Chat case
AI for tasks: turn a week of team chat into a list of tasks with owners and deadlines
Tasks in a team group get lost between jokes, “will send by Wednesday” and deadlines that quietly move. We copied a week of an ops group, 35 messages from six people, and sent it to our test agent. 27 seconds later it returned every task with an owner, a date and a status, and a separate block for what nobody took. Three more messages updated the list after the weekend, wrote a post for the group and built a Google Sheet. Below are the real replies, the cost and the requests to copy.
The short answer
Copy the week's messages from your team group, send them to your AI agent and write one line: “Pull out every task: who owns it, the deadline, and whether it's done, in progress or still open. Flag anything nobody took.” Add today's date and say who you are. The agent sends back a task list sorted by status, with a name and a date on every line.
Then keep using the same chat. Send what came in after the weekend and ask what changed. Ask for a post for the group, with the orphaned tasks turned into direct questions. Ask for a Google Sheet the team can tick off.
That is what people mean by AI for tasks: the agent reads the thread the way a careful manager would, and you keep the decisions. The agent starts in Telegram and works in DMs, groups and forum topics, so it can also sit in the team group itself.
The test: a week in the ops group of a bakery chain
We wrote the thread ourselves. Brightloaf is a made-up chain of four bakeries opening a fifth, on Canal Street, on Monday 9 November. Sara runs operations. In the ops group with her: Omar from purchasing, Jess who manages the Riverside bakery, Ken from maintenance, Lucy who does hiring and rotas, and Ben from marketing.
One working week, Monday 26 to Friday 30 October, 35 messages. It reads like a real group. Tasks hide inside replies (“Will send by Wednesday”), a deadline moves because a letter arrived, Sara signs off a text a day late, two people joke about a coffee machine, and three things are raised with “someone needs to” or “can someone…”. Some of them get picked up later in the week, and some never do.
A few messages from the thread, so you can see what the agent had to work with:
We copied the whole thread and sent it to our test agent on 30 October 2026, in one message with the request on top. The company and the people are invented. The agent's answers are real and quoted as they came; we trimmed a few words in two replies.
Step 1. A week of messages in, a task list out
The request does three things: says who is asking, asks for a status on every task, and asks for the tasks with no owner. Today's date turns “Wednesday” into a real day.
The list came 27 seconds after the message landed. We checked it against the thread line by line: every owner and every date matches, and the jokes about the coffee machine did not turn into a task.
Three things in this reply are worth more than the formatting.
Statuses come from the thread, not from good intentions. Omar wrote “Then I sign with Millbrook tomorrow”. The agent put the contract under “In progress” and added “No confirmation in the thread that he did.” The same with Lucy's rota: Jess sent the names, but nobody said the rota was built. A list that marks promises as done is the kind that fails a week later.
“Someone needs to” gets its own block. Uniforms for the new hires and the Saturday delivery have no owner, and both are close: the new hires start on Monday. The agent also caught a subtler one. Ben agreed to post on Instagram, but the photos depend on Ken's sign, and nobody owns getting them taken.
The dates are checked against each other. Nobody in the group noticed that the new hires start on Monday and their food safety training is on Tuesday. The agent did, and it saw that the sign, the photos and the post all have to fit into three working days before the opening.
Step 2. Monday: new messages, an updated list
Lists go stale by Monday. Over the weekend and on Monday morning six more messages came in: Omar signed, Lucy handed out uniforms, Jess checked the logs, and Ken's part slipped. We sent the new messages and asked only for what changed.
Twelve seconds. The agent remembered Friday's list and returned the difference: three closed, one moved, one new. It did not repeat the whole week.
The useful part is the consequence of the move. Ken's message only says the part is late. The agent connected it to Jess's message from last week, so the reduced baguette batch now runs until Thursday. And it noticed that Wednesday collects four things at once: the sign, the flyers, the photos and the opening hours. That is the day Sara should keep free.
The Saturday delivery is still in the list with “still no name on it”. A list that quietly drops what nobody took is the reason things fall through. This one keeps asking.
Step 3. A post for the group that gets the gaps taken
Sara needs the team to see the list, and she needs two names. We asked for a post with direct questions.
Thirty-three seconds. The post has three parts, and each answers a question the team actually has: what is mine, what needs a volunteer, what the week looks like. The two open items end with a question, so the natural reply in the group is a name.
The “What the week looks like” block was not in the request. The agent added it, and it is the part people will screenshot: five dates from Wednesday to the opening, in order.
Step 4. The list in a Google Sheet
A post scrolls away. For the run-up to the opening Sara wanted a table the team can update.
Two minutes and thirteen seconds, and the sheet sat in Drive with the link set to edit. Two decisions the agent made on its own make it a working tool rather than a copy of the list. The status column is a dropdown, so people change a status with one tap. And the source column says who wrote the message and when, so any row can be traced back to the thread.
While building the sheet, the agent looked at the list once more and found a third gap: the health inspection on Friday 6 November is in the thread, but nobody said who will be at Riverside for it. That is a fair question for Sara to settle before Friday.
How this works with the agent in your group
In our test we copied the thread by hand: select the messages in Telegram, copy, paste. It takes a few seconds and works with any group, including one where the agent is not a member.
The next step is to put the agent in the group itself, as one more participant. The agent walks you through it: create a group, add your agent, turn on Topics if you want them, and give it admin rights to manage topics. In a group the commands that matter are short:
/mention onkeeps the agent quiet in the everyday flow: it answers only when someone mentions it or replies to it./trust grouplets it keep your teammates' messages in the group, so it has the thread to work from when you ask.- In a group with forum topics, each topic is a separate session with its own memory. One topic per project means one task list per project, with no mixing.
/allow @usernamegives a teammate access to the agent in a topic, so Ben can ask it for the opening hours himself.
Then the request is the same as in our test, written in the group: mention the agent and ask for this week's tasks, owners, dates and gaps.
Reminders work the same way. Ask for a one-off reminder before a deadline and it arrives in the chat where you set it, on every plan, Free included. Pro adds recurring reminders and scheduled tasks: for example, a task round-up from the group every Friday at 16:00, or a Monday-morning list of what is due this week.
What a week of messages turned into
Four messages and under four minutes of waiting gave Sara:
- a task list from 35 messages: done, in progress and raised-but-not-taken, with an owner and a date on every line;
- two timing risks nobody in the group had spotted;
- a Monday update that showed only what changed, including the knock-on effect of a late part;
- a post for the group with two direct asks and the week's key dates;
- a Google Sheet with a status dropdown, a source for every row and a link the team can edit.
Habits that keep the task list reliable
The agent does the reading and the sorting. A few habits on your side make the list solid enough to post straight to the group.
Say who you are and give today's date. “I'm Sara” lets the agent write the post in your voice and keep your own tasks apart. “Today is Friday, 30 October” turns every “by Wednesday” into a real date.
Read the post once before you send it. Messages from different people end up in one text, and a phrase like “I'm away that weekend” belongs to whoever wrote it. In our first draft the agent kept Omar's line in Sara's voice; we trimmed it before posting. Give the sheet the same quick look. It takes a minute and keeps the post accurate.
Ask for the source of each task. “Next to each task, give the name and time of the message it comes from” makes the list easy to check and settles “I never said Wednesday” in seconds. Our sheet has that column.
Send updates, not the whole thread again. The agent keeps the list in the conversation. New messages plus “what changed?” gives a short diff, as in step 2.
One project, one topic. If your group has forum topics, keep each project in its own. The agent keeps memory per topic, so the Canal Street opening and the summer menu never end up in the same list.
What it costs
We read the balance in the agent's status line after each message. The task list cost 3 , the Monday update 2 , the group post 1 and the Google Sheet 8 . That is 14 for the whole case.
Text work on the default model costs 1 per 20,000 tokens, so a weekly round-up of a busy group is a few magnets. The Free plan gives 200 in the first month and 100 every month after, with no card. Pro costs $13.90 a month for 400 , or $1.39 for the first month if you subscribe within 7 days of signing up, and adds recurring reminders, scheduled tasks and top-ups. All rates are on the Pricing page.
Your messages stay in your agent's isolated container. To answer, the agent sends requests to the model: on the default model they go through DeepSeek's infrastructure in the People's Republic of China, which you confirm during onboarding, and on Pro you can attach your own Claude or Codex, so requests go to that provider under your own account.
If your tasks come out of calls rather than chats, the same chat works with a recording: see AI meeting minutes: send the recording of a team call, get decisions, owners and deadlines. And when the month closes, the agent turns the numbers into a deck: AI for presentations: send a spreadsheet, get a quarterly report deck for the owner.
Messages worth copying
Here is this week's thread from our team group. Pull out every task: owner, deadline, status. Flag what was raised but nobody took. Today is [date], I'm [name].
Here is what came in since Friday. Show only what's closed, what moved and what's new, then the open tasks.
Write a post for the group: open tasks by person with dates, and a direct question for each task with no owner.
Put the open tasks into a Google Sheet: task, owner, deadline, status, source message. Set the link to edit.
Next to each task, give the name and time of the message it comes from.
Create your agent, copy this week's messages from your team group and ask who owes what by Friday.
Try it with your own agent
Free, no card. The agent starts working in the chat straight away.
Create your AI agentQuestions
Can an AI pull tasks out of a group chat?
Yes. Copy the messages from your team group and send them to the agent with one line: who owns what, by when, done or open. In our test a week of an ops group, 35 messages from six people, became a task list in 27 seconds: six done, six in progress and four raised with no owner, each with a name and a date.
Does it notice when a deadline moves?
Yes, and it says so. When the delivery of the oven part in our thread slipped from Monday to Wednesday, the update put it under “Moved” and worked out the knock-on effect: the other oven stays in use two days longer, and the baguette batch stays cut until Thursday.
What about “someone needs to…” messages?
That is where the agent helps most. It keeps a block called “Raised, nobody took” and puts every “someone needs to” and “can someone…” there until a name appears. In our test it caught uniforms for new hires, a Saturday delivery and the opening hours for an Instagram post.
Can the agent sit in our team group?
Yes. Add it to the group like any participant. With /mention on it answers only when someone mentions it or replies to it, so it stays quiet in the everyday flow. In a group with forum topics, each topic keeps its own memory, so every project gets its own task list.
Can it put the tasks into Google Sheets?
Yes, once Google is connected. In our test the agent built a sheet with five columns, task, owner, deadline, status and source message, made the status column a dropdown and set the link to edit, all from one message.
How much does it cost?
Our four messages cost 14 in total, including the Google Sheet. The Free plan gives 200 in the first month and 100 every month after, with no card. Pro adds recurring reminders and scheduled tasks, such as a task round-up every Friday.
