The first week of running an AI assistant on a real Telegram account produces a predictable list of complaints, and almost all of them are variations of one thing: it said something we do not do. Occasionally the persona is at fault. Far more often the assistant was asked a question nobody had answered anywhere, and it did what models do with gaps.
Scalegram splits what the assistant knows into two stores that behave differently. Facts are things it may answer from, in its own words. Snippets are blocks it sends exactly as written. Getting the split right removes most improvisation.
Facts: written as answers, not as notes
A fact is a short statement of something true about the business: what the product does, what is included, what is not included, how support works, what happens after a purchase, which countries are served. The assistant reads them and answers in its own voice, so the wording does not have to be customer-ready. The content does.
Write them as complete sentences rather than fragments. "Refunds: see policy" is not a fact, it is a note to yourself, and the assistant will either skip it or invent the policy. "Subscriptions activate immediately and fees are non-refundable" is a fact, and it produces the same answer every time regardless of how the question was phrased.
Negative facts matter as much as positive ones. Write down what you do not offer, in the same plain style. Most bad answers are about absent things: a feature the customer assumed existed, a payment method you dropped last year, a market you do not serve. If the absence is not written, the assistant has nothing to refuse with.
Snippets: sent as written
Some text must arrive intact. Payment instructions, a wallet address, the exact wording of a policy line, a link to a bridge page, the onboarding steps in order. Paraphrasing any of those is a support ticket at best.
Snippets are the store for that material. The assistant selects the right one and sends it verbatim, which turns a class of errors into an impossibility rather than an unlikelihood. Keep them short enough to read on a phone, one job each, and name them for the situation rather than the content, so selection is easy.
Anything with a number, an address or a legal consequence belongs in a snippet, not in the facts. Facts are rephrased by design, and the one time you do not want rephrasing is the time it costs you money.
Where prices belong
Prices sit awkwardly between the two, and the answer depends on how much the price moves. Stable pricing works as facts, because customers ask about it in a dozen different shapes and a fact answers all of them. Pricing with tiers, conditions and dates works better as a snippet, because the conditions are the part that gets mangled.
Either way it should live in exactly one place. Pricing copied into a persona, a snippet and a fact will drift within a month, and the assistant will pick whichever it read first. If you run more than one bot, updating the same pricing block into all of them from a single source is the only version that stays consistent, which is the subject of updating pricing and policy.
What does not go in either store
Message content from customer conversations. Scalegram stores no message content at all: conversations are never saved, only contact recency and the link back to the real Telegram chat. That is a deliberate design decision and it means the fact store cannot quietly become a transcript archive.
The practical consequence is that everything the assistant knows about your business has to be written by you, on purpose. Nothing is learned by absorbing what customers said. Some operators find that inconvenient for the first week. It is also the reason the assistant cannot leak one customer's details into another customer's chat, and that trade is not close.
Maintenance is the whole job
The stores are easy to build and easy to abandon. A price changes, a policy changes, a payment method is dropped, and the fact store still carries last quarter's version. The assistant will answer from it without hesitation, because a stale fact and a fresh one look identical from the inside.
The habit worth building is small: when anything customer-facing changes, the fact or snippet changes in the same sitting. Not a weekly review, not a quarterly audit. The same sitting. Teams that do this have assistants that stay accurate for months. Teams that promise to review later end up with a bot quoting prices nobody has charged since spring.
How to find the gaps
Use the handoffs. Every conversation the assistant escalates because it did not know something is a gap with a label on it. Read the handoff queue once a week, and for each one either add the missing fact or confirm that the question should always reach a person.
That loop is the whole improvement mechanism. It is unglamorous and it works, because it fills gaps in the order real customers found them rather than the order you imagined. Within a few weeks the queue thins out to the genuinely human questions, which is exactly where you wanted it. What that queue looks like in practice is covered in handing a conversation to a human.
"Every wrong answer we have ever had to fix traced back to something nobody wrote down. The bot did not invent it out of malice. It filled a gap we left."
— Roman Onta, Executive Director, SINGUARD
Key Takeaways
- Facts are answered from in the bot's own words; snippets are sent verbatim. Anything with a number or an address belongs in a snippet.
- Write negative facts too, because most bad answers are about things the business does not offer.
- Scalegram stores no message content, so everything the assistant knows has to be written deliberately rather than absorbed from chats.
- Update a fact in the same sitting as the change it describes, and mine the handoff queue weekly to find what is missing.
Frequently Asked Questions
What is the difference between a fact and a snippet?
A fact is information the assistant may rephrase when it answers. A snippet is a block of text it sends exactly as written, used for payment instructions, links and policy wording where paraphrasing causes problems.
Does the bot learn from customer conversations?
No. Scalegram stores no message content at all, only contact recency and the link to the real Telegram chat, so everything the assistant knows comes from facts and snippets you wrote.
How do I find out what is missing from the fact store?
Read the handoff queue. Every conversation the assistant escalated because it did not know something points at a specific gap, in the order real customers found it.
About the Author
Roman Onta is an Executive Director at SINGUARD. He builds the Prop Firm CRM, the Broker CRM, Scalegram and CopySignals side by side with his brother Alex Onta, and he helped on the design of eTrader, the division Alex built and leads. His ground is worldwide payment processing, AML compliance and the corporate structures brokers are built on, work the two of them carry together, shaped by executive roles in the UAE and international corporates. He lives and works in Dubai for most of the year. Meet the executive duo leading Singuard's five divisions.