Leads: a client-search pipeline run by agents

Task

I was approached for a well-thought-out chain of agents that together give the staff of a client-search team extra advice, and sometimes make routine decisions themselves. Key decisions and sending stay with the staff: the agents help people as much as they can, but do not replace them.

Clients for bots and automation have to be found where they sell: in the open Telegram channels of shops, showrooms, resellers and wholesalers. By hand that is thousands of channels. What is needed is a pipeline that finds live channels, filters out dead and unsuitable ones, scores each by clear rules and prepares a personal first message based on a recent post. Only open data is used.

Solution

A factory with five belts. Between the belts are warehouses with the funnel counts, on the belts are agent and script machines, and two belts have quality control with a reject chute. The diagram is further down the page; here is what rides on each belt, in order.

• Belt 1, collection. Raw material comes in: open channel catalogues and search, 54 queries across three segments (shops and showrooms, wholesale and markets, resellers). The machine is a collector plus a snowball that pulls mentions of other channels out of posts. Output: warehouse A, 1,270 channels.
• Belt 2, liveness. Control at the entrance: 300 to 60,000 subscribers, clothing and footwear, geography. Rejects go to the reject chute but are never deleted: the tier is lowered and the reason recorded. The machine reads the channel’s public page: a post within 7 days, 3-4 posts in two weeks, 8 signs that orders are taken by hand. Output: warehouse B, 488 live channels.
• Belt 3, profile. Agent machines, the analyst and the lead manager: a channel profile by written rules and five priority tiers, from “write right now” to “not yet”. Output: warehouse C, 210 channels reviewed by hand, 100 leads in the first hundred.
• Belt 4, messages. The copywriter machine assembles a message from six parts: greeting, an observation from a recent dated post, a pain stated as an assumption, a solution and one benefit, a line about experience, one easy question. 350-650 characters, no prices, deadlines, pressure or made-up features; six variants of every part so that no phrase repeats in more than 20 % of messages. Control at the exit: an 8-point self-check, 14 table format checks, an independent sales critic. Rejects go back onto the belt: rewrite. Output: warehouse D, messages ready to send.
• Belt 5, sending. The human gate: an employee’s button in Shtab, no more than 20 first messages a day, one every 3 minutes, a freshness check of the hook before the click. Replied: a client card, the employee takes over, the assistant suggests replies “as me” from the sales cheat sheet. Silent: a follow-up after 3 days, no more than two, then the archive for 30 days.
• The machine crew. 12 agents: head of sales, analyst, lawyer, researcher, copywriter, lead manager, reply triage, assistant, two critics and an external review team. Each has its own model by task complexity and its own boundaries.
• Learning by rules. Every remark from an employee or a critic becomes a check for all messages, so one mistake is not repeated.

What waits for what

• A message waits for liveness and freshness. It is written only for a channel from warehouse B and only from a post within 7 days; if the post goes stale before sending, the message returns to belt 4.
• Sending waits for three things at once: the human’s button, a free slot in the daily quota of 20 and a passed freshness check. If any of the three is missing, the message stays in warehouse D.
• A follow-up waits for 3 days of silence after the last message; there is never a third follow-up, and a new attempt with the same channel comes no sooner than 30 days later.
• A reply stops the follow-up belt: the lead becomes a client, the series goes to the log, the employee answers from there.
• Sending via the server waits for a live bridge: if the personal Telegram bridge on the server is off, Shtab sends by itself, with the same limit of 20.
• A rule waits for a remark: every remark from an employee or a critic becomes a check for all messages on belt 4.

Result

A complete chain was built: search, liveness check, channel profile, personal messages and sending. The agents give staff advice and drafts and settle part of the routine themselves, while sending and key decisions stay with people.