An unattended sales assistant needs a hand brake. Access to a paid group, an upgrade, a discount, anything that costs the business something: none of it should happen because a language model felt confident. In Scalegram those moments stop and come to you as an alert in your own Telegram, and you answer with a tap or a short reply. That mechanism is described in ask-me checkpoints.
This article is about the unglamorous half of it: how your answer gets attached to the right client.
The failure mode nobody expects until it happens
Picture a busy hour. Four alerts arrive within ten minutes. One asks whether to admit a trader who says he deposited. One asks whether to honour a discount somebody was quoted last week. Two are routine. You read them on a phone, in a queue, and reply "yes" to the third one.
A naive implementation now has a problem. It has a word with no address. The obvious repair is to have a model read your reply, look at the quoted text and work out which request you meant. That works most of the time, which is exactly what makes it dangerous, because the times it fails are the times two alerts look alike. Two people called Andrei, both asking about the same VIP group, twenty minutes apart. A model matching on quoted text will pick one of them, and it will be confident, and there is no signal anywhere that it picked wrong.
Admitting the wrong person to a paid group is a small event with an outsized cost. You either take access back from somebody who did nothing wrong, or you carry a member who never paid.
Binding by message id instead
Scalegram does not infer. Every alert the console sends is filed against the client it concerns and against the Telegram message id of the alert itself. Telegram's reply feature carries that id back with your answer, so a reply is not a piece of text to be interpreted. It is a pointer to one specific outstanding request, resolved by identity rather than by resemblance.
The result is that "yes" is a complete answer. You do not have to name the client, restate the request or worry about which alert was on screen when you started typing. Reply to the alert and the answer lands on that alert, whether it arrived thirty seconds ago or an hour ago while you were away.
The block that tells you who this is about
Binding solves the machine's side. The human side needs the opposite: you should never have to guess who you are approving. So each alert carries a fixed block naming who this is about, along with what is being asked for and the facts the bot is relying on.
What that means in practice is that the alert answers the questions you would otherwise ask before deciding. Which contact. What they want. What the bot has already established, for instance that a UID was checked against a broker connection and came back registered under your affiliate link, as described in purchase sync. Whether a purchase is on the record.
The alert shows the facts the decision rests on, not a transcript. Scalegram does not store the text of your conversations, so an alert cannot quote one back at you. What it can do is link you into the real chat in Telegram, which is one tap away when you want the full context before answering.
What happens after you answer
A yes carries out the action and files it. The client record moves, the grant is recorded, and the bot resumes the conversation with the thing you authorised. A no is equally recorded, and the bot returns to the customer without inventing a reason of its own or promising a review it cannot deliver.
Either way the outstanding request is closed against that message id and cannot be answered twice. If you reply again, the console already knows the question was settled.
Silence is its own state. An unanswered alert stays open rather than defaulting to approval, which is the only defensible behaviour when the thing on the other side is access to something you sell. The queue behaviour, and how to work through a backlog without approving on autopilot, is in reviewing bot alerts as an owner.
Refusals are worth as much as approvals, and for a reason that is not obvious. Every no you send is a data point about where the flow is asking customers for the wrong thing, or granting on evidence that is too thin. A checkpoint that has never produced a refusal in three months is not proof of a well-behaved bot. It is a sign the check is placed where nothing contentious ever happens, and the genuinely contentious moments are going through somewhere else unwatched.
Why this belongs in Telegram rather than a dashboard
The console lives where you already are. An approval that requires opening a web application will be answered in twenty minutes at best, and a lead who is ready to pay does not wait twenty minutes. Putting the request in the same app you read your messages in is what makes an unattended assistant workable rather than theoretical, and it is the same reasoning behind keeping the real conversation in Telegram instead of mirroring it into a separate inbox, covered in the Telegram CRM.
SINGUARD's Executive Directors, Alex Onta & Roman Onta, built this part after watching operators approve requests from a phone on a train. The design target was one tap, on a bad connection, with no risk of the tap landing on the wrong person.
"A model guessing which alert you answered is fine until two clients ask the same question on the same evening. We bind the answer to the message id so it is never a guess."
— Alex Onta, Executive Director, SINGUARD
Key Takeaways
- Anything that grants access or costs money stops and asks the owner in Telegram before it happens.
- Each alert is filed against the client and the Telegram message id, so replying to it resolves that exact request rather than a text match.
- A who this is about block names the contact, the request and the facts the bot checked, so you decide without opening anything else.
- An unanswered alert stays open instead of defaulting to approval, and a resolved one cannot be answered twice.
Frequently Asked Questions
Do I have to type the client's name when I approve something?
No. Reply to the alert message and Telegram carries its id back, which is what identifies the request. A single word is a complete answer.
What if I never reply?
The request stays open. Nothing is granted by timeout, because approval by silence is the wrong default for anything that costs you money or gives away access.
Can the alert show me what the client actually said?
It shows the facts the bot verified rather than the conversation, since message text is not stored in the platform. A link opens the real chat in Telegram if you want the full context before deciding.
About the Author
Alex Onta is an Executive Director at SINGUARD. He built eTrader, the terminal, the mobile apps, eTrader Broker, Copytrading, Business and Community, along with the worldwide clustered-server infrastructure it all runs on, with his brother Roman Onta helping on the design, and he leads that division today. Together with Roman he builds the Prop Firm CRM, the Broker CRM, Scalegram and CopySignals, and the two of them carry worldwide compliance, payment processing and international business structuring side by side. He lives and works in Dubai for most of the year. Meet the executive duo leading Singuard's five divisions.