Polylog FAQ
Frequently asked questions
The questions every team asks before a pilot — answered plainly. If yours isn't here, request a pilot and ask a human.
Is this against Telegram's terms of service?
We answer this one first because you'll ask anyway: automating a personal Telegram account is not something Telegram's terms explicitly bless, and we won't pretend otherwise. Polylog is built to keep an account's behavior within human patterns — real response delays, human-scale message volume, hard rate limits, and an automatic stop if anything starts to look non-human. Every pilot starts with a frank conversation about this trade-off, and we never run an account without the owner's informed consent. [Draft answer — pending legal review.]
What happens if the AI says something wrong?
Several layers exist exactly for this. Anything the creator shouldn't promise is blocked in the prompt and filtered again after generation. During early weeks, every reply sits in a shadow-review window where a human can rewrite or cancel it before it sends. If something still slips through, a human can take over the conversation instantly, and the full audit trail shows exactly what was sent, when, and why.
Does this replace our operators entirely?
No — it changes their job. Instead of typing hundreds of messages a day, your best people review the conversations that matter: approving sensitive replies, stepping in on high-value leads, and closing. One reviewer can oversee the volume that used to take a full shift of operators.
What data do you store, and for how long?
To build a persona we process the creator's exported chat history; to run conversations we store the active dialog state and an audit log of every AI action. Data is retained for the duration of the engagement plus a defined wind-down window, is not used to train shared models, and is deleted on request. The specifics live on our legal page — and in the pilot agreement in plain language.
Which languages and verticals does this work for?
Any language the creator actually texts in — the persona is learned from their own history, not from an English-first template, so slang and mixed-language chats carry over. It fits any funnel where leads expect a personal 1:1 conversation over weeks: coaching, communities, info-products, expert services, personal-brand offers.
See also: how control & safety works andhow we handle data.