ArlemInsights / Legatus
An AI agent in the group chat: a Telegram walkthrough
What it looks like when an AI agent joins the team as a member rather than a tool: taking dictation, checking the database, issuing documents, and asking before it writes. A short walkthrough with a tea wholesaler as the example.
A member, not a menu
Most "AI features" are a button inside an app the team has to open. An agent is the opposite: it lives where the team already talks, in this case a Telegram group, and it does the work a back-office person would do if the company had one. It has a name, a set of things it is allowed to do, and a database it reads and writes through code.
Below is a compressed version of a real morning for a two-founder specialty tea importer. The numbers are illustrative. The behavior is exactly what runs today.
The walkthrough
Sale · Café Norte · 5 kg Ceremonial @ list · total 1,250,000 ARS
Collection · 750,000 ARS → Galicia ARS · today
Balance Café Norte after: 500,000 ARS (was 0)
Stock Ceremonial after: 41 kg (≈ 38 days at current pace) Confirm?
Tea Shop Palermo · 380,000 ARS · 44 days
Casa Kyoto · 1,200 USD · 37 days
Café Sur · 210,000 ARS · 31 days
Total ≈ 2.0 M ARS at today's rate.
What happened underneath
Each message went through the same four steps.
- Understand. The model reads the message with the company's vocabulary loaded: product names, customers, accounts, currencies. "Galicia" resolves to a specific peso account; "ceremonial" to a specific product.
- Ask the database. Before proposing anything, the agent calls tools that read the record: list price, current balance, stock, days of stock. These are functions in code, not guesses. The same tools are available in Claude through MCP for whoever prefers a desktop.
- Preview. The agent shows exactly what it intends to write, including the consequences (new balance, new stock). This is the moment a human catches "no, that was 4 kg".
- Write, then act. On "confirm", code writes the rows inside one transaction, and only then produces the document, the PDF or the reminder.
Other jobs the same agent takes
- Data entry from documents. A supplier's price list arrives as a PDF or an Excel in the chat; the agent turns it into rows, flags what changed since last week, and asks to load it.
- Quotes and orders. "Price for 20 kg culinary to a new café in Rosario" returns list, discount rule and delivery cost from the database, and an order draft if you say yes.
- Documents. Delivery notes, invoices with tax, statements per customer, produced from the record with the numbering the accountant expects.
- Reading the inbox. Supplier confirmations, bank notifications and customer emails classified and summarized into the same chat, with the relevant rows updated after confirmation.
- Questions. Cash by account, what a customer bought last quarter, exchange difference this month. Same figures as the dashboard, because it is the same code.
What it is not allowed to do
Write without a preview. Talk to anyone outside an allowlist of people and groups. Exceed a rate of actions per minute. Compute a price, a margin or a balance itself; it must call the function. Delete history; corrections are new rows. Every action is logged with who asked, what was proposed and what was written.
The result is an agent the founders trust with money, because it behaves like a careful new employee: it reads the file, shows its work and asks before it commits.