How to build a WhatsApp agent for an accounting firm

January 2026 · 8 min read

In practice, WhatsApp is the number one communication channel between Costa Blanca accounting firms and their clients. More than email. More than the phone. And that means it's also where the most time is lost — answering the same questions over and over.

In this note I'll explain, without jargon but without skipping anything important, how to build an AI agent for WhatsApp in an accounting firm. What it does, what it doesn't, how it integrates with your software, and where AI ends and the human begins.

Which queries can be automated

The first thing is knowing what to leave to AI and what not to. In a typical firm, 70-80% of messages are of the same type:

All of these are automatable. The AI has access to the tax calendar, each client's documentation, the status of their file, and the FAQs the firm has answered a thousand times. It responds in seconds, accurately, 24 hours a day.

What is not automated: any query that requires real advice, professional judgement, or sensitive information that needs human contact. Those cases are escalated to the accountant with full context prepared.

The architecture, in five steps

Without jargon, the flow is:

1. Message reception

The client writes to the firm's WhatsApp as always. Doesn't install anything, doesn't use a different app. For them nothing changes.

2. Classification

The system reads the message and classifies it: question about deadlines, document request, status query, complex query, complaint, other. This classification is done with a language model, and is correct more than 95% of the time if properly trained.

3. Querying your systems

Here's the trick. AI alone knows nothing about your business. It needs to be connected to your accounting software, document manager, client database. When a client asks "did you receive my invoice?" the AI goes to the system, checks, and answers with real data, not made up.

4. Response generation

With information in hand, the AI drafts a response in the client's language, with the firm's tone, citing specific data. If it's a German client, it answers in German. If it's an English one, in English. Without anyone having to translate.

5. Human escalation

If the message exceeds what the AI can handle (complex legal query, a complaint, a decision requiring advice), the agent steps back and hands the case to the accountant with a history summary. The client notices the tone change but not the friction: the conversation continues in the same thread.

What works well and what's tricky

Things that work especially well with this architecture:

And things to be careful about:

How much does it cost? How much do you save?

For reference, a project like this in a mid-sized accounting firm typically takes 4-6 weeks of implementation and costs between €4,000 and €8,000, plus a small monthly maintenance fee. The typical return is saving 15 to 25 hours of staff time per week. In months, not years.

If you want to talk about how something like this would fit in your specific firm, write to me. The first conversation is free and lasts half an hour.