AutomatizaLead

AutomatizaLead does the job of an SDR (sales development rep) on WhatsApp: it imports a list of companies, decides which ones are worth contacting, reaches out, holds the conversation, books the meeting and follows the deal until it closes. The client stays anonymous here.

Prospecting was manual and few messages turned into meetings. The gain wasn’t in sending more messages, but in talking to the right companies, replying quickly and never losing a meeting to a slow answer.

Before writing code, I wrote the rules with the client — ideal customer profile, conversation script, contact frequency, opt-out and scheduling — and the client approved them in writing. Then I built it in phases, each one delivering something that already worked.

1. Choosing companies

   Imports the provider’s list, validates the data and scores each company, with the reason written down.

2. Outreach

   Campaigns with A/B-tested messages that respect business hours and a daily sending limit.

3. AI agent

   The AI reads the lead’s reply and picks the next step from the ones the system allows at that moment (a state machine). Before anything is sent, the code blocks prices, deadlines and legal opinions.

4. Scheduling and human handoff

   The system offers three time slots and holds each one for 60 minutes. When the AI is unsure, the conversation goes to a person, with the reason and the draft.

ALLOWED = {
    "awaiting_reply": {"qualifying", "opted_out", "human"},
    "qualifying":     {"offering_slots", "human", "lost"},
    "offering_slots": {"booked", "human"},
}

def apply_ai_decision(lead, decision):
    # The AI proposes; the system decides.
    if decision.next_state not in ALLOWED[lead.state]:
        return hand_off(lead, reason="step_not_allowed")

    # Prices, deadlines and legal advice are never sent.
    if breaks_rules(decision.message):
        return hand_off(lead, reason="rule", draft=decision.message)

    return advance(lead, decision.next_state, decision.message)

• The system decides, the AI follows

  The AI chooses from a list of allowed steps. It can’t invent steps, book meetings or write time slots.

• Rules in code, not in the prompt

  A prompt guarantees nothing. What must never be said is checked after the AI answers and before anything is sent.

• When in doubt, a person

  If the AI fails, answers in the wrong format or isn’t confident, the conversation goes to someone. No lead is left without an answer.

Built and working in simulation mode, with nearly 400 automated tests. What’s missing doesn’t depend on the code: an approved official WhatsApp number, CRM access and the client’s legal sign-off.

What I took from it

In an AI project, most of the work isn’t the AI: it’s the rules, the permissions, the scheduling and what to do when something goes wrong.

Private client project: no public demo or source code.