How to Draft Legal Responses with AI Without Hallucinating Case Law
Dutch lawyers are increasingly using AI for legal writing. But there’s a problem: AI tools are generating citations to cases that don’t exist, or that say something completely different from what’s claimed.
This damages your professional reputation when opposing counsel checks your references. More importantly, it undermines your argument when the rechtbank discovers your key authority doesn’t exist.
The core problem: AI is a reasoning engine, not a legal research tool. When you ask it “what cases support my position,” it generates plausible-sounding legal text based on patterns. It doesn’t retrieve actual jurisprudence. It predicts what legal language should look like.
The solution: Never give AI discretion to invent the law. You do the research. You control what sources AI can reference. AI applies those sources to your facts.
The Concept: RAG-Inspired Workflow
In AI systems, RAG (Retrieval-Augmented Generation) refers to a technical architecture where the system automatically retrieves relevant information from a knowledge base using vector embeddings and semantic search, then uses that retrieved context to generate responses.
What we’re doing here is simpler but follows the same principle: You manually retrieve the legal sources, then provide them to AI as context. No vector databases or automated retrieval—just you doing the research and explicitly controlling what sources AI can reference.
Think of it as “manual RAG” or “lawyer-controlled context injection.”
The workflow:
- You research and identify the relevant cases and statutory provisions
- You provide those sources to AI along with your facts
- You tell AI exactly what each source should be cited for
- AI applies those sources to your facts and structures the argument
- You verify and edit
AI never decides what the law is. It only explains how the law you provided applies to your case.
This eliminates hallucination because AI can’t cite jurisprudence you didn’t give it.
The Process
Step 1: Do Your Legal Research
Read the opposing party’s dagvaarding or conclusie. Identify the legal issues. Then:
- Find the relevant Hoge Raad decisions and lower court jurisprudence
- Pull the applicable BW, Rv, or other statutory provisions
- Check rechtspraak.nl for recent developments
- Verify cases are still good law (not overturned or distinguished)
- Download everything as PDFs
This is traditional legal research using the tools you already know: rechtspraak.nl, Navigator, Kluwer, etc.
Step 2: Define Your Argument
Before involving AI, figure out what you’re arguing:
- What’s your counter-argument to their legal position?
- How do the facts favor your interpretation?
- What’s your theory of defense?
Write this out in 2-3 paragraphs for clarity. This requires legal judgment—something AI can’t provide.
Step 3: Structure Your Prompt
Most lawyers fail here. They write: “Draft a response arguing we should win.”
That’s not a prompt. That’s an invitation to hallucinate.
The structure that works:
Paragraph 1 - Role and Format: Tell AI it’s drafting the legal argument section of your conclusie. Specify tone, length, and formatting (numbered paragraphs, formal Dutch legal style).
Paragraph 2 - The Facts: Provide the relevant facts. Be specific. AI needs facts to apply law to.
Paragraph 3 - The Law (CRITICAL): State the legal standard explicitly. List every case (with ECLI number) and statutory provision you want cited, and specify what each source should be used for.
Then add:
“Do NOT cite any cases beyond what I have provided. Do NOT search for additional authorities. Do NOT suggest alternative legal frameworks.”
This is the core of the RAG-inspired approach—constraining AI to only the sources you control.
Paragraph 4 - The Argument: Explain your legal reasoning from Step 2. Tell AI what you want articulated and how the provided sources support it.
Step 4: Upload Context
Give AI:
- Your structured prompt
- Their dagvaarding or conclusie (so AI understands what it’s responding to)
- PDFs of your cases (from rechtspraak.nl)
- The relevant statutory provisions
Most AI tools (Claude, ChatGPT, specialized legal AI) can handle PDF uploads and long context.
Step 5: Generate and Edit
Run the prompt. You’ll get a draft that:
- Cites only the sources you provided
- Applies those sources to your facts
- Follows your legal reasoning
- Uses appropriate Dutch legal writing style
Edit for emphasis and tone. If major revision is needed, your prompt wasn’t specific enough—revise and regenerate.
Step 6: Verify Citations
Check that:
- ECLI numbers are correct
- Case names and courts are accurate
- The propositions match what the jurisprudence actually says
- Article numbers are correct
This should be quick since you already researched these sources.
The Prompt
See the code block at the top of this post for the full template.
Concrete Example
Scenario: You represent a buyer in a commercial dispute. Seller claims breach because buyer refused final payment of €180,000 after machinery was delivered three weeks late. Buyer claims the delay constitutes breach and entitles withholding payment.
Your research: You’ve found relevant Hoge Raad jurisprudence on delivery dates and whether they constitute fatale termijnen, lower court decisions on accepting delayed performance, and the applicable BW provisions on good faith and contract performance.
You structure the prompt:
You are drafting the legal argument section of a conclusie van antwoord on behalf of buyer (defendant). Seller (plaintiff) claims buyer breached the purchase agreement by refusing to pay €180,000 for custom machinery, arguing the three-week delivery delay entitled buyer to withhold payment.
TONE AND FORMAT:
- Professional Dutch litigation writing style
- Focus on contractual interpretation and application to facts
- Target length: 800-1000 words
- Use numbered paragraphs as customary in Dutch proceedings
THE FACTS:
Parties entered into a purchase agreement on 1 March 2024 for custom machinery (€600,000 total, €180,000 final payment). The contract stated "Levering uiterlijk 1 juni 2024" but contained no language making time of the essence. Seller notified buyer on 15 May that delivery would be delayed due to supplier bankruptcy. Seller delivered on 22 June. Buyer accepted delivery, installed the machinery, and began using it in production. Two months later, buyer refused final payment, claiming the delay caused substantial losses.
THE LAW (CRITICAL - YOU MUST FOLLOW THIS EXACTLY):
The applicable legal standard is that specifying a delivery date does not automatically make time of the essence (fatale termijn) unless the contract explicitly provides this or circumstances clearly indicate this was the parties' intention.
You must cite ONLY the following sources I have provided:
- [Hoge Raad decision on delivery dates, ECLI:NL:HR:YYYY:XXXX]: Use this for the principle that delivery dates don't automatically constitute fatale termijnen
- [Appellate court decision on acceptance of performance, ECLI:NL:GH:YYYY:XXXX]: Use this for the principle that accepting delayed performance precludes later claims the delay was material breach
- Art. 6:265 BW: Use this for the good faith principle that parties must act consistently with their prior conduct
Do NOT cite any cases or authorities beyond what I have provided above. Do NOT search for additional legal sources. Do NOT suggest that other legal frameworks might apply.
THE ARGUMENT:
Our position is: The contract did not make the delivery date a fatale termijn, so the delay cannot constitute material breach justifying non-payment. Moreover, buyer's conduct after delivery—accepting the goods, installing them, and using them in production—constitutes acceptance of the delayed performance. Under art. 6:265 BW, buyer cannot claim the delay was material after acting inconsistent with that position. Buyer's proper remedy, if any, is a separate claim for proven damages, not withholding the purchase price for accepted and functioning goods.
Draft the legal argument section showing why seller's claim should succeed. Apply the legal standard I provided to our facts, citing the specific cases and authorities I listed above for the propositions I indicated.What you get:
AI generates a properly structured Dutch legal argument that:
- Uses numbered paragraphs following Dutch conventions
- Cites only the cases and provisions you provided (with correct ECLI numbers)
- Applies those authorities to your specific facts
- Maintains formal Dutch legal writing tone
- Structures the reasoning logically from general principle to specific application
You then verify the citations match what the cases actually say, make any edits for emphasis, and you have a solid draft argument section ready for your conclusie.
Why This Works
AI is genuinely good at:
- Applying established legal principles to facts
- Structuring legal arguments according to Dutch conventions
- Articulating reasoning clearly
- Maintaining consistent formal tone
AI cannot:
- Conduct legal research
- Determine what the law is
- Verify case citations or ECLI numbers
- Exercise legal judgment about strategy
This RAG-inspired approach lets you use AI for what it’s good at while you handle what requires professional judgment: research, verification, and strategy.
The key principle: Never give AI discretion on the law. You tell it what the law is. AI shows you how that law applies.
When to Use This
Use this methodology for:
- Conclusies van antwoord in bodemprocedures
- Kort geding proceedings where speed and accuracy both matter
- Written submissions where citation accuracy is critical
- Any litigation writing where professional reputation depends on accuracy
Don’t use this for:
- Exploratory research where you’re still figuring out the law
- Situations where you’re uncertain what legal framework applies
- Initial case assessment before you’ve identified the relevant jurisprudence
In those cases, use traditional research tools (rechtspraak.nl, your legal database) first. Once you know the law, use this approach to apply it.
We’re building Mino for lawyers who want to use AI without compromising professional standards. Our tools are designed around this principle: you control the inputs, AI assists with the reasoning. If you want to work more efficiently while maintaining the quality your clients and the rechtbank expect, join the founding members.
Don't paste confidential client information into AI tools without proper safeguards
Dutch bar rules and GDPR require you to protect client data.
Practical workaround: Anonymize before you prompt.
- Replace names with generic labels ("Party A," "the manufacturer," "the employee")
- Remove identifying details (specific amounts, dates, locations)
- Keep only what's needed for the legal reasoning
For these reasoning tasks, the logic works the same whether you're analyzing "Holding B.V.'s distribution agreement" or "a distribution agreement." You're structuring the thinking, not processing the raw case file.
Or consider on-premise deployment where your data never leaves your infrastructure. We wrote about practical options for running AI locally - including cost analysis and which models actually work for legal work.