AI for Tax Advisory Firms: How German Steuerkanzleien Beat the 10,000-Position Shortage With Custom AI Agents

On any given Tuesday morning in a German Steuerkanzlei, a Steuerfachangestellte opens DATEV Unternehmen online for the fourteenth Mandant of the week. There are 42 new Belege in the Postkorb, a Lohnabrechnung that needs to go out today, a USt-Voranmeldung that should have been done yesterday, and an email from a client that starts with “ich hatte nur eine kurze Frage”. The deadline is, as always, gestern.

The Mittelstand has been talking about a Fachkräftemangel for a decade. Inside the German tax-advisory profession, the math has finally caught up with the talk. The Bundessteuerberaterkammer reports roughly 105,000 active Steuerberater serving more than 3 million businesses, and the DStV and Visionarydata count over 10,000 unfilled positions across roughly 65 percent of firms that report acute shortages³. Hiring has stopped being a strategy. Throughput has become the constraint.

This guide is for the Kanzleiinhaber, the Sozius, or the Steuerfachwirt who has watched the AI hype cycle for two years and now wants a concrete answer to one question: can a custom AI agent actually take over Belegerfassung, Mandantenkommunikation, USt-Voranmeldung prep, and the Buchhaltungs-Monday inside DATEV, and if so, how?

The Kanzlei Crunch

The reason tax work consumes so much time is rarely the tax law itself. It is everything around it: chasing missing Belege from Mandanten, deciphering supplier formats, classifying expense photos, looking up cost centres, translating a free-text email into a booking, drafting Mandantenkorrespondenz, and re-confirming what the Mandant already confirmed three weeks ago. Multiply that by 30 Mandanten and the math gets ugly fast.

• 10,000+ unfilled positions - The DStV and Visionarydata together count more than 10,000 open Stellen across German Steuerkanzleien, with around 65 percent of firms reporting acute Fachkräftemangel³.
• Generation change at the top - The BStBK reports that more than 35 percent of practising Steuerberater are over 55, and Kanzlei-Nachfolge has become its own crisis category - many Kanzleien close not because they fail, but because no one wants to take over without modernisation¹.
• The 41 percent AI threshold - Bitkom’s 2026 Pressemitteilung “Durchbruch bei Künstlicher Intelligenz” finds that 41 percent of German firms now use AI actively, up from 17 percent in 2024. Within professional services - including Steuerberatung - adoption is even higher¹⁶.
• Manual cost per invoice - Industry benchmarks put fully manual AP at EUR 10 to 15 per invoice; AI-assisted processing brings this to roughly EUR 2 or less - an 80 percent saving for high-volume Buchhaltung²⁰.
• OCR is now solved - Modern AI extraction reaches 97 to 99 percent accuracy on header fields (Lieferant, Datum, Betrag, USt) and adapts to layout changes without template rebuilds²¹.
• The Mandant is changing too - SME owners have been trained by ChatGPT to expect answers in seconds. The Kanzlei that takes two weeks to reply on a routine question loses retention to one that replies in two hours.

The result is a Kanzlei where the most senior Steuerberater spend most of their time on the most repetitive work, while Mandantenberatung, Bilanzanalyse, and Jahresabschlussplanung get squeezed into the last hour of the day. AI agents do not fix this by being clever. They fix it by removing the repetitive 60 percent that nobody, on either side of the desk, wants to do.

What an AI Agent Actually Does in a Steuerkanzlei

The market is full of products labelled as “AI für Steuerberater”. Most of them are OCR with a polish. A real AI agent goes much further: it owns the workflow, not just the data extraction. Here is the difference, in plain Kanzlei terms.

The agent loop in a Kanzlei

1. Capture - Pulls documents from Mandanten-Postkorb, DATEV Belegtransfer, email inboxes, supplier portals, scanner folders, the bank feed, and the Mandanten-App.
2. Classify - Reads the document - PDF, XRechnung XML, ZUGFeRD hybrid, photo, screenshot - identifies the Mandant, the Beleg-Typ, and routes it to the right pipeline.
3. Extract - Pulls structured fields: Lieferant, Datum, Brutto, Netto, USt-Satz, Konto, Kostenstelle, Projekt - with a confidence score for each.
4. Enrich - Looks up the Lieferant in the Mandanten-Stammdaten, checks against open Bestellungen, applies the Kanzlei-specific Kontenrahmen rules.
5. Propose - Generates the full Buchungssatz (Soll, Haben, USt-Schlüssel, Kostenstelle, Projekt, Fälligkeit) for the Steuerfachangestellte to approve.
6. Decide - Above the confidence threshold the agent posts directly into DATEV; below it the case lands in a human review queue with a one-click approve or correct.
7. Learn - Every correction feeds back into the agent. Patterns it sees three times for a specific Mandant become rules it applies automatically.
8. Audit - Every step is logged with timestamp, document hash, agent version, and (on review) the human who approved it - the GoBD audit trail is built by the system, not by the Steuerfachangestellte after the fact.

The difference from what you have today

The category matters. DATEV Copilot, available in the DATEV KI-Werkstatt since February 2026, is a horizontal assistant for tax firms - text creation, document analysis, prompt library, research⁵. It is excellent at making people faster. A custom agent goes one layer further: it removes the work entirely from the human review path until something looks unusual or judgement-dependent.

6 Use Cases That Work Today

Not every Kanzlei process is a good first AI candidate. The ones below are - they are high volume, repetitive, well documented, and have a clear correct answer. Start with one Mandant in one use case. Add the next once the first runs cleanly.

1. Belegerfassung (Document Capture and Booking)

The biggest queue and the easiest first win. An agent reads incoming Belege from the Mandanten-Postkorb, classifies them, extracts header and Position items, matches against open Bestellungen, and posts the Buchung into DATEV with the Kostenstelle already assigned.

• Volume is high - A typical Mittelstand-Mandant generates 300 to 1,500 Belege per month, mostly from the same 50 to 200 recurring Lieferanten
• Patterns are stable - Once the agent has seen a Lieferant three times in a specific Mandant, it books the fourth one automatically
• ROI is immediate - EUR 10 to 15 manual cost per Beleg drops to EUR 2, and Buchungs-Zykluszeit falls from 10+ days to 1 to 3²⁰
• Mandant side effect - Faster Buchung means the Mandant sees an up-to-date BWA, not one from six weeks ago
• Compliance side effect - Every Beleg gets timestamped, hashed, and archived to GoBD spec on entry¹³

2. Mandantenkommunikation (Email and Inquiry Triage)

The hidden time-sink. A 10-person Kanzlei receives 200 to 400 Mandanten-emails per day. Most are routine: “wann ist meine Lohnabrechnung fertig”, “wie war nochmal mein Steuersatz”, “ich habe noch eine Frage zu dem Beleg vom 14.”. An agent reads each email, classifies the intent, drafts a reply grounded in the Mandantenakte, and queues it for human send.

• Intent classification - Frage zur Steuererklärung, Beleg-Nachfrage, Fristverlängerung, Vollmachtswechsel - each routes differently
• Grounded answers - The agent looks up the Mandantenakte before drafting, so it answers from facts, not from invention
• Tone matching - The draft uses the Kanzlei’s standard tone, not a generic AI voice
• Human-send guardrail - Nothing leaves the Kanzlei without a human pressing send; the agent prepares, the Berufsträger reviews
• Time saved - 40 to 60 percent of email handling time disappears within the first month

3. Lohnvorbereitung (Payroll Prep)

Payroll is high-stakes, high-frequency, and dominated by the same data exchanges every month. An agent collects the inputs (Stundenzettel, Abwesenheiten, Bonus-Tabellen, neue Mitarbeiter, Kündigungen), validates them, applies the rules each Mandant uses, and prepares the LODAS or DATEV Lohn payload for the Lohnsachbearbeiter to confirm.

• Input aggregation - Excel, scanned Stundenzettel, Zeiterfassung-Exports, email mentions - the agent pulls them all into one structured input
• Plausibility checks - The agent flags Mitarbeiter with unusual deviations from prior months (e.g., 200 hours when the norm is 160)
• Master-data hygiene - New hires, leavers, Lohnsteuerklasse changes get logged with the source document
• Pre-validation - The agent runs the inputs through DATEV-equivalent plausibility checks before the Sachbearbeiterin opens the file
• Time saved - Lohnvorbereitung that previously took 2 to 3 hours per Mandant per month drops to 30 to 45 minutes of review

4. USt-Voranmeldung Preparation

The monthly or quarterly USt-Voranmeldung is mostly mechanical, but easy to get wrong: § 13b, Reverse Charge, innergemeinschaftliche Leistungen, Reisekosten with split VAT - the edge cases multiply. An agent prepares the Voranmeldung, flags the edge cases, and hands the Steuerberater a draft to sign.

• Automatic Konten-Aggregation - All VAT-relevant Konten aggregated by USt-Schlüssel with cross-check to the GuV
• Reverse-charge handling - § 13b Sachverhalte automatically classified with the rule reference
• Innergemeinschaftliche Lieferungen - Cross-border line items aggregated with USt-IdNr validation
• Skontoberichtigungen - Skonto-changes that affect VAT recomputed automatically
• ELSTER-ready output - Output formatted for direct ELSTER submission after Steuerberater approval

5. Jahresabschluss Vorbereitung

Year-end is when the Kanzlei hours peak and the senior team disappears for two weeks. An agent prepares the Vorarbeit - Abstimmung der Konten, Abgrenzungen, Sonderposten, Anlagenspiegel - so the Bilanzbuchhalter and Steuerberater spend their time on judgement calls, not on data prep.

• Konten-Abstimmung - Open items per Lieferant and Kunde, Saldenbestätigungen drafted automatically
• Abgrenzungen - Recurring accruals (Miete, Versicherung, Wartung) prepared with the right reversal entry for the new year
• Anlagenspiegel - Zugänge, Abgänge, AfA recomputed and presented for review
• BWA draft - First draft of the year-end BWA ready before the team sits down on 5 January
• Time saved - Jahresabschluss-Vorbereitung that previously took 8 to 12 hours per Mandant drops to 3 to 4 hours of review

6. Kanzleisteuerung (Practice Management Insights)

Most Kanzleien fly blind on their own margins. An agent watches time logged per Mandant, fees billed, exceptions handled, and DATEV postings made, then produces a weekly Kanzlei-Kennzahlen-Dashboard - which Mandanten are profitable, which are slipping, which are about to need a fee renegotiation.

• Mandanten-Profitabilität - Time spent per Mandant in EUR vs. fees billed, surfaced weekly not yearly
• Throughput per Steuerfachangestellte - How many Buchungen, USt-Voranmeldungen, Lohnabrechnungen per person per week
• Exception trending - Mandanten where the share of unusual postings is rising - usually the first sign of operational change
• Fee-renegotiation triggers - Mandanten where time-spent has grown more than 25 percent year-over-year
• Capacity forecasting - Projection of next-month workload based on year-over-year patterns

The DATEV-First Architecture

In Germany, the Steuerkanzlei stack is dominated by DATEV: roughly four of every five Kanzleien run on it⁵. Any AI agent that wants to be useful in a Kanzlei has to live on top of DATEV without trying to replace it. Here is what the working architecture looks like.

The five integration layers

1. Mandanten-Eingang - DATEV Belegtransfer, DATEV Unternehmen online, scanner folders, Mandanten-App, email, supplier portals - everything routed into one normalised inbox.
2. Klassifikation und Extraktion - The agent identifies the Mandant, the Beleg-Typ, and extracts the structured fields with confidence scores.
3. Buchungsvorschlag - The agent applies the Mandanten-specific Kontenrahmen (SKR 03, 04, BGA) and proposes the full Buchungssatz.
4. DATEV-Schnittstelle - Approved Buchungen flow into DATEV via the standard import interface or the new DATEV API for Pilot-Kanzleien; nothing bypasses the system of record.
5. Audit und Reporting - Every action is logged with timestamp, document hash, agent version, and reviewer name - the GoBD audit trail and the Kanzleisteuerung dashboard share the same event log.

What sits where

The principle is simple: DATEV stays the system of record. The agent is the operator. Any architecture that tries to replicate DATEV in a side-system breaks GoBD on day one and breaks the Steuerberater’s sign-off on day two.

Where the data lives

• Mandantendaten - In DATEV, untouched. The agent works on a controlled copy via DATEV-approved interfaces.
• Agent state - In an EU-resident environment (Frankfurt, Berlin or similar) operated under a BStBK-compatible AV-Vertrag.
• LLM inference - Either via EU-resident endpoints (Azure OpenAI EU, AWS Bedrock EU, Anthropic Claude on AWS EU, or Mistral) or on a Kanzlei-private deployment for sensitive segments.
• Prompt/output retention - Logged for audit. Not used for vendor model training - the AV-Vertrag must say so explicitly.
• Backups - In EU, encrypted at rest, with documented retention aligned to GoBD requirements (10 years for relevant Belege).

The 60-Day Pilot Playbook

The biggest mistake Kanzleien make with AI is starting with a strategy. The right path is the opposite: start with one Mandant, one use case, and 60 days.

The phases

1. Days 1-10: Scope - Pick the Mandant. The right one is mid-size (300-1,000 Belege per month), recurring patterns, and a CFO or Geschäftsführer who is willing to be the first pilot client. Map the current Belegerfassung process step by step.
2. Days 11-20: Datenfundament - Audit Mandantenstammdaten, Lieferantenliste, Kontenrahmen, recurring Buchungsmuster. Clean the obvious mess (duplicate Lieferanten, ambiguous Konten). The agent will be only as good as this input.
3. Days 21-35: Build - Connect the agent to DATEV via Belegtransfer or the API. Train it on the last 90 days of bookings. Run a shadow mode where every booking is generated but not posted - the Steuerfachangestellte sees proposals next to her own bookings.
4. Days 36-50: Parallel Pilot - Go live with human approval required on every Buchung. The agent posts; the human reviews. Track accuracy, time-per-booking, and the exception rate. Tune the confidence threshold.
5. Days 51-60: Confidence Raise - For Lieferanten where the agent has been right 20+ times in a row, auto-approve. For new or unusual cases, keep human review. Document the Verfahrensdokumentation.
6. Day 60+: Expand - Once one Mandant runs cleanly, add the next two. Once five Mandanten run cleanly, add the second use case (typically Mandantenkommunikation triage).

Checklist before you start

What to measure

• Time per Beleg - Baseline manual time (typically 90 to 180 seconds), target with agent (typically 15 to 30 seconds of review)
• Accuracy - Booking proposals that the human accepts unchanged - target 80%+ after week 4, 95%+ after week 8
• Exception rate - Belege the agent kicks back as low-confidence - track the trend, not the absolute
• Cycle time - Beleg-Eingang to fertig gebucht - target 1-3 days
• Mandantenzufriedenheit - Spot-check the pilot Mandant after 60 days - up-to-date BWA, fewer back-and-forth Belege-Anfragen

BStBK, Verschwiegenheit and the EU AI Act

Compliance is the make-or-break in a Kanzlei. Every other consideration sits below it. The Steuerberater profession has three overlapping rule sets that all touch AI: BStBK Berufsrecht, GoBD, and the EU AI Act. Each has a concrete answer; none of them is a blocker.

BStBK and § 57 StBerG

The Verschwiegenheitspflicht in § 57 StBerG is the spine of the profession. The BStBK’s 2025 Berufsrechtskommentar explicitly addresses AI tools: they are permissible where the Steuerberater retains the Verantwortung and the technical arrangement preserves Verschwiegenheit. § 203 StGB makes the criminal-law side explicit - unauthorised disclosure of client data is a Straftat.

• Hosting in the EU - All Mandantendaten processed by the agent stay in EU data centres. No US-CLOUD-Act exposure under Schrems II.
• AV-Vertrag with Berufsgeheimnis clause - The Auftragsverarbeitungsvertrag must explicitly cover Berufsgeheimnisträger and reference § 203 StGB and § 57 StBerG.
• No training on Mandantendaten - The vendor agreement must include an explicit no-training clause covering prompts and outputs.
• Berufsträger remains responsible - The Steuerberater signs off on every Mandanten-Output. The agent prepares; the human approves.
• Audit logs available to Mandanten on request - On request, a Mandant can see exactly which AI actions touched their data.

GoBD - the bookkeeping rulebook

GoBD (Grundsätze ordnungsmäßiger Buchführung und Datenzugriff) applies to every booking the agent makes. The three Grundsätze - Nachvollziehbarkeit, Unveränderbarkeit, Vollständigkeit - have concrete technical translations¹³.

• Nachvollziehbarkeit - Every booking has a logged source document (Beleg), an extraction record, a proposal record, and an approval record
• Unveränderbarkeit - Once a booking is posted, it can only be reversed, not edited. The Beleg-Archiv is write-once.
• Vollständigkeit - The agent must process every Beleg in scope; nothing silently dropped
• Verfahrensdokumentation - Written documentation of how the agent works, when it auto-books vs escalates, what model it uses, who reviewed the deployment¹⁴
• Aufbewahrung 10 years - Belege, logs, and Verfahrensdokumentation all retained for the statutory 10 years

EU AI Act - what applies, what does not

The EU AI Act becomes fully applicable in August 2026. For a Steuerkanzlei deploying AI agents, the relevant articles are mostly limited-risk: transparency, AI literacy under Article 4, and the general-purpose model obligations on the vendor side. The agent itself is typically not a high-risk system unless it makes decisions about employment, credit, or essential services.

• Article 4 (AI Literacy) - Every Mitarbeiter using the agent receives basic AI literacy training, documented²⁵
• Transparency obligations - Mandanten are informed that AI is used in the preparation of their work; the Steuerberater remains the responsible person
• Provider obligations - The LLM vendor (OpenAI, Anthropic, Mistral, etc.) carries the general-purpose model obligations; the Kanzlei is the deployer, not the provider
• No high-risk classification - Belegerfassung, USt-Voranmeldung prep, and Mandantenkommunikation are not Annex III high-risk use cases
• Documentation - The same Verfahrensdokumentation that satisfies GoBD also covers the EU AI Act provider/deployer documentation

The Growth Question: More Mandanten With the Same Staff

The Fachkräftemangel is not going away. Hiring two more Steuerfachangestellte is harder, slower, and more expensive every year. The question that matters for Kanzleiinhaber over the next three years is not “wie spare ich Zeit” - it is “wie wachse ich ohne neue Köpfe”.

What an agent does to the capacity math

The economic logic

• The Mandant pays for outcomes - A Mandant pays the same fee whether the prep was manual or automated. The agent shifts cost without shifting price.
• Throughput becomes the moat - In a Fachkräftemangel market, the Kanzlei that can take on more Mandanten without proportional hiring wins the marginal client
• Beratung is the high-margin work - Hour-for-hour, Mandantenberatung bills at 2-3x the rate of Buchhaltung. Freeing senior time for Beratung shifts the mix.
• Retention compounds - Mandanten who get next-day responses and 3-day BWAs do not switch Kanzlei. The retention curve flattens.
• Succession planning gets easier - A Kanzlei running 50 Mandanten with 10 people on modern infrastructure is salable. A 30-Mandant Kanzlei running on paper is not.

How Superkind Fits

Superkind builds custom AI agents that sit on top of the systems Kanzleien already use - DATEV, Addison, Stotax, Agenda, Wolters Kluwer - without forcing a switch. The deployment model is process-first: we map your Belegerfassung, your Mandantenkommunikation patterns, and your USt-Voranmeldung workflow before we touch a line of code. The agent is built around your Kanzlei’s reality, not a generic template.

What sits in the Superkind agent for a Kanzlei

What Superkind brings to a Steuerkanzlei

• Process-first deployment - We map Belegerfassung, Mandantenkommunikation, and USt-Voranmeldung before we build, so the agent fits your Kanzlei’s reality
• DATEV-native architecture - The agent works through Belegtransfer, Unternehmen online, and the DATEV API. No replacement, no parallel system.
• BStBK-ready compliance - EU hosting, AV-Vertrag with explicit Berufsgeheimnis clause, no-training guarantees, full audit logs
• Mandanten-specific learning - Each Mandant’s Lieferanten, Kontenrahmen, and posting patterns become rules the agent applies automatically
• Verfahrensdokumentation included - The compliance paperwork ships with the agent, not as an afterthought
• Human-in-the-loop by design - Confidence thresholds are configurable; nothing auto-books below the line you set
• 60-day pilot scope - One Mandant, one use case, written success criteria - go or no-go after 60 days
• Long-term partnership model - We stay involved beyond the pilot; the agent evolves as the Kanzlei evolves

Decision Framework

Not every Kanzlei is ready, and not every Kanzlei needs a custom agent. The framework below helps locate which one you are.

Start with DATEV Copilot if

• Solo or 2-person Kanzlei - Throughput is not yet the constraint; the gain from a horizontal assistant is enough
• Under 200 Belege per month - The math does not yet justify a custom build
• No succession or growth pressure - Status quo is acceptable for the next 12 months
• You want to test AI broadly before committing - Copilot is the lowest-friction starting point

Move to a custom agent if

• 5+ employees - Coordination overhead and Mandanten count justify the build
• 500+ Belege per month per Kanzlei - Volume produces the ROI mathematics
• Hiring is blocked - You have tried to hire and the candidates are not arriving
• BWA delivery is slow - Mandanten complain about timing; you have lost or are about to lose Mandanten over it
• Senior team is doing data entry - Steuerberater are spending evenings on what Steuerfachangestellte should be doing
• Succession or sale on the 3-year horizon - Modernised Kanzleien sell for higher multiples

Wait if

• Mandantenstammdaten are a mess - Clean them first; the agent will not save you
• No DATEV access for the IT partner - Sort the technical access before the project
• Resistance from the Steuerberater - The Berufsträger has to sponsor the project; without that, it stalls
• No budget for AV-Vertrag review - The legal foundation is non-negotiable; budget the lawyer hours

Frequently Asked Questions

Related Articles

• AI in Bookkeeping: How the Mittelstand Automates Receipts, DATEV Booking, and the End of Buchhaltungs-Mondays
• E-Rechnung 2025 with AI: How the German Mittelstand Meets the Mandate and Makes More of It
• AI as a Compliance Assistant: How the Mittelstand Automates Audit Trails, Policies, and Reporting
• Which LLM Should the Mittelstand Choose? GPT, Claude, Gemini and Mistral Compared
• AI Agent ROI: The KPI Framework That Convinces CFOs in 90 Days
• Intelligent Document Processing: How the Mittelstand Replaces OCR With AI Agents

Henri Jung

Co-founder of Superkind, where he helps SMEs and enterprises deploy custom AI agents that actually fit how their teams work. Henri is passionate about closing the gap between what AI can do and the value it creates in real companies. He believes the Mittelstand has everything it needs to lead in AI - it just needs the right approach.