana and Tom each get a line with their actual progress. Then the agent comes back to the conflict from the morning, now sharper: 14 customers already have the message, so if the courier allows 24-hour pickups, those 14 need a correction. Its advice is to get Tom's answer earlier.

This is where the list from Task 2 pays off. The agent kept it for the topic, compared it with the new messages and reported only the changes. You do not paste the morning thread again.

Task 4. A teammate asks for a customer reply

The agent works for the whole team once you allow it. Sam has access to the Returns topic and brings a real edge case: boots bought 40 days ago, a 30-day policy, worn once, box intact.

Sam gets a reply he can send: it states the rule, passes the case to the manager as a one-off and asks the customer for the size, so the next step is ready. The agent explains the wording in one line.

Then it gives Nora's side of it. Its recommendation is to grant the exception: the boots can be resold, an exchange keeps the sale, ten days over is the edge of the window. And it names the bigger issue: with 64 open tickets, decide the rule once instead of case by case. The agent also says clearly whose decision it is. Sam gets the words, Nora keeps the call.

Task 5. A pinned message with the rules

The last task is the one that makes the rest work. People need to know how to call the agent and what to bring it.

Five lines, nine seconds. How to call it, how the topics split the work, what to bring, how to ask for a reminder. The note under the draft is worth passing on to the team: in mention-only mode the agent does not see untagged messages, so if you want it to know something, tag it.

The agent offered to post the message itself and waited for the word. Pin it, and the setup is done.

What to hand the agent in a department group, and what to keep

Good jobs for the agent in the group:

• recaps of a busy thread: what was agreed, who took what, what nobody took;
• the list of assignments for a topic, updated from new messages;
• one-off reminders to check a deadline, and on Pro, recurring ones;
• draft replies to customers and partners for the person who has to send them;
• group rules, short instructions, a checklist for a new process;
• a meeting summary: send the call recording or a voice note to the topic.

What stays with the people:

• decisions and exceptions, like the 40-day boots;
• priorities between topics;
• conversations about performance and workload;
• anything that goes out under your name: the agent shows a draft and waits for your yes.

If you want to see a single thread turned into a full task list with a Google Sheet, read AI for tasks: turn a week of team chat into a list of tasks with owners and deadlines. For a project with a launch date, AI for project management: a project status from the team chat in a minute, with risks and who owes what shows the status, the risks and the critical path. And if your team lives on calls, see AI meeting minutes: send the recording of a team call, get decisions, owners and deadlines.

How this differs from a regular group bot

Telegram groups have plenty of good bots: for polls, moderation, welcome messages, reminders on a fixed schedule. They follow commands and do one job well.

The agent works with the content. It reads the messages it is given, understands who promised what, notices when two deadlines conflict, and writes in plain language. It keeps a separate memory for each topic, so “where are we?” in Returns is answered from Returns. It is the same agent you use in your DMs: it knows you, works with files and voice notes, searches the web and, once Google is connected, writes documents and tables in your Drive.

It is your own agent, with a name and avatar you choose, not a shared bot for everyone. It starts in Telegram, and other messengers can be connected. For the full picture of what it does outside the group, see An AI agent in Telegram that does the work.

Tips for the first week

• Start with one topic. Put the agent where the most threads get lost, Returns in our case, and add the others after a week.
• Write assignments as “name, task, date”. The agent keeps the list more precisely when every line has all three.
• Say who you are and what the topic is about in the first message. The agent uses it for every answer in that topic.
• Allow only the people who need the agent, topic by topic. Check the list with /allowed once in a while.
• Tag the agent when you want it to know something. In quiet mode it sees only messages addressed to it.
• Ask for “short” when you want a status, and for “with reasons” when you want a recommendation.

What it costs

Groups and forum topics are included on both plans. Free gives 200  in the first month and 100  every month after, with no card. Pro costs $13.90 a month with 400 , or $1.39 for the first month if you subscribe within 7 days of signing up. Pro adds recurring reminders and scheduled tasks.

In our test, by the balance in the agent's status line, the recap of the thread cost 4 , the other five tasks 1–2  each, about 11  in total. On the default model 20,000 tokens of text cost 1 , and a message later in the conversation costs a fraction of a magnet. The first reply of a new session costs 3–4 , because the agent reads its whole setup, about 90,000 tokens. After that the setup comes from the cache about 50 times cheaper. The full rate card is on the pricing page.

Messages worth copying

Create your agent, add it to one topic of your team group and ask it what was agreed today.