What a German tax advisory firm lets run on its own today

If you search for AI agents for accounting firms, you'll mostly find tool lists and vendor pages, many of them written for US firms. What's often missing is a sober look inside a real firm. Which work does it hand over to AI employees today? How were they trained? And where does a person deliberately stay in charge? That's what we show here, using one example: a German tax advisory firm (a Steuerberatung) where two processes run on their own after training. We don't name the firm. Where this article touches on rules, they are German rules for tax advisors and EU rules that apply in Germany.

Where a firm starts: work that simply has to get done

At a tax advisory firm, there are roughly two kinds of work. The first is the professional work: advising clients, preparing tax returns, checking tax assessment notices, lodging objections, talking to clients. That takes expertise, judgment and responsibility. The second kind is the work around it: opening and sorting mail, filing assessment notices, sorting incoming items, getting documents to the right place. It has to be done every day, follows fixed rules and still ties up time in the office.

That's why we ask the same question for every process: thinking or doing? Where someone's judgment is needed, the work stays with a person or in conversation with an AI employee that person steers. Where something simply has to get done, an AI employee can take over the process and let it run on its own, level by level.

Good first processes at a firm have three things in common:

• They come up often. Every day, not once a quarter.
• They follow rules. Even if there are many variants, each variant can be described.
• They arrive digitally or can be made digital. An assessment notice as a PDF, an email with an attachment, a scanned letter.

Filing tax assessment notices ticks all three boxes. In our example, it was the first process.

The example: a tax advisory firm, two processes

The firm has two AI employees: one AI employee for filing and a second one for the inbox. They share two processes:

• Process no. 1: filing digital assessment notices. The first AI employee handles this on its own.
• Process no. 2: filing assessment notices that arrive on paper. Here, both work together. The second AI employee finds the paper notices in the inbox and hands them over; the first one files them.

Both processes are at Autonomy Level 2, Autopilot, as of September 2026. By our definition, that means:

• Autonomy Level 2, Autopilot: The AI employee starts on its own on a schedule or trigger, runs through to the final report and hands anything uncertain to a person.

What the other levels mean and how a process moves from one to the next is explained in The four autonomy levels: how a process in your business runs on its own.

Process no. 1: filing digital tax assessment notices

Before, the office team filed every notice by hand. Up to 30 to 40 notices came in digitally per day, and the office team needed about 2 minutes for each one.

That sounds like a simple task. It is, but only in the standard case. The assessment notice process distinguishes around 45 rules in total, for the digital notices and the ones on paper together. They come from several directions:

• Many types of assessment. Each type of assessment has its own filing rule.
• Different types of tax and types of letter. Not every letter from the tax authority gets filed the same way as an assessment notice.
• Letters from the city on paper, with different layouts. What sits at the top right in one letter is somewhere else entirely in the next.

These variants are exactly what makes the process demanding. For the office team, they're routine. An AI employee has to be taught every single one. That's why the list of variants is the core of every process description. How to write a process down so an AI employee can take it over is covered in Documenting processes so an AI employee can take them over.

Process no. 2: assessment notices on paper

Assessment notices that arrive on paper don't reach the filing on their own. They come into the firm's inbox by email, among lots of other emails. Before, filing them took the office team up to 2 hours, depending on volume.

That's why the second AI employee belongs to process no. 2. It checks the inbox every 30 minutes, spots the emails with paper notices and hands them over to the first AI employee, which files them. So the task has two steps: the inbox spots and hands over, filing matches and files.

The reason for the split: the paper notices have a different start. That gives each AI employee a clearly defined task. In our program, that's precisely the goal for the second month: process no. 2 comes on board, and where a task passes through several steps, the AI employees hand it off to each other.

How the training worked: four steps

No AI employee runs on its own from day one. At the tax advisory firm, training happened in four steps that we apply the same way to every process:

1. Watch. First, Kevin watched the office team: which types of assessment there are and how each one gets filed.
2. Sit next to it. Then the AI employee was trained, one type of assessment at a time, with someone sitting next to it. Each variant was demonstrated once.
3. Remember and ask. The AI employee saves every variant it has learned. Anything it doesn't know goes to a person through the escalation path and gets trained in the feedback session.
4. Sign off. It was signed off after four clean runs. After that, the team watched alongside it for a while. Only then did the process run on its own, and since then it has reported back with a completion email.

At this tax advisory firm, it took 2 weeks from the first round of watching to sign-off. That's what happened at this one firm, not a time promise for yours.

The first steps correspond to Autonomy Level 1: someone starts the AI employee, and the team checks every run. What happens on this level and why it isn't skipped is described in Autonomy Level 1, Assisted: when your team checks every run.

What happens when the AI employee is unsure

No process knows every case in advance. Sooner or later, a type of assessment comes along that the AI employee hasn't seen, or a letter from which the filing can't be worked out. That's what the escalation path is for.

Anything the filing employee can't match with certainty goes to the office team by email. It doesn't guess. The office team decides the case, and in the feedback session, the new variant gets trained. Next time, the AI employee knows it.

Then there's the report. After every run, filing reports back with a completion email. Whoever reads it can see what's done and what's still open, without having to look through the filing system themselves.

The escalation path and the report are the reason a process on level 2 may run on its own. The process is managed: it has rules, a contact person and a path for anything it doesn't know. More on level 2 in Autonomy Level 2, Autopilot: the process starts on its own and reports back at the end.

Where the line is: advice and responsibility stay with people

The two AI employees in this example file and hand over. They don't give advice.

Responsibility stays with the tax advisor. The German Tax Advisors Act (Steuerberatungsgesetz) requires tax advisors to practice their profession on their own responsibility (Section 57 StBerG), and the German Federal Chamber of Tax Advisors (Bundessteuerberaterkammer) writes in its AI guidance: AI is a tool, and the professional remains responsible for the result. That's why our AI employees at the firm handle incoming items, sorting and filing. The professional work and the final check stay with people.

That filing involves no professional assessment is our own classification, not a statement by the chamber. Whether a notice is correct in substance, whether an objection is filed and what a client should do is decided by a person at the firm.

Confidentiality and data protection

Tax assessment notices contain confidential client data. Confidentiality applies to AI too. Under German law, anyone who lets an AI provider work with client data needs, among other things, a contract in text form with a confidentiality obligation (Section 62a StBerG), has to choose the provider carefully and, for services abroad, has to check whether protection there is comparable. Data protection comes on top, usually with a data processing agreement. What that looks like at a specific firm, we work out in the project with the firm and its chamber, not on a blog.

We don't give legal advice. For individual cases, the firm's own chamber or its own legal counsel is the right place to ask.

AI literacy on the team

One point concerns the whole team. Anyone using AI systems at the firm has to take measures under the European AI Act to support the development of AI literacy among their staff (Art. 4 AI Act, newly worded since July 2026). The law doesn't require a specific level.

In daily work, part of this literacy comes from training. The team watches how the AI employee works, checks its first runs, decides the uncertain cases and teaches it new variants in the feedback session. Anyone who has done this once knows what the AI employee can do, where it becomes unsure and when a person has to take over.

What further measures a firm takes is its own decision. We don't assess that.

What your firm can take from this

The example can't be copied one to one. Every firm has different clients, different workflows and a different office team. The approach carries over, though:

1. Draw up the map. Write down every recurring process at the firm with the team: who, how often, how long, through which input.
2. Separate thinking from doing. Advice and professional review stay with people. The levels are for processes that simply have to get done.
3. Pick process no. 1. The process that saves the most time or money, or brings in the most revenue, with little risk, that comes up often and follows rules.
4. Collect the variants. Watch the team and write down every variant before training begins.
5. Set the escalation path and the report. Who gets the uncertain cases? Who gets the report?
6. Train, and only sign off after clean runs. Then watch alongside it, then let it run on its own.

If you're more interested in the bookkeeping itself, meaning receipts, account assignment and matching, Kevin's blog has a separate article on that: AI Bookkeeping: What Actually Works Today.

What this looks like for us

This whole article is our example. The current state in brief:

• What's running: Two processes, filing the digital tax assessment notices and the notices on paper, are at Autonomy Level 2, Autopilot, as of September 2026. They start on their own and report back with a completion email.
• Who does what: One AI employee files, a second one checks the inbox. The office team gets the uncertain cases.
• How long it took: 2 weeks from the first round of watching to sign-off.
• What isn't running yet: As of this state, the two processes haven't reached Level 3, Audited, or Level 4, Autonomous.

In our program, level 3 is planned from month 3: an AI manager bundles the reports into a morning report and flags exceptions. Kevin tries this level in his own business first, before it comes to yours. In Kevin's own business, an AI manager reviews the work of his AI employees. That role is still on trial.

Frequently asked questions

Does AI replace the office team at a tax advisory firm?

Not in our example. Filing runs on its own, but the office team decides the uncertain cases, and in the feedback session, the AI employee learns new variants.

May a firm let an AI employee process client data?

What that requires under German law is described above in the section "Confidentiality and data protection". We don't give legal advice.

How long does it take until the first process is signed off?

At the tax advisory firm, it took 2 weeks from the first round of watching to sign-off. In our program, the goal for the first month is for the first process to start on its own and report back with a completion email. How quickly it goes at your firm depends on the number of variants and on how much time your team brings. That's why it's a goal, not a promise.

Which other processes are suited to a firm?

Any that come up often, follow fixed rules and arrive digitally or can be made digital. Which ones those are at your firm, the map will show. Advice, review and decisions with consequences for clients aren't among them.

Which process would be no. 1 at your firm?

At many firms, work that simply has to get done crosses the office every day. On the call, we look at whether Owner Autonomy is a fit for your firm and which process is worth starting with. That's also where we go over how it works and what it costs.