Practice Areas
AI for Tax Lawyers and Tax Advisers: Research, Rulings and Returns
AI for tax lawyers works where sources are supplied and numbers are checked. What tax authorities, Big Four and Blue J do, plus a worked withholding example.
On this page
- The tax authorities got there first
- Where AI saves a tax team time, and where it does not
- A worked example: dividend from an Austrian subsidiary to a German parent
- Numbers, thresholds and the tax-specific hallucination problem
- Transfer pricing, rulings and the agents that actually help
- Who is allowed to give tax advice, and what the tools are
- A setup that works for a tax team
AI is useful to tax lawyers in exactly the places where the source text is in front of the model and the numbers are checked afterwards: analysing a treaty article you pasted in, structuring a ruling request, drafting transfer-pricing narratives, monitoring guidance. It is dangerous where it is asked to remember a rate, a threshold or a deadline. The tax authorities, meanwhile, have already automated the other side of the table.
The tax authorities got there first
Austria's Amt für Betrugsbekämpfung reported in February 2026 that it had brought in more than €154m for the Republic in 2025, an increase of over €47m, secured 92.5 terabytes of evidence and runs a dedicated "Predictive Analytics Competence Center" (BMF press release, 10 February 2026). North Rhine-Westphalia's tax administration has reportedly used a self-developed AI tool since May 2025, the first German Land to do so, according to a secondary source that could not be independently verified (ADS Steuer Forum).
The practical consequence is real. Risk scoring of returns, matching of automatically exchanged cross-border information and pattern detection across a client's group are now the baseline against which your positions will be read. Documentation that explains the commercial rationale in plain terms, with the numbers reconciled, has become more valuable, and that is precisely the kind of drafting a well-supervised model does quickly.
The Big Four are building the same machinery for advisers. EY launched "EY.ai Tax Labs" and an "EY.ai Tax Agent Factory" for multistage tax processes on 25 November 2025 (EY newsroom); secondary reports of "150 agents for 80,000 tax professionals" are unverified. PwC licenses its Harvey-built "Tax AI Assistants" to smaller UK accountancy firms through The Corporate Finance Network (April 2026) (Harvey blog). The Big Four as software vendor to the mid-market is a new competitive fact for tax boutiques and law-firm tax teams alike.
Where AI saves a tax team time, and where it does not
The Blue J and CPA.com survey of more than 1,000 US tax practitioners (8 June 2026) found that 60% now use AI for tax research at least weekly, up from 33% a year earlier; 84% say it saves time; and 69% expect to move to value-based, hybrid or fixed fees (CPA.com). Those numbers describe research. The gains are not evenly distributed across a tax practice.
| Task | Gain as of September 2026 | Condition |
|---|---|---|
| First-pass issue spotting on a fact pattern | High | Give the model the facts and the relevant provisions; ask for questions, not conclusions |
| Treaty article analysis (OECD Model commentary, protocol, MLI position) | High | Paste the article, the protocol and the commentary paragraphs; ask for a structured reading |
| Transfer-pricing local file narratives (functions, assets, risks) | High | Feed the group's master file, intercompany agreements and interviews; draft, then have the economist check the benchmarks |
| Ruling-request drafts (EAS in Austria, verbindliche Auskunft in Germany, APAs) | Medium–high | The model structures; the adviser supplies the legal analysis and the precise question |
| Monitoring guidance, ministry decrees and case law | High with an agent | See the regulatory-monitoring agent pattern; verify every item before acting |
| Client explainers and cover letters | High | Tone control and plain language; no numbers the model was not given |
| Retrieving a rate, threshold or filing deadline from memory | None; net negative | Never |
| Multi-jurisdictional questions | Low–medium | Every tool in the Vals research benchmark lost 11–14 points on multi-jurisdiction questions (October 2025) |
| Numerical computations across a return | Low without code | Use a spreadsheet or a code-execution tool; check the model's arithmetic line by line |
The Vals finding deserves emphasis for cross-border practitioners: in its October 2025 legal research study, every system tested, including ChatGPT with web search at 80% overall, dropped roughly 11–14 points on questions spanning more than one jurisdiction (LawSites summary). Cross-border tax is nothing but multi-jurisdictional questions.
A worked example: dividend from an Austrian subsidiary to a German parent
Take the most routine cross-border question a Vienna tax team gets: an Austrian GmbH is about to distribute a dividend to its German corporate parent. What is withheld, under which instrument, and what procedure applies?
Ask a frontier model cold and it will produce a fluent answer that typically contains at least one of four errors. It quotes the Austrian corporate income tax rate from before the two-step reduction that took the rate to 24% in 2023 and 23% from 2024, because most of its training text predates the change. It conflates the Austria–Germany treaty rate for qualifying holdings with the exemption under the EU Parent-Subsidiary Directive as implemented in the Austrian Income Tax Act, and forgets that the domestic exemption has its own minimum-holding and holding-period conditions. It skips the procedural question entirely, which is where the money actually sits: relief at source against the required documentation versus a refund claim afterwards. And it invents or overstates the anti-abuse test.
None of this is a reason to avoid the model. It is a reason to run it the right way round. Here is the prompt a tax associate should actually use, with the sources attached as files:
Role: You are assisting an Austrian tax adviser. Work ONLY from the attached
documents. Do not use any rate, threshold, holding period or procedural rule
from memory. If a needed fact is not in the documents, write "NOT IN SOURCES"
and list what is missing.
Attached:
A. Extract of the Austrian Income Tax Act (EStG) provisions on capital
yields tax and the exemption for EU parent companies (current text)
B. Austria–Germany double tax treaty, dividend article and protocol
C. Ministry of Finance guidance on relief at source and refund procedure
D. Fact sheet: the shareholding percentage, acquisition date, legal form of
the parent, tax residence certificate status, intercompany agreements
Task:
1. Identify every condition in A and B that determines the withholding
outcome. Quote each condition verbatim with its paragraph reference.
2. Apply each condition to the facts in D. Mark each as MET / NOT MET /
UNCLEAR, with a one-line reason.
3. State which instrument gives the lowest withholding and why the other
does not apply, citing A or B for each step.
4. Describe the procedure from C: relief at source or refund, documents
required, deadlines. Quote C for each deadline.
5. List anti-abuse or substance conditions in A, B or C and flag which
facts in D are relevant. Do not speculate beyond the sources.
6. Produce a table: Condition | Source paragraph | Fact relied on | Status.
Output: structured memo, max 900 words, then the table. No conclusions that
are not traceable to a quoted paragraph.
The output is then checked against the sources, not against the model. The check takes ten to fifteen minutes and looks like this:
| What the model asserted | Where to verify | Typical failure |
|---|---|---|
| The withholding rate under domestic law | Statute extract, current consolidated text | Old rate from pre-2023 text |
| The exemption conditions (minimum holding, holding period, legal form) | Statute plus the directive annex | Holding period omitted or wrong |
| The treaty rate and its qualifying-holding threshold | Treaty article and protocol | Treaty and directive thresholds merged |
| Relief-at-source documentation and the refund route | Ministry guidance | Procedure invented or deadline missing |
| Substance and anti-abuse conditions | Statute, guidance, current administrative practice | Overstated or asserted without a source |
| Every date and every percentage | The paragraph the model quoted | Quote is accurate but the paragraph says something different |
This is the three-check verification that applies to all legal AI, adapted to a field where the errors are numbers rather than case names.
Numbers, thresholds and the tax-specific hallucination problem
The hallucination debate in law has centred on fabricated cases. In tax the more common failure is a real provision applied wrongly, or a correct rule with the wrong number attached. The HAQQ benchmark of 3,000 graded answers across 51 practice areas (5 June 2026) found that 24% of all answers "cited or misapplied law that didn't support claims", even though the best model's outright citation fabrication rate was as low as 3% (HAQQ). Misapplication is what tax advisers should fear, and it is harder to spot than an invented case.
Two further findings matter for tax work specifically. First, knowledge cutoffs: as of September 2026, Claude Fable 5.1 stops at June 2026, Claude Opus 5 at May 2026 and GPT-6 Astra at 30 April 2026. Anything enacted, decreed or decided after those dates is invisible unless you paste it in. Tax law changes every year; budget laws, ministry decrees and court rulings arrive continuously. Second, sycophancy: Stanford's 2026 AI Index reports that when a false premise is framed as the user's own belief, model accuracy collapses, in one test from 98.2% to 64.4% (Stanford HAI). "My client's structure qualifies for the participation exemption, please confirm" is exactly that framing. Ask neutrally, and ask the model to argue the authority's side too.
The adviser's liability does not shrink because a model drafted the memo. Deloitte Australia partially refunded an AU$440,000 government report in October 2025 after it was found to contain a fabricated court decision and non-existent academic papers (The Register). That was an advisory report, not a court filing; a tax opinion a client relies on for a transaction sits in the same category.
Transfer pricing, rulings and the agents that actually help
Transfer-pricing documentation is the tax task with the best ratio of drafting volume to judgment. The functional analysis, the narrative of the intercompany agreements, the consistency check between master file and local file across fifteen entities: this is document analysis and structured drafting at scale, and it is where an agent with a 1M-token context (roughly 555,000 words on Claude's current tokeniser) starts to pay for itself. Feed it the agreements, the interviews and last year's file; ask for inconsistencies, missing risk allocations and every place where the narrative and the numbers disagree. The benchmarking itself stays with the economist.
Ruling requests reward the same discipline. Whether it is an Austrian EAS enquiry, a German verbindliche Auskunft or an advance pricing agreement, the request has a fixed anatomy: facts, question, applicant's legal view, reasons. A model drafts that structure well from a good briefing note and will produce three alternative framings of the question. The legal view is yours. Keep a Claude Project per group client with the structure chart, standing facts and house style, so the model stops asking for the same context.
Monitoring is the third place. A weekly agent that reads the ministry's decree list, the fiscal court's new decisions and EU Council tax announcements, and returns a one-page digest with links, is a workflow that firms build in an afternoon on n8n or Zapier and that pays back the first time it catches a decree affecting a live matter. Module 4 of the AI Lab for Lawyers course builds exactly this kind of monitoring automation live, without code. The rule for every item the digest returns is the same as for research: open the source before you tell a client.
Who is allowed to give tax advice, and what the tools are
In May 2026 the Steuerberaterkammer Berlin sued the start-up Accountable at LG Berlin II (Az. 61 O 90/26) over its "KI-Steuerberater" label, the first German court case on whether AI tax software amounts to unauthorised Hilfeleistung in Steuersachen and whether software can carry a protected professional title (Rose & Partner). The outcome will draw the German line between assistive software used by a licensed adviser and software that advises the public directly; the same line-drawing shapes unauthorised-practice debates in the US. For a licensed adviser using AI internally the question does not arise, provided the adviser signs off and the client's data is handled in line with professional rules. Germany's BRAK guidance (December 2024) treats tools that use client data for training as a breach of § 43a Abs. 2 BRAO; Austria's ÖRAK rules require EU servers, disclosure to clients, the lawyer's final responsibility and no training on client data. The Austrian and German bar guidance page covers both.
As for tools: Blue J raised a $122m Series D in November 2025, reports 3,500-plus customers and 3m-plus queries in 2025, prices at roughly $1,500 per seat per year and has added IBFD cross-border content; its CEO's pitch is "15 hours to 15 seconds" for a research memo (VentureBeat). Thomson Reuters sells CoCounsel Tax on Checkpoint content; Bloomberg Tax has an AI assistant. For European practitioners the honest position, as of September 2026, is that the dedicated tools are US-centric, and that a frontier model with the right sources attached, run inside a confidentiality-compliant setup, does most of the analytical work. None of them removes the adviser's duty to read the provision.
Dr. Niklas Schmidt, who heads the firm-wide tax team at Wolf Theiss in Vienna and is admitted as both Rechtsanwalt and Steuerberater (profile), works across tax, private clients and crypto assets; the crypto and private client pages on this site extend the tax methods here to those fields.
A setup that works for a tax team
- Choose a model tier that does not train on client data and, where the matter warrants it, one that supports zero data retention; check the current terms rather than the marketing page.
- Build one Project or workspace per group client with the structure chart, standing facts, house style and the current-year rate card you maintain yourself.
- Adopt the rule "no number from memory": every rate, threshold and deadline comes from an attached source and is quoted with a paragraph reference.
- Use the withholding-tax prompt pattern above for any technical question: sources attached, conditions tabulated, status marked, gaps flagged.
- Run a weekly monitoring agent for decrees, fiscal-court decisions and EU tax measures; verify before you circulate.
- Keep a research log for each AI-assisted memo showing which sources were supplied and who checked which numbers, which is also what your professional-liability insurer will ask for.
- Read the AI legal research and deep-research pages before choosing a tool; for tax, closed-universe work with your own sources beats open web search almost every time.
Frequently asked questions
Can I use ChatGPT or Claude for tax research?
Yes, for issue-framing, structuring a memo and analysing sources you paste in. No, for retrieving rates, thresholds or case citations from memory. Frontier models have knowledge cutoffs between April and June 2026 and still misapply law in roughly a quarter of legal answers on independent tests, so every number and citation must come from a source you supplied and then be checked against it.
Which AI tools are built specifically for tax research?
Blue J is the largest dedicated tax research platform, with 3,500-plus customers, IBFD and Tax Notes content and about $1,500 per seat per year. Thomson Reuters offers CoCounsel Tax on Checkpoint content, Bloomberg Tax has an AI assistant, and PwC licenses a Harvey-built Tax AI Assistant. All of them still require the adviser to open and read the primary source.
Is AI tax advice legal in Germany and Austria?
Assistive software used by a licensed adviser is uncontroversial. Software marketed directly as a tax adviser is contested. In May 2026 the Steuerberaterkammer Berlin sued the start-up Accountable at LG Berlin II over its 'KI-Steuerberater' label, the first German court test of whether AI tax software constitutes unauthorised assistance in tax matters.
How do tax authorities use AI?
Austria's Amt für Betrugsbekämpfung runs a Predictive Analytics Competence Center and reported €154m recovered in 2025. North Rhine-Westphalia's tax administration reportedly deployed a self-developed AI tool in May 2025. Advisers should assume risk scoring of returns and cross-border data matching are routine, and document positions accordingly.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Use Cases
AI Legal Research in 2026: What Works, What Hallucinates
AI legal research now beats a lawyer baseline on accuracy yet misapplies law in 24% of answers. Which tools are grounded, where errors hide, a routine that holds.
11 min read
Playbooks
How to Verify AI Legal Research: A Citation-Checking Protocol
Verify AI legal research with a five-check protocol: open every authority, match the quote, confirm holding and status, log it. Template, time budgets, tools.
11 min read
Regulation & Ethics
AI Rules for Austrian and German Lawyers: ÖRAK, BRAK and CCBE Guidance
ÖRAK's four AI rules and § 40 Abs 3 RL-BA bind Austrian lawyers; BRAK's § 43a BRAO line binds German ones; CCBE guides cover both. With a bilingual checklist.
11 min read
Practice Areas
AI for Crypto and Blockchain Lawyers: MiCA, Tax and On-Chain Evidence
AI crypto lawyers use models for MiCA authorisation files, whitepaper gap checks, crypto-tax memos and on-chain evidence, plus agents that watch regulators.
8 min read
Practice Areas
AI for Private Client and Estate Lawyers: Wills, Trusts and Family Offices
AI private client lawyers draft from precedent, build asset schedules and explain plans to families; formalities, cross-border succession and tax stay with you.
8 min read
Guides
Deep Research Tools for Lawyers: Claude, ChatGPT, Gemini and Perplexity Compared
Deep research AI for lawyers compared: what Claude, ChatGPT, Gemini and Perplexity research modes do, where they fail on law, and a verification table.
10 min read