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.
On this page
- Why crypto practice is unusually AI-friendly
- MiCA files and whitepapers: the gap-check prompt
- Crypto tax: where the model must not do the arithmetic
- On-chain evidence: an agent with a tool, not a chatbot with an opinion
- Regulatory monitoring that actually runs
- Confidentiality: wallets, keys and consumer chatbots
- Where this leaves a crypto practice
A MiCA authorisation file runs to hundreds of pages, a token whitepaper has a fixed list of required contents, a crypto-tax ledger can have 40,000 lines, and the evidence in a dispute sits on a public blockchain that any tool with an API key can read. That combination makes crypto practice unusually suited to AI, provided the model reads and drafts while a lawyer characterises, computes and verifies.
Why crypto practice is unusually AI-friendly
Three features of the work line up with what models do well as of September 2026. The regulation is new and structured: MiCA (Regulation (EU) 2023/1114) sets out authorisation requirements for crypto-asset service providers and mandatory whitepaper contents in annexes that read like checklists, and the transitional periods for existing providers ended by 1 July 2026 at the latest, so every operator is now inside the regime. The data is public and machine-readable: transactions, contract code and token flows sit on chains that expose APIs. And the documents are long: 1M-token context windows (Claude Opus 5 and Fable 5.1, GPT-6 Astra, Gemini 3.1 Pro) hold an entire authorisation file at once.
Dr. Niklas Schmidt, who teaches the AI Lab for Lawyers course, wrote a German-language book on crypto assets and blockchains, co-edited an English-language book on crypto taxation covering more than 40 countries, sits on the advisory board of the Digital Asset Association Austria and founded the WT Crypto Academy, which has trained more than 1,000 lawyers. The workflows below are the ones that survive contact with that kind of practice.
| Workflow | What the model does | What stays with the lawyer | Tool type |
|---|---|---|---|
| CASP authorisation file | Gap-checks the draft against the requirement list; drafts policy sections from your templates | Substantive adequacy; regulator engagement | Long-context model with your templates in a Project |
| Whitepaper review | Maps required contents to paragraphs; flags missing or inconsistent disclosures | Token characterisation; exemptions; liability language | Long-context model |
| Crypto tax | Drafts the memo around figures; explains the method to the client | Cost basis, event characterisation, the numbers | Tax computation tool plus adviser |
| On-chain evidence | Queries the chain via a tool; builds timeline; returns hashes | Interpretation; chain of custody; expert evidence | Agent with block-explorer tool (Claude Code, Cowork) |
| Regulatory monitoring | Diffs regulator pages daily; summarises changes; drafts client alert | What matters and what to tell whom | Scheduled agent (Claude for Legal, Harvey, Legora, n8n) |
MiCA files and whitepapers: the gap-check prompt
An authorisation application is a set of documents mapped to a set of requirements: programme of operations, governance, ICT and security, safeguarding of client assets, complaints, conflicts, AML, prudential requirements. The same is true of a whitepaper. The model's job is the mapping.
You are assisting a crypto-regulatory team reviewing a draft document
against a list of mandatory contents. Work only from the two uploads:
(1) the draft; (2) the requirement list.
For every item in the requirement list, return a row with:
- requirement (verbatim from the list);
- location in the draft (section and page) or "NOT FOUND";
- a one-sentence assessment: satisfied / partially addressed / missing;
- the exact quotation from the draft you relied on.
Then list:
- internal inconsistencies (for example, token supply figures or
governance descriptions that differ between sections);
- statements that read as promises of value, return or liquidity;
- any reference to a law, regulation or guideline, so a lawyer can
verify it.
Do not assess whether the token is in scope of the regulation or
whether an exemption applies. Do not cite law from memory.
Two habits keep this honest. First, the requirement list is pasted in from the regulation, not left to the model's recollection: models cite law from memory with confidence and with errors, and in the HAQQ benchmark (5 June 2026, 3,000 graded answers across 51 practice areas) 24% of all answers cited or misapplied law that did not support the claim. Second, the output is a review aid, not a filing. The characterisation questions that drive a MiCA analysis are researched in primary sources, and anything that reads like a regulator's position is checked on the regulator's site. The same method scales to due diligence on a target's compliance file.
Crypto tax: where the model must not do the arithmetic
Tax is the part of crypto practice where AI is most tempting and most dangerous. A client sends exports from three exchanges and two wallets and asks for the gain. A general model will produce a number. It will not produce a defensible working, because the method (which cost-basis rule, how staking rewards and airdrops are characterised, what a token-to-token swap triggers, how a lost-key event is treated) differs by jurisdiction and changes with practice, and because a model doing arithmetic across thousands of rows is a model guessing at scale.
The tax profession's own numbers show where this is going. The Blue J/CPA.com survey of more than 1,000 US tax practitioners (8 June 2026) found weekly AI use for tax research at 60%, up from 33% a year earlier. Tax authorities are ahead: Austria's Amt für Betrugsbekämpfung reported €154m recovered in 2025 and runs a "Predictive Analytics Competence Center". From 1 January 2026 the EU's DAC8 rules require crypto-asset service providers to report client transactions to tax authorities, so the authority will have the ledger too. A crypto-tax workflow that holds up has three parts: a purpose-built computation tool that exports the full ledger with its method; a tax adviser who checks the characterisation and signs the return; and a general model used only to draft the explanatory memo around figures that are already fixed. The tax lawyers page covers the research tools; Germany's first test of whether software can call itself a tax adviser (Steuerberaterkammer Berlin v Accountable, LG Berlin II, 61 O 90/26, filed May 2026) is a reminder that the label matters as much as the maths.
On-chain evidence: an agent with a tool, not a chatbot with an opinion
Asking a chatbot "what did this wallet do" produces fiction. Giving an agent a tool that queries a block explorer or a node produces data. That distinction is the whole method for on-chain evidence in a fraud, insolvency or asset-tracing matter.
- Preserve first. Export the raw transaction data for each address from a primary source (node or explorer API) with timestamps, and hash the export. This is the evidence; everything after is analysis.
- Give the agent a read-only tool. In Claude Code or Cowork, connect an explorer API through a Model Context Protocol server or a small script, restricted to read calls. No wallet-signing capability of any kind lives in the same environment.
- Ask for a timeline with hashes. Every entry in the output carries the transaction hash, block number and the addresses involved, so a paralegal can verify each line against the explorer.
- Flag patterns, do not conclude. Deposits to known exchange addresses, interaction with mixers, bridge transfers, rapid split-and-merge sequences: the agent flags; a forensic expert interprets.
- Record the method. Which tool, which API, which queries, when. Expect to explain it. Courts treat AI output as a tool's output, and the proposed US Federal Rule of Evidence 707, which would apply expert-reliability standards to machine-generated evidence, was published for comment in June 2025 and reportedly sent back for further study in June 2026; either way, the reliability question will be asked.
The scripting side of this is easier than it sounds and is covered in Claude Code for lawyers. The security side is not optional: an agent with an API key and network access is exactly the kind of system that the agent security guide exists for, and a July 2026 wire-level analysis showed the Grok Build CLI uploading entire repositories including an unredacted .env file that users had told it to ignore. Keys go in a vault, never in a prompt.
Regulatory monitoring that actually runs
Crypto regulation moves weekly: ESMA and EBA guidance, national-authority Q&As, US agency statements, enforcement actions. The most useful automation in a crypto practice is a scheduled agent that fetches a defined list of pages every morning, diffs them against yesterday, summarises what changed with a link, and drafts a two-paragraph client note for a lawyer to approve.
Three ways to build it as of September 2026. Anthropic's open-source Claude for Legal repository ships a regulatory-legal plugin (feed watcher, policy diff, comment deadlines) and a headless "reg-monitor" agent cookbook that runs on Managed Agents at $0.08 per session-hour plus tokens; the plugin's redrafts are framed as proposals, not edits. Harvey's Agent Builder and Legora's Monitors do the same inside those platforms for firms already paying for them. A self-built n8n workflow with an HTTP node, a diff step and a model call costs almost nothing: a three-lawyer German firm runs its whole intake on a self-hosted n8n server for €20–30 a month. The design pattern and the traps are in AI regulatory monitoring. Module 4 of the AI Lab for Lawyers course builds exactly this monitor live in n8n, with an approval step before anything reaches a client.
Confidentiality: wallets, keys and consumer chatbots
Crypto clients are often technically fluent and often use consumer AI tools themselves. Two rules for the practice, and one conversation with the client.
Nothing identifying goes into a consumer tier. Wallet addresses linked to a named client, seed phrases (never, anywhere), KYC files, exchange account details, and the fact of a regulatory enquiry are all client confidential. ChatGPT's consumer plans train on conversations unless the user opts out; consumer Gemini keeps activity for 18 months by default with human review of a subset; consumer Claude asks the user to choose, with five-year retention if training is enabled. Business, Team, Enterprise and API tiers do not train by default. The comparison is in the confidentiality checklist.
The client's own chatbot use is not privileged. In United States v. Heppner (S.D.N.Y., bench ruling 10 February 2026, opinion 17 February 2026), Judge Rakoff held that a defendant's research memos generated with consumer Claude were neither privileged nor work product: "Claude is not an attorney" and the terms allow logging and disclosure. Privilege might attach where counsel directs the AI use. That is a conversation to have at the first meeting with any founder or investor who has been "researching" their position with a chatbot; the detail is in privilege and AI. Scanned exchange statements and screenshots go through an OCR pass before upload, as described in extracting data from PDFs, and for figures Claude is the model most likely to say "illegible" rather than invent a number.
Where this leaves a crypto practice
- Adopt the gap-check prompt for authorisation files and whitepapers, with the requirement list pasted from the regulation every time.
- Fix the tax division of labour in writing: tool computes, adviser signs, model drafts the narrative.
- Stand up one read-only on-chain agent in a sandbox with keys in a vault, and document its method once.
- Run a regulatory monitor with a human approval step; review its output for a month before trusting the summaries.
- Add two lines to the engagement letter: what the firm does with AI, and what the client should not do with it.
Crypto lawyers spent a decade explaining a new asset class to sceptical regulators. Explaining a new tool class to sceptical partners is a smaller job, and the workflows above are the evidence that it works.
Frequently asked questions
Can AI review a MiCA whitepaper?
Yes, as a first-pass gap check. Give a long-context model the draft whitepaper and the required content headings from MiCA's annexes, and ask it to map each requirement to the paragraph that satisfies it and to list what is missing. The legal characterisation of the token, and whether an exemption applies, remains your judgment and should be researched in primary sources.
Is AI reliable for crypto tax calculations?
Not on its own. Cost-basis methods, staking and airdrop treatment and token-to-token swaps vary by country and year, and a general model will produce a plausible number without showing a defensible working. Use a purpose-built tax tool that exports the ledger, check it, and have a tax adviser sign off. A model is fine for drafting the explanatory memo around the figures.
Can an AI agent analyse blockchain transactions as evidence?
It can, if it queries the chain through a tool rather than from memory. An agent connected to a block explorer or node can pull transactions for an address, build a timeline and flag mixers or exchange deposits, returning hashes you can verify independently. Preserve the raw data, record the query method, and expect to explain it to a court or tribunal.
Which AI tools do crypto lawyers use in 2026?
Mostly general models in business tiers: Claude and ChatGPT for drafting and long-document review, Claude Code or Cowork for agents that call APIs, Gemini Notebook for closed-universe case files, and Perplexity for fast sourced research. Regulatory-monitoring agents come from Claude for Legal's regulatory plugin, Harvey or Legora Monitors, or a self-built n8n workflow.
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
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.
10 min read
Use Cases
AI Regulatory Monitoring: Agents That Track Law Changes for You
AI regulatory monitoring works as a scheduled fetch, diff, classify and route workflow on official sources. How to build a weekly digest and cut false alarms.
10 min read
Tools
Claude Code for Lawyers: Should Non-Programmers Use a Coding Agent?
Claude Code is a terminal coding agent. When lawyers should use it for bulk document work and small tools, two worked examples with guardrails, safer routes.
9 min read
Playbooks
Extract Data from PDFs and Scans with AI: Contracts, Invoices and Court Files
How lawyers extract data from PDFs and scans with AI: OCR first, a fixed schema, a full lease-extraction prompt, batch runs in Cowork or n8n, and sampling.
11 min read
Guides
Agent Security for Lawyers: Prompt Injection, Data Leaks and Guardrails
AI agent security for law firms: prompt injection via hostile documents, the Grok Build and Cowork incidents, and the allowlists and gates that contain them.
11 min read
Glossary
Model Context Protocol (MCP) — Definition for Lawyers
Model Context Protocol (MCP) is the open standard that lets Claude, ChatGPT and Copilot read your DMS or docket feed through one connector. What a server exposes and what to check.
4 min read