Back to the blog

An AI agent that sent tax advice by itself: four limits to set before you start

· By Bee Flow · 4 min read

A Dutch tax adviser told Taxence how his own AI agent answered client e-mail with advice that read well and was wrong. Nothing said it had to wait for him. Four limits to set first, and how to start small.

An envelope waiting in front of a closed barrier with a check mark above it

On 10 November 2026 the Dutch tax journal Taxence published an interview with a registered tax adviser who had built an AI agent to answer his e-mail. It worked very well, he says: his inbox was empty. Then he read the sent items. The agent had sent clients tax advice that looked convincing and was wrong on the substance.

The cause was small. He had never laid down that the agent could only send a message after he had approved it.

It is a generous story to tell about your own practice, and a useful one. The agent did what it was built for. The limit was missing.

Why this is about your firm too

AI is daily work in Dutch tax practice by now. Research by the NOB, the Dutch association of tax advisers, published on 11 September 2026, had 799 respondents. About 80 percent of junior advisers use AI every day, and 53 percent of seniors. Trust is another matter: 88 percent name the quality of AI output as their greatest concern, and 42 percent name the confidentiality of client data and the GDPR.

The same report shows where the rules are thin. Of the advisers at smaller firms, which the report defines as up to 15 tax specialists, 38 percent say their organisation has no specific AI policy.

A chat window that gives a wrong answer costs you a few minutes. An agent is different, because it acts: it sends, files, replies. That is why the limits come first.

Four questions to answer on paper

The article lists four questions, and they are the right ones.

  1. May the agent send messages by itself? For anything that goes to a client, our answer would be no. Let it write the reply and put it in your drafts. You read it and you press send.
  2. May it give tax advice? An agent that asks for a missing annual statement does different work from one that explains how a scheme applies to this client. Write down which of the two it is allowed to do.
  3. Which sources may it use? Your own knowledge files and the client file are one thing. The open web is another. An agent that can only draw on sources you approved can also show you where an answer came from.
  4. When does a person step in? Name the moments: an amount above a limit, a deadline, a new client, a question the agent is unsure about.

The article adds a rule of thumb that we share. Simple routine work can do with a light check. As soon as an agent does something with consequences for clients, the check has to be heavier.

Start small

The advice in the article for the first weeks is practical. Test with test data and not with current client data. Record what the agent is allowed to do, log what it does, build in human approval and a way to undo things. Make sure someone can stop it at once.

Two more points belong on the list. If the agent processes personal data, the GDPR applies, and an outside supplier needs a data processing agreement. And the responsibility for the advice stays where it was: with the adviser. This post is no legal advice.

If your firm has no AI policy yet, start there. One page is enough, as we described in a one-page AI policy for a small company. For agents in particular, least privilege for AI agents explains how to give an agent only the access its task needs.

Where Bee Flow fits

In Bee Flow these limits are settings. An agent that reacts to new mail in Gmail or Outlook gets one of three levels of mail access: read only, write a draft, or send a reply. With the draft level, every answer waits in your own drafts folder until you send it. In a run started by an incoming e-mail the agent may only use the tools you allowed for such runs.

Automations work the same way. A new automation is inactive until a person switches it on, a dry run shows what it would do without sending or writing anything, and every run is logged step by step. A step that needs a colleague's approval first is available in the paid plans, see plans. Knowledge bases answer from your own documents and show the source with each answer.

The page for accounting and tax firms shows what that looks like for client work, how we work describes how an automation is built and tested, and you can plan a call to go through your own first agent.

#accountancy #automations #ai-at-work