An AI agent for tax advisors pays off first where work arrives in the same form every day and gets handled by fixed rules. At a tax advisory firm I built AI employees for, two such processes now run on autopilot: tax assessment notices, both digital and by mail, including filing, and so-called click payrolls (Klicklöhne), meaning payroll runs with no change from the previous month. On the digital side alone, 30 to 40 assessment notices arrive there every day. For the assessments, sign-off came only after four clean runs and a phase in which people watched in parallel. Today the AI employees work on their own there and report back with a completion email at the end, and anything uncertain goes to the front office by email. They cannot transmit filings to the tax office or trigger payments, because their user account in DATEV simply doesn't have those rights.
I run my own business with a workforce of fourteen AI employees, and I build the same kind of employees for clients. What an AI employee is and how it differs from a chat tool is collected on the overview page on AI employees, the full definition in What Is an AI Employee?. The firm stays unnamed in this post, and I describe its AI employees by role, not by name. The sibling page for another industry is AI Employees for Consultants and Agencies.
Why a tax firm is a good place for AI employees
Most businesses have to write their processes down before an AI employee can take them over. A tax firm already has most of that: deadlines, one central system, clear responsibilities, and rules for handling every case. Whether an assessment matches the return, where it gets filed, and who needs to see it isn't decided fresh every day. It exists somewhere, in the head of an experienced colleague or in a work instruction, and can therefore be written down.
That's exactly what makes the handover possible. An AI employee doesn't need creativity for this work, it needs reliability: apply the same rule to every case and report when a rule doesn't fit. Which tasks generally suit that kind of handover, I've laid out as a framework in Which Tasks You Can Hand Off to AI.
Process 1: Tax assessment notices, digital and by mail
Before. Every notice went through the office, by roughly 45 rules. A digital notice took about 2 minutes, the paper mail up to two hours depending on volume. With 30 to 40 digital notices a day, that's not a side issue but a fixed block of office time before the first letter is even opened.
Today. Two AI employees work through the notices. One watches the mailbox and hands over whatever comes in. The other checks the notice against the firm's rules and files it. Each has a narrow job, and that has a practical advantage: an employee who only hands over can't get the checking wrong, and one who only checks doesn't have to deal with the mailbox. Errors can be traced to one place instead of a vague "the AI".
What autopilot means here. Nobody starts the process. It's triggered automatically, runs through, and ends with a completion report by email. Anything no rule covers, or anything uncertain, the AI employee doesn't file by guesswork. It sends it to the front office by email. That's the emergency route, and it matters as much as the normal case.
Process 2: Click payrolls
The second autopilot process is the click payrolls: payroll runs with no change from the previous month. The name already says why they're a good fit. The work there isn't judgment, it's execution, in the same form every month.
The important boundary is built into the term itself. A click payroll is a run with no change. As soon as something has changed, it's no longer a click payroll, and so no longer part of this process.
The path to sign-off: runs first, then parallel, then alone
The assessment process didn't go solo on day one. The path there had three stages, and I consider it the most important part of this example:
- Four clean runs. Only after that was sign-off even on the table. One good run is an event, four clean runs are a pattern.
- Watched in parallel. After that, the AI employees did the work and people watched in parallel. This phase shows whether the process holds on ordinary days too, not just on the days when everyone is looking.
- Alone, with a completion email. Today the process runs without supervision. The completion email says what happened, and the emergency route through the front office catches whatever wasn't clear.
The same principle applies in my own workforce. Peter, my AI employee for receipts, deliberately leaves unclear cases open and reports them instead of forcing them to fit. An employee who hands a case back is worth more than one who guesses.
What else the AI employees take on there
Besides the two autopilot processes, AI employees at the same firm take on further tasks:
Sending out the completeness declarations. Annual financial statements and returns come with a completeness declaration from the client (in German practice, the Vollständigkeitserklärung) that has to come back signed. The AI employee sends it out and keeps track of what's been returned. It's diligence work where the main thing is that nothing gets left lying.
Email monitoring and simple tasks. An AI employee monitors the mailbox and works through simple cases. What counts as "simple" is defined by the firm, not by the AI. How that kind of pre-sorting is set up in general is in AI Email Triage: Agent or Tool?, and in my own office it's done by Frieda, my employee for the inbox.
An AI assistant for staff questions. For day-to-day technical questions, the firm's staff use an AI assistant on the platform of DeutschlandGPT, a third-party provider. The provider advertises itself as GDPR-compliant from Germany ("DSGVO-konform aus Deutschland"), with hosting in German data centers, and states that content isn't used to train AI models. For professionals bound by confidentiality, DeutschlandGPT describes a dedicated mode and names part of its terms of use as the agreement under § 62a StBerG, the German tax advisors act; at the same time the provider itself writes that the service does not replace professional judgment. These are the provider's statements, not my own audit. In practice that means: the answers are a research aid, not tax advice, and the professional review stays with the firm. On top of that come specialized applications we develop ourselves and introductory workshops, so the staff know how to use it day to day.
What this means for the team
No job was cut because of the AI employees. What was saved is work, and the firm can put that time to new use elsewhere. That's the point: work that repeats every day runs at a consistently high quality without a person having to do every step. The person checks the result at the end via the completion email and can focus on other things. The firm is moving into a change, not at the end of one. How to bring your team along on that path is in AI Colleague or Competitor?.
Control: own user, limited rights, audit log
The AI employees run on a server owned by the firm. In the DATEV workplace they have their own AI user. That sounds like a technical detail and is the most important decision in the whole setup: the log shows who did what, a person or the AI. If the AI worked under a colleague's account, that distinction would be gone.
That user's rights match those of a clerk, without rights to transmit filings or make payments. So the boundary isn't just described in an instruction, it's drawn in the system. An AI employee can't transmit anything to the tax office or trigger a payment even if it tried, because its user isn't allowed to.
For a firm, that's three questions to answer before the first run: Which user does the AI work under? Which rights does that user have, and which explicitly not? Where in the log does it show what the AI did? How bookkeeping stays traceable instead of being silently changed is covered in AI Bookkeeping: What Actually Works Today.
Privacy and professional secrecy
Up front: this isn't legal advice, it's a pointer to the rules you should know before client data goes into an AI system. The assessment for your firm belongs with a specialist or your chamber. The rules below are German law.
For tax advisors, unauthorized disclosure of third-party secrets is a criminal offense under § 203(1) no. 3 of the German Criminal Code (StGB). Under § 203(3) sentence 2 StGB, they may disclose such secrets to people who contribute to their professional work, but only as far as necessary. For service providers, § 62a StBerG sets the conditions: access only as far as the service requires, careful selection, a contract in text form with a confidentiality obligation, and for services abroad only if secrecy protection there is comparable. On top of that comes the GDPR with its data processing agreement. How I decide in general what data an AI gets to see at all is in AI and Privacy: What the AI Gets to See.
For my own AI employees that process client data with personal information, I use a business subscription, meaning a commercial Claude product with a data processing agreement; what applies there for training and contracts is in Claude Privacy and GDPR: What Applies. Depending on the need, two further routes come in: Claude via AWS Bedrock in EU regions, described in Claude on AWS Bedrock: EU, GDPR and Cost, and open models, locally or with German providers, see Local AI and Privacy: Cloud or Server? and Your Own AI Server: Hardware and Cost. Which AI access the AI employees at this firm run on, I don't publish. The choice depends on the data and the contracts, not on a general recommendation.
Where a firm should start
Not with the hardest process, and not with five at once. The right first process has three properties: it comes up daily, it has clear rules, and its result can be checked quickly. Assessment notices and click payrolls meet all three, which is why they came first in this case.
Here's how to go about it:
- Write the rules down. In this case there were roughly 45. What only exists in a colleague's head, no AI employee can apply.
- Define the emergency route. Where does anything uncertain go? At this firm, it's an email to the front office.
- Create a dedicated user, with a clerk's rights and without rights to file with the tax office or make payments.
- Count the runs, then watch in parallel, and only then let it run alone.
What an AI employee generally costs, I work through in What an AI Employee Really Costs. A firm setup with its own server and a DATEV user is more than a subscription, and I don't publish the prices from this engagement. Whether an AI employee ends up replacing positions or relieving colleagues is a separate question, which I answer in AI instead of employees?. For businesses without firm software there's the page on AI employees in the trades.
Common questions
Is there an AI agent for DATEV?
DATEV has built in its own AI features, such as booking suggestions in its invoice automation service (as of September 2026, context in AI Bookkeeping). An AI employee is something different: it works with its own user in the DATEV workplace, follows a written process description from the firm, and can work across programs, for example from the mailbox all the way to filing.
Can an AI agent do the tax return?
The groundwork yes, the responsibility no. In Germany, professional help in tax matters may only be provided by those authorized under § 5(1) StBerG, and the information in a tax return must be given truthfully to the best of one's knowledge and belief under § 150(2) of the Fiscal Code (AO). In a firm, an AI employee therefore does the preparation, and the licensed professional decides.
What is a click payroll?
A payroll run with no change from the previous month. Because there's nothing to judge, it suits an AI employee. If something has changed, it's no longer a click payroll.
May a firm give client data to an AI employee?
That depends on the route the data takes and on the contracts. Relevant in Germany are, among others, § 203 StGB, § 62a StBerG for service providers, and the GDPR with a data processing agreement. An AI employee on the firm's own server with its own user is a different case from a chat in a consumer account. This is a pointer, not legal advice; the review for your firm belongs with a specialist.
Do AI employees replace tax clerks?
No, in this case no job was cut. The AI employees take over work that comes up the same way every day: processing and filing assessments, click payrolls, sending documents and tracking returns. Cases no rule covers land with the front office, and professional decisions stay with people. The firm can use the time saved elsewhere. The general answer is in AI instead of employees?.
How to take this further
Take one day and count how many assessment notices arrive at your firm, digital and by mail, and how long one takes on average. Then write down the rules your most experienced colleague uses to handle them. Once that list exists, you're halfway to the process description for your first AI employee. How an AI employee comes together from picking the task to sign-off is in Hiring an AI Employee: The Process, and how I write up and train such processes, I show in my community Claude Practitioners.
