An AI assistant for the owner in Telegram: what it actually takes off your plate

Almost every owner has the same leaky layer: verbal commitments. Something promised in conversation, something that arrived as a voice note, something discussed in chat and never written down. A week later it turns out the contractor was waiting on a file from you while you were waiting on an invoice from them, and both of you were sure the ball sat on the other side. A calendar doesn't catch this — only what you deliberately entered gets in. A task tracker catches it even less: using one takes a discipline that real businesses rarely have. What follows is about a tool that closes exactly this layer: an assistant in your messenger that remembers commitments and reminds you about deadlines. How it differs from a normal chatbot, what it does over a day, where it errs, how it works inside, and who actually needs one.

This is not a chatbot with clever answers

A normal bot runs on question — answer — forget. You asked, it replied, the story ended. Useful for looking something up, but irrelevant to management: tomorrow it remembers neither your question nor its own promise.

An assistant is built differently: it holds state that survives between conversations. It knows which tasks are open, who holds them, by when, what is already overdue and what people expect from you personally. A conversation isn't a closed episode but a source of changes to that state: you said “I'll send the quote by Friday” and a line appears with an owner, a deadline and a marker that it was promised out loud.

Hence the criterion when choosing one: not how cleverly it answers, but how precisely it remembers. A good-looking answer is available in any chat with a language model for free. Remembering what exactly you promised on the eighteenth is the thing the whole system exists for.

Memory is the hard part, not the easy one

From the outside, remembering looks trivial: write it to a database and you're done. In practice the difficulty sits elsewhere — deciding what counts as a fact. Hundreds of phrases pass through a day of messages and only a few are commitments. Is “sure, I'll look” a promise or politeness? Is “we should get on a call” a task with an owner or a thought out loud?

If the assistant records everything, within a week the list is landfill and you stop opening it — at which point the tool is dead. If it records too cautiously, you lose exactly the commitments you installed it for. The working setting sits in between and is tuned to the person: for one owner “fine, I'll take a look” is a firm yes, for another it is a polite no.

We handle this three ways. First, every record carries an explicit state — who holds it, what we're waiting on, whether the assignment has been lifted. Second, borderline cases are shown before they're written, not after. Third and most important, the system is set to “when in doubt, don't write”. A missed task surfaces and gets entered by hand; ten invented tasks destroy trust in the whole list, and rebuilding that trust costs far more.

A day with the assistant

The morning starts with a brief — a short message timed to when you usually pick up the phone. It carries three things: what is due today, what is already overdue, and what people are waiting on from you. Not a list of forty items, but exactly what needs a decision today. It replaces the familiar routine of opening five chats to remember where you left off.

Through the day the assistant works on reactions. A commitment appears in a conversation — it offers to create the task. A deadline approaches — it reminds the person holding it, not everyone. You reply “done” — it closes and removes the item. One small detail decides everything here: the reminder has to arrive where you already are. A notification in a separate app you open once a week does not work — tested on real people, myself included.

In the evening, or at the end of the week, the other side: what closed, what stalled and with whom. This is the layer that usually goes missing — not “what tasks do we have” but “what was promised and never done, and by whom”. Management conversations grow out of that, not out of a pretty board of cards.

Voice: not decoration, but the condition for it being used at all

An owner is rarely sitting comfortably at a desk. They're in the car, on site, between meetings, phone in one hand. Typing in that situation is friction, and friction is where most task systems die: it's easier not to record it.

So voice isn't a nice option, it's a condition of viability. Dictate on the move — “remind me Thursday to send the contract” — and the task is recorded with an owner and a date. Talk for five minutes after a meeting and you get back a parsed list of commitments to confirm.

The side effect mattered more than we expected: when the cost of recording drops to almost nothing, people start capturing what they used to hold in their heads. Not because they became more disciplined, but because writing it down became cheaper than remembering.

Where it gets things wrong — and what we do about it

Error one: an invented task built out of a polite phrase. Fixed with strict parsing rules and confirmation of borderline cases before writing. Error two: duplicates — one commitment mentioned in three places becomes three tasks. Fixed by checking against the existing list and merging by meaning rather than by matching words.

Error three, the nastiest: the assistant treats a task as open long after it closed and keeps reminding. The person stops trusting reminders and starts scrolling past them — at which point the system has effectively ceased to exist while formally still running. Only one thing helps here: any explicit sign of closure must lift the reminder immediately, with no exceptions.

The general rule we arrived at: in ambiguous situations the system should err toward silence. A surplus reminder irritates more than a missed one, because it damages trust in the entire list at once. And to be plain: the assistant does make mistakes. That is a property of the tool, not an unfinished feature. The question isn't whether errors occur but whether they're cheap and quickly corrected.

How it works inside — without the hype

Technically this is not “one neural network”. It is several parts: a collector that reads what you allowed it to read; a parser that turns conversation into structured records; storage where those records live; and a scheduler that watches deadlines and sends reminders. The language model is responsible only for understanding text — everything else is ordinary code, and that is deliberate: deadlines and task states cannot be entrusted to something that might invent them.

Data lives in your storage, not in someone else's cloud with unclear rules. Access keys stay with you and can be revoked at any moment. If the assistant reads work conversations, you define the scope — specific chats, not everything; private correspondence stays outside by default.

In the commercial version each client is isolated: their own keys, their own storage, no shared pot of other people's tasks. It's the boring engineering part nobody advertises, and it's exactly what decides whether such a tool can be let into a company's working conversations at all.

Who needs this, and who is too early

It fits if you carry dozens of promises in your head, work in messengers, have several contractors or staff, and catch yourself thinking “I've definitely forgotten something, I just don't know what”. The more commitments pass verbally and in chat, the bigger the return: the tool closes precisely the verbal layer that CRMs and trackers never reach.

It's too early if you run one or two projects and it all fits in your head — then it's an extra moving part. It's also too early in the opposite case: if processes already live in a CRM, everyone enters tasks diligently and nothing gets lost. The assistant does not manufacture discipline where there is none — it relieves the person already carrying the coordination.

And a word on expectations. This is not “an employee for pennies” and not a replacement for an assistant. It is a tool that removes one specific class of losses — forgotten commitments and missed deadlines. That is usually enough to pay for itself, but I won't promise that “AI will run your business”: phrased that way, it isn't true.