MyCustomerPilotMyCustomerPilot
Log inTalk to salesGet started
Get started

How it works

MyCustomerPilot runs an AI agent on your WhatsApp number. You describe the agent and connect your APIs as tools; we handle the conversation, the model, the tool calls, and the WhatsApp plumbing.

The shape

WhatsApp is the frontend; your API is the backend. There is no app to build and no inference loop to run. You configure an agent and register tools — HTTPS endpoints we call on its behalf — and we sit in the middle: receiving messages, deciding what to do, calling your endpoints, and replying in the chat.

The players

  • Agent — the configured worker: a goal, a persona (tone/style), and the tools and knowledge it can use. One agent answers on one WhatsApp number.
  • Knowledge base — documents you upload; we chunk and embed them so the agent can answer from your content (retrieval-augmented).
  • Tools — your API endpoints, described so the model can call them: check stock, create a payment link, look up an order.
  • Contacts & conversations — the customers messaging you and the threads with them, each bounded by WhatsApp’s 24-hour window.

What happens when a customer messages

Every inbound message runs the same pipeline — you don’t orchestrate any of it:

  1. 1
    Receive. Meta delivers the message to our webhook; we verify its signature and open (or roll) the 24-hour conversation window.
  2. 2
    Build context. We assemble the system prompt from the agent’s goal and persona, retrieve relevant knowledge, and select the tools worth offering this turn.
  3. 3
    Decide. The model reads the conversation and chooses whether to answer directly or call one or more tools (capped per turn).
  4. 4
    Call your endpoints. Each tool call goes to your URL over HTTPS, HMAC-signed so you can verify it came from us, with the result fed back to the model.
  5. 5
    Reply & meter. We send the reply on WhatsApp, persist the turn, fire any webhooks, and record usage for billing.

The 24-hour window

WhatsApp only lets a business message a customer freely for 24 hours after their last message. We track that window per conversation, warn before it closes, and gate outbound sends accordingly — so you never accidentally send outside it.

Where you fit
You bring three things: the agent’s goal and persona, your knowledge, and your APIs as tools. Everything between an inbound message and the reply — the model, retrieval, tool selection, signing, the window, billing — is ours.

Next

Ready to see it move? The Quickstart gets an agent replying in the sandbox in a few requests.