QWE AI Academy If you’ve ever copy-pasted last month’s NDA, scrubbed out the client name, and prayed you caught every reference to “Acme Corp” – you already know the real problem with contract templates. It’s not drafting. It’s consistency. Every variation introduces risk. Every new clause someone quietly adds becomes the new “standard.” And when you finally try to use AI to fix this, you run into something nobody warns you about.
Ask any LLM to draft the same NDA twice. You get two different documents. Legal tech researchers documented this directly – ChatGPT produces a different version of the same clause on every run. That’s the opposite of “standard.” So before we get into prompts and tools: the goal isn’t to generate a contract. It’s to generate a system.
Why most AI contract tutorials miss the point
Jotform, Venngage, Legitt, Lumin, Otto – the workflow is identical across all of them. Describe what you want, click generate, edit, export. Legitt AI produces clause-structured drafts in under 60 seconds. Fast. But what you get is a one-off draft, not a template. There’s a difference, and it matters at scale.
A real template is reused fifty times – different parties, different dates, different fees – with the legal structure staying identical each time. AI tools don’t produce that by default. They produce a fresh draft on each run, with subtle variations in clause order, defined terms, and risk allocation. For a freelancer that’s annoying. For a company sending 200 contracts a month, it’s a compliance problem.
Build the template once. Lock the structure. Then fill.
The approach that actually works uses generic LLMs (ChatGPT, Claude, Gemini) deliberately – not to invent your contract, but to handle the structural reformatting while you control the substance.
- Start from a known-good base. Use a contract you’ve already relied on, or pull a vetted template from your industry association, a previous lawyer-reviewed agreement, or official government templates where they exist. Don’t start from a blank prompt. That’s where hallucinations live.
- Mark your variables. Every spot that changes per deal – party names, addresses, fees, dates, jurisdiction, deliverables – becomes a placeholder:
{{CLIENT_NAME}},{{FEE_AMOUNT}},{{GOVERNING_LAW}}. - Hand the marked-up version to the AI with one specific instruction: “Rewrite this contract preserving every clause exactly as written. Replace the bracketed terms with the placeholder syntax. Do not add, remove, or rephrase clauses. Output only the template.”
- Save the output as your master. This is your template. From here, you fill it – manually, or with a tool – instead of regenerating it each time.
Step 3 is the key move. You’re not asking the AI to draft. You’re asking it to reformat. Reformatting is something LLMs do reliably. Drafting legal language from scratch is something they get wrong at surprisingly high rates – which brings us to the numbers.
The hallucination tax
Turns out the risk is higher than most people expect. A Stanford RegLab study (published January 2024, arXiv:2401.01301) found legal hallucinations occur 58% of the time with GPT-4 and 88% with Llama 2 on specific verifiable legal queries. That’s not edge-case territory. And purpose-built legal AI doesn’t fully escape it – a follow-up Stanford HAI study found Lexis+ AI produced incorrect information more than 17% of the time; Westlaw’s AI-Assisted Research crossed 34%.
One in three clauses drafted from scratch may cite a statute that doesn’t apply, use a defined term inconsistently, or apply a rule from the wrong jurisdiction. The “reformat, don’t draft” approach sidesteps most of this – the AI is operating on text you already trust. Its job is structural, not legal.
Worth checking: After generating your template, paste it back into a fresh session and ask: “List every clause. For each, summarize in one sentence what it commits each party to.” If any summary surprises you, that clause needs human review before the template goes live.
There’s a broader question this data raises that the tool vendors don’t really answer: if even specialized legal AI hallucinates 17-34% of the time on benchmarked queries, what does that mean for the contracts nobody benchmarks? The honest answer is we don’t know yet. Which is exactly why the reformat-not-draft approach matters – it limits how much you’re relying on the model’s legal knowledge in the first place.
A worked example: freelance design service agreement
You have an old service agreement from a previous client engagement. Here’s the prompt:
You are a contract formatting assistant. Below is a service
agreement that I want to convert into a reusable template.
Rules:
1. Preserve every clause word-for-word. Do not rewrite.
2. Replace party-specific details with placeholders:
- Client name → {{CLIENT_NAME}}
- Client address → {{CLIENT_ADDRESS}}
- Project description → {{PROJECT_SCOPE}}
- Fee amount → {{FEE_AMOUNT}}
- Payment terms → {{PAYMENT_TERMS}}
- Start date → {{START_DATE}}
- Delivery date → {{DELIVERY_DATE}}
3. Do not add new clauses. Do not remove clauses.
4. At the end, list every placeholder you used.
Contract to convert:
[paste your existing contract here]
Output: your template. Every future project – copy it, find-and-replace the placeholders, send. Same structure every time. No drift.
No starting contract? Stop here. Either consult a lawyer or use a vetted source (your jurisdiction’s small-business administration often publishes them free). Generating from a blank prompt is exactly the scenario the Stanford research flags.
The traps
Confidentiality leak.Per Gavel’s analysis (verify current OpenAI data policy – this may have changed), public ChatGPT operates on shared infrastructure and may retain user inputs unless specific settings are enabled. Paste a real signed contract with client names and pricing into a chat window and you’ve potentially exposed all of it. Strip identifying details before pasting, use a paid tier with data retention disabled, or run a local model.
Jurisdiction blindness. The Stanford researchers specifically noted LLMs know California law better than Wyoming law – training data isn’t evenly distributed across jurisdictions. If you operate somewhere less represented, the model will confidently apply rules from somewhere else. Fix: specify governing law explicitly in your template, and have a local lawyer check the clauses that interact with it – typically limitations of liability, indemnification, and dispute resolution.
Quota traps. Jotform’s free Starter plan caps at 5 AI document generations as of current pricing (check their site – tiers change). Easy to burn through in one afternoon of iteration. Their Gold plan raises that to 100. Do your iteration in ChatGPT where prompts are unmetered on paid tiers, then use the contract platform only for final output.
Choosing your approach
| Approach | Best for | Main risk |
|---|---|---|
| General LLM (ChatGPT/Claude) reformatting an existing contract | Freelancers, small teams with a vetted base contract | Confidentiality if pasting real data |
| Hosted generator (Jotform, Venngage, Lumin) | Users who need built-in e-signature and storage | Quota limits; less control over clause text |
| Purpose-built legal AI (Juro, Legitt, Gavel) | Teams running 50+ contracts/month with a defined playbook | Cost; still hallucinates per Stanford data |
| Lawyer-drafted template + AI for variable filling only | High-stakes contracts (M&A, regulated employment) | Upfront cost |
For most people reading this: option one. You probably have an existing agreement that works. You don’t need a new platform. You need a system for not breaking what already works.
What to do next
One contract. The one you send most often. Open ChatGPT or Claude, run the prompt above – ten minutes to a reusable template. Use it for the next deal. The one after that. Stop regenerating from scratch.
FAQ
Can I just ask ChatGPT to write an NDA from scratch?
Yes. It’ll look fine. Per the Stanford data, there’s a real chance something is subtly wrong – wrong jurisdiction, a missing protection, a defined term that drifts. Low-stakes situation: probably acceptable. Real money or IP involved: start from a vetted base.
Do I need a lawyer to review every AI-generated template?
For a master template you’ll reuse repeatedly – one review is worth it. A freelance designer who pays a lawyer to review a service agreement once, then uses that template for two years across 50+ clients, has effectively paid a small fraction of one hourly rate per contract. The math works. What doesn’t work is paying for a fresh review every time, or skipping review entirely on a contract that actually carries risk. (Note: lawyer fees vary – treat any specific figure you’ve seen elsewhere as illustrative, not a quote.)
What about confidentiality when using ChatGPT for this?
Three options, and they’re not equal. Running a local model like Llama on your own machine means the text never leaves your machine – highest protection, highest setup friction. A paid ChatGPT tier with chat history and training disabled is the practical middle ground for most people. Scrubbing identifying info before pasting (replace real names with “Client A”, real numbers with placeholders) is the fastest option and works fine for template-building since you’re working with the structure, not the live data anyway. One common misconception: “enterprise” plans at various vendors don’t all offer the same data protections – check the specific terms, not just the tier name.