Most lawyers who "tried AI" tried it the wrong way. They opened the public ChatGPT website, typed a generic question, got a generic answer, and concluded that AI "isn't ready for real work yet." That experience is so widespread it's caused entire firms to write AI off altogether.
The problem wasn't AI. The problem was five concrete mistakes that repeat for everyone working with AI alone for the first time. Here they are — and how to do it differently.
1 Public chat instead of the API
Your contracts don't belong in the free version.
The free versions of ChatGPT and Claude default to settings where your conversations are visible to the provider — and in some modes may be used to train the model. For an exploratory chat about the difference between the Czech Civil Code and the BGB, that's fine. For a client contract on a EUR 10 m acquisition it's a problem.
The fix: API access via a business account, where Anthropic (and OpenAI) contractually guarantee that your data is not used for training. The cost is comparable to a regular team subscription. The difference is the contractual guarantee — what would otherwise be an awkward audit conversation becomes a simple answer.
2 Too little context in the prompt
"Draft a service agreement" vs. an actual prompt.
The classic opening: "Draft a service agreement." Output: a generic boilerplate from the internet, jurisdiction-agnostic, statutory references probably wrong. Lawyer's reaction: "See, AI doesn't get it."
Reality: AI does get it; you just didn't say what you needed. Compare with this:
Draft a service agreement under § 2586 et seq. of the Czech Civil Code. Client: a Czech property developer. Contractor: a Czech construction company. Subject: renovation of a residential building in central Prague within a heritage zone. Price: CZK 15 m fixed, excl. VAT. Term: 9 months from site handover. Warranty: 60 months. Include: late-payment penalty 0.05 % per day capped at 25 %, judicial moderation right preserved, change-of-scope mechanism, materials price indexation given current volatility, damages capped at twice contract value. Tone: professional, not overly defensive — the parties have a good working relationship.
Same model, different output. A draft that's 80 % there. The difference: two minutes spent writing a contextual prompt.
3 No system prompts
Are you starting every conversation from scratch?
Most people using chat in legal practice open a new conversation and re-explain context from zero. "I'm a Czech lawyer, I work on…" — every single time.
A professional setup uses a system prompt: framing instructions that apply to every conversation. Something like:
You are a legal assistant specialised in Czech private law. Always cite specific statutory provisions (article and number). When citing case law, always verify the citation — if you're not sure the decision exists, say so. Reply in English in legal style, not conversational. For complex questions, structure: 1) legal framework, 2) application to the case, 3) risks, 4) recommendations. If you're not sure, say so explicitly.
You set this once. From that moment every conversation starts with this context. Output quality jumps an order of magnitude. Because we work in teams, you can run different system prompts for different use cases — contract review, tax queries, client correspondence, and so on.
4 Blind trust in hallucinations
AI invents things. Sometimes with full confidence.
AI models occasionally hallucinate — invent facts that sound plausible but don't exist. In a legal context this is critical: AI may cite Supreme Court ruling 25 Cdo 1234/2020, which doesn't exist but reads precisely. If you rely on it in a memorandum, you have a problem.
Rules to defend against this:
- Always verify exact citations. AI cites § 2586 of the Czech Civil Code? Open the statute and check that it says what AI claims. Thirty seconds. AI cites a precedent? Find it on nsoud.cz or in your case-law database.
- Demand explicit confidence. Add to your prompt: "For each statement, indicate confidence: high / medium / low. High only when there is a specific source."
- Critical passages get a second pass. A memorandum that goes to a client passes through a second round — a human reads the output and verifies. Like an associate's work.
AI is a powerful assistant — not an autonomous lawyer. Treat it like a very fast but very confident first-year associate.
5 Ad-hoc use instead of workflow integration
"I open chat sometimes" is a different discipline than "AI is in my workflow."
Most lawyers with some AI experience use it ad hoc. Monday morning, a question pops up, open chat, ask, get an answer, close chat. Maybe re-open it next week. This is AI as an occasional tool.
A professional setup looks different. AI is in every recurring workflow:
- Claude for Word during every drafting or review session
- Email assistant that drafts replies from your bullet-point notes
- Knowledge base of your templates and memoranda that AI uses as reference
- Automated first-pass review of every incoming contract against a custom checklist
- Tax research with statutory database integration
The difference: ad-hoc use saves minutes per week. Systematic integration saves hours per day. Between the two sits an order of magnitude — the difference between a curious gadget and a competitive advantage.
None of these mistakes are because you're slow or bad at this. They're because you started with AI the way anyone reasonable would when they don't have a specialist on the team. The specialist does it differently — and once you've seen the setup, you don't want to go back.