Accounting firm automation usually starts with software: practice management, client bookkeeping that syncs into the firm's ledger, document storage, AI suggestions inside the accounting tool. What software rarely covers is the work between those systems: reading the inbox, checking each tax notice against the return, filing it, chasing the missing receipt, answering the same question for the twentieth time. That's where AI fits, not as one more tool but as an AI employee: an AI role with a fixed job, written rules, its own user account and a human who signs off. I built AI employees like this for a tax advisory firm; on top of that come roles from my own business, and here is how they work in a firm.
This isn't legal or tax advice. The legal points refer to German law and professional rules as of November 2026. The full case is in AI Agents for Tax Firms: A Real Case, other industries on AI by Industry.
What accounting firm automation covers today, and what it leaves out
I split a firm's digital setup into three layers: the firm's own software with document storage; the digital route from client to firm, in Germany mostly DATEV, compared in DATEV Integration: The Basics; and AI built into the software, such as booking suggestions, covered in AI Bookkeeping: What Actually Works Today.
What these layers mostly don't do is work across program boundaries. The German Federal Chamber of Tax Advisers (BStBK) lists in its AI FAQ (as of 27 October 2026) receipt processing, data extraction and pre-checking returns as things AI can automate, and adds that decisions and responsibility always stay with people. An AI employee is built for that gap: it follows a written process description, and what isn't in it, it doesn't do.
Four places where AI fits in a firm
Where
What gets prepared
Who does it in my business
What stays with people
Inbox and intake
sort emails, route requests to the right person, draft replies
Frieda runs in my own business. She sorts the inbox into four piles, Now, Today, Can wait and Trash, plus a fifth pile for cases that need my judgment. She drafts replies and has never sent anything on her own. How the piles work day to day is in Sorting Your Inbox: The Pile System.
In a firm, an AI employee with Frieda's rules sorts requests by topic (the BStBK names tax type, legal form and deadline pressure) and routes them to the right person. Drafts only is more than caution there: under § 3 of the professional code (BOStB), tax advisers sign essential correspondence personally and keep their own responsibility in electronic correspondence too.
Notices and routines: the case from a tax firm
At the tax advisory firm I built AI employees for, two processes run on autopilot: tax assessment notices, digital and by mail, and click payrolls, meaning payroll runs with no change from the previous month. 30 to 40 digital notices arrive there every day, and each went through the office by roughly 45 rules. For the notices, sign-off came only after four clean runs and a phase in which people watched in parallel. Anything uncertain goes to the front office by email.
The AI employees have their own user in DATEV with the rights of a clerk, without filing or payment rights. Details are in the tax firm case.
Client receipts: Peter
Peter has been running in my business since August 19, 2026. In my bookkeeping interface in the browser, he screens newly arrived receipts, assigns each one the correct receipt type, matches the credit card statement line by line and reports what doesn't add up. He doesn't book entries; final approval stays with me and my accountant.
With Peter I sit on the client side, and from there it's clear: if the accountant's side stays entirely manual, the bottleneck simply moves into the firm. In a firm, Peter checks the receipts that arrive through the integration, lists what's missing and drafts the request to the client. The routes out of clients' bookkeeping tools are in Lexoffice Integration: The Main Routes and sevDesk Interfaces: API and Integrations.
Conversations: Gustav
I hand Gustav a call recording. He pseudonymises it first, that's the fixed first step, then distils a pattern register: the questions clients keep asking, each with my clearest reviewed answer so far. The raw transcript never leaves the machine, and he doesn't invent advice.
For a firm, that's the questions clients ask every year, as the basis for a handout or a first-meeting guide. Consent comes first: under § 201 of the German Criminal Code, the BStBK notes, recording private conversations requires everyone's prior consent, and consent after the fact doesn't count. More in Transcribing Calls: What German Law Says and Call Analysis with AI: The Pattern Register.
Professional secrecy: what the rules demand
Confidentiality. Under § 57(1) of the German Tax Advisers Act (StBerG), tax advisers practise independently, on their own responsibility, conscientiously and confidentially. § 5(3) BOStB requires them to make sure unauthorised persons get no view of client records. Unauthorised disclosure of client secrets is a criminal offence under § 203(1) no. 3 of the Criminal Code.
Service providers. An AI provider the firm engages falls under § 62a StBerG. Access only as far as the service requires, careful selection, a contract in text form with a confidentiality obligation, and services abroad only where secrets are protected comparably. Paragraph 5 adds: where a service directly serves a single client engagement, the client has to consent. Data protection law, meaning the GDPR and a data processing agreement, applies on top (paragraph 8).
Own judgment. Under § 3(1) BOStB, tax advisers form their own judgment. Only those authorised by law may provide business assistance in tax matters (§ 5(1) StBerG). So an AI employee prepares, and the professional decides.
Rules inside the firm. The BStBK recommends a written AI policy, an approval process for new tools, and banning unauthorised installations and private AI accounts for firm work. What Claude's plans cover is in Claude Privacy and GDPR: What Applies, how to lock down Claude Code extensions in Claude Code Mods for Tax Firms: Uses, Risks.
In the US, the rule to start with is 26 U.S.C. § 7216: a return preparer who knowingly or recklessly discloses return information, or uses it for anything other than preparing returns, commits a misdemeanor, with exceptions in the statute and Treasury regulations.
How you put this into practice at your firm, we can work out together.
The data lock: pseudonymise before anything reaches Claude
For client data there's what I call the data lock. It sits between your team and Claude. Before a request goes to Claude, it replaces names and other details that point to a person with placeholders: "Ms Schneider called on Tuesday" reaches Claude as "[PERSON_1] called on Tuesday". Claude gets the placeholders instead of the names. The list of which placeholder stands for whom stays on our server in Germany, and when the answer comes back, the data lock puts the real names back in. It works with Claude Code and the AI employees built on it.
The data lock is ready for use and already in operation. I set it up for each client individually and tailor it to their work; what your case needs, we work out together in a call. Pseudonymised is not anonymous: the GDPR still applies. The data lock lowers the risk but takes no obligations off your plate; the contract with the AI provider and the confidentiality obligation under § 62a StBerG are still needed. No automatic detection finds every detail, so when we set it up we also decide which content shouldn't go to Claude at all. How I decide what an AI gets to see is in AI and Privacy: What the AI Gets to See.
Where to start
Count one day. How many emails, notices and receipts arrive, and how long does one take? Without that number you won't see whether anything changed.
Write the AI policy: which data may go where, which tools are approved, who approves new ones.
Pick one process that comes up daily, follows clear rules and is quick to check. Write the rules down the way your most experienced colleague works, plus an emergency route for anything uncertain.
Create a separate user with limited rights, so the log shows what the AI did.
Count runs, then watch in parallel, and only then let it run alone.
No. The BStBK calls AI no substitute for tax expertise but an amplifier, and responsibility stays with the professional. In the firm case, no job was cut.
Can I put client data into ChatGPT or Claude?
Not without more: under German rules you need the contract, the data processing agreement and, for services tied to a single engagement, the client's consent (see above). The BStBK recommends banning private AI accounts for firm work.
Do I need the client's consent?
Under German law, for services that directly serve a single client engagement, yes (§ 62a(5) StBerG). For other services the provision names no consent; paragraphs 1 to 4 apply. Which of your processes fall under paragraph 5 belongs in your AI policy.
Which task should I hand to AI first?
One that comes up daily, follows fixed rules and is quick to check, like the notices and click payrolls in the firm case. Which tasks suit a handover in general is in Which Tasks You Can Hand Off to AI.
What to do next
Before any AI service sees data from your firm, go through the AI privacy check. If you want to look at your first process with me, book a strategy session: in 20 minutes we look at the one workflow in your business that depends on you the most.
Kevin Welter
Developer, IT architect, author of technical books (Kubernetes, cloud infrastructures) and speaker. Runs his business with an AI workforce of fifteen AI employees and shows solo business owners in his community how to hire their first AI employee.
