A law firm handling personal bankruptcy cases. The flow of ad-driven inquiries was steady, cost per lead was known and tolerable, and roughly one in seven inquiries turned into a signed contract.
The problem wasn't sales — it was the schedule. Someone under debt pressure looks for a solution in the evening and at night — exactly when nobody's in the office. By morning they'd already reached out to three other places, and whoever answered first got the conversation. From the outside it looked like a weak sales team. In reality, sixty percent of requests were lost before the first live contact.
The owner was skeptical of the idea, and reasonably so: the topic is sensitive, the person is in crisis, and an imprecise answer costs reputation and complaints. The condition he set was strict — the assistant answers only from an approved knowledge base, any question outside it goes to a human, and he personally reads every conversation for the first two weeks. We set exactly the same frame on financial projects: a narrow knowledge base, explicit boundaries, a full log.
We built an assistant that ran the entire chain: it took requests at any hour, gave a substantive answer within thirty seconds, gathered the details of the situation — debt amount, assets, number of creditors — created the CRM record with the fields filled in, assigned a task to the lawyer, and booked the person into an open consultation slot. Sensitive or emotionally difficult requests it handed off to a live specialist along with a summary of the conversation.
After two weeks of reading the conversations, the owner dropped the daily oversight on his own. Requests lost before first contact fell from sixty percent to fifteen — what remained were people who wouldn't have continued with a human either. Conversion to a signed contract rose from fifteen to thirty-five percent — not because the machine sells better, but because the lawyer was now reaching people who'd already had a substantive conversation. Additional revenue came out to about three million rubles a month, on the same ad budget.
"I was afraid a robot would be talking to people in a hard situation. Turns out the worse problem was that nobody was talking to them at all until morning."
This is a real case; the company's name is withheld under NDA. It's not a bank — the industry is adjacent: financially sensitive requests, high cost of an imprecise answer, strict knowledge-base boundaries. The mechanics are the same as what we propose for a bank or MFI on the first line: round-the-clock intake, answers strictly from approved sources, escalation to a human with context. At a bank, requirements around the deployment perimeter and logging get added on top, covered below. Your numbers depend on your current response speed, the share of routine questions, and the quality of your knowledge base.