What a live WhatsApp AI agent actually does: invoicing, proposals, CRM and client comms from one chat

The starting point: the business already runs on WhatsApp

Field services businesses live in WhatsApp. Engineers confirm jobs there, directors approve things there, clients chase updates there. The admin, meanwhile, lives everywhere else: invoices in Xero, jobs and pipeline in the CRM, proposals in whatever document tool was nearest at the time. Every task means leaving the conversation, opening another system, doing the work, then coming back.

The build fixes that by putting an AI agent inside the channel the team already uses. Nobody learns new software. Nobody logs into a portal. You message the agent the way you would message a very fast, very reliable office manager, and it does the work in the systems behind the scenes.

Invoicing through Xero, from a message

The business invoices for cleaning and installation jobs. With the agent live, an engineer or director sends a message along the lines of "invoice the deep clean at the Croydon site, usual rate" and the agent raises the invoice in Xero against the correct contact, with the correct line items and references. No retyping job details into accounting software at the end of the day, and no invoices going out three days late because admin got buried.

That last point is the commercial one. In trade businesses, the gap between finishing a job and invoicing it is dead cash flow. Closing that gap to minutes rather than days is worth more than most owners expect.

Internal approval routing, without forwarding chains

The agent handles internal communication properly rather than just passing messages along. When a proposal is ready, it goes to the director for approval as a structured request: here is the document, here is what it covers, approve or amend. That is genuinely different from someone forwarding a PDF into a group chat and hoping it gets seen. The approval step is part of the workflow, so nothing goes to a client without sign-off, and nothing sits in limbo because the request got lost under forty other messages.

Client communication: reminders and emails on instruction

The team can instruct the agent to handle routine client contact. "Send a payment reminder to [client]" triggers exactly that, through the right channel: a WhatsApp message on the client-facing number, or an email sent either directly from Xero or through the individual team member's own Gmail account, so it arrives from a real person's address rather than a noreply.

The human-in-the-loop principle holds throughout. The agent does not decide to chase a client on its own. A team member tells it to, and it executes. That distinction is what keeps client relationships owned by people while the admin behind them is automated.

Quotes and proposals: generated, approved, then sent as live documents

This is the part most operators have not seen before. Ask the agent to prepare a quote or proposal and it drafts one using Claude, deploys it via GitHub, and sends back a link to a live, interactive HTML proposal. Not a PDF attachment. A proper web page, branded, readable on a phone, with the scope and pricing laid out cleanly.

The director reviews it at the link, approves it in the chat, and only then does it go to the client. The client receives something that looks like it came from a business twice the size, and the whole cycle from "we should quote for this" to "quote sent" happens inside one conversation.

A CRM assistant in the same thread

The agent also manages the CRM directly. Update the status of a job, add a record, delete one, create a new board for a new workstream, or ask for a dashboard-level read on the pipeline, all by message. "What's still awaiting payment this month?" gets an answer pulled from live data, not from whoever last remembered to update a spreadsheet.

This matters because CRMs fail in small businesses for one reason: nobody updates them. If updating the CRM is as easy as sending a WhatsApp message you were going to send anyway, the data stays current, and every report built on top of it becomes trustworthy.

Set up for four users, extendable in minutes

The system is configured for four users at launch, each with their own identity, so emails go from the right person and actions are attributable. Adding another user takes minutes, not a new project. That is a deliberate design decision: the businesses we build for grow, and a system that needs a consultant every time someone joins the team is a liability, not an asset.

What holds it all together

Under the surface this is Make.com orchestrating Xero, the CRM, WhatsApp and the Claude API, with GitHub and Cloudflare serving the proposal documents. None of it is exotic. What makes it work in production is the plumbing: correct contact matching in Xero, approval gates that cannot be skipped, and predictable behaviour when someone sends a message the agent cannot action. An agent that fails loudly and asks for clarification is useful. One that guesses is dangerous.

We will be publishing a video walkthrough of the live system shortly, showing the full flow from message to invoice, and from quote request to approved proposal in a client's hands.

The takeaway

A WhatsApp AI agent is not a chatbot that answers questions. Built properly, it is the operational layer of the business, reachable from the app the team already has open. The measure of it is not how clever the conversation feels. It is whether invoices go out the day the job finishes, proposals get approved before they get sent, and the CRM tells the truth. That is what this build is going live to do.

Related reading: WhatsApp AI agent for business operations · WhatsApp automated messages for business · How to automate baseline sales operations for an HVAC business