WhatsApp AI Agents
A lead who asks a question on WhatsApp at eleven at night and gets a reply the next afternoon is already cold. Most small businesses know this and cannot do anything about it: the pace of the ads runs ahead of the pace of a human, and every day of delay is money spent on traffic that does not convert. We build a WhatsApp AI agent that replies in seconds in your voice, asks the questions you would ask, qualifies the conversation, and passes the client to you only when they are ready to talk.
What we build
The agent runs on Twilio and connects to your WhatsApp Business number. It replies to every new message in under thirty seconds, asks the questions we define with you during discovery, classifies the lead by intent and budget, and leaves a summary of the conversation in your database or CRM. When a lead crosses the "serious" threshold the agent hands them off to a human on your team with context. The rest get archived without noise.
The code is yours. Node.js, PostgreSQL, deployed on Coolify over Hetzner by default, or on the infrastructure you prefer. No ChatGPT wrappers, no agency middlemen, no lock-in to a no-code platform that later charges you by volume.
How we work
We start with a thirty-minute call where Javier maps your operation: how leads come in today, who answers them, what questions your clients ask, what answers you already have written down somewhere. That produces a base prompt and a conversational flow that Daniel ships the same week.
From launch onward we iterate in production. Every Friday we read the real conversations of the week with you, mark where the agent sounded forced, where it lost the lead, where it over- or under-qualified, and the next week we come back with the flow tightened. We built this for a Spanish fitness coach with 1105 dormant leads, and by the fifth round the agent was holding conversations the coach approved without edits. The full case study is published.
screenshot
The live agent state for the fitness coach: 11 active conversations, 34 leads in the current batch. Each bar is a lead moving through the funnel without anyone typing a message by hand.
FAQ
How do you integrate with our current CRM?
It depends on the CRM. For HubSpot, Pipedrive, or Airtable we write against their API and qualified leads land with the tags we agree on. If you use a Google Sheet or a Notion base that also works. If you have a custom CRM, you send us the docs and we wire up against its endpoints.
What happens when the agent gets something wrong?
Every conversation is logged in the dashboard with the agent reply, the lead context, and a button to flag it as incorrect. That flag goes into the Friday iteration queue. In practice the serious errors get fixed in the next weekly round, and the tone nuances in the two or three after that.
How do you train it to sound like us?
During the discovery call we capture how you reply to the typical messages. From that we write the base prompt with your expressions, your level of formality, and the phrases you already use. The first five production conversations you review line by line and we tune them together.