Your AI contract drafting tool just regenerated your payment terms. Thirty days became net 60. A ₹50,000 minimum became ₹500,000. Nobody caught it until the client signed. This is not a corner case—it's the default behavior of any LLM given a contract to 'improve' without guardrails. The fix is not to reject AI contract drafting. The fix is to separate what the AI can draft from what it absolutely cannot touch. Lock your critical business terms into a frozen template structure before you ever hand the contract to an LLM. Mark those zones as untouchable. Audit the output against the locked template. Only then does AI contract drafting become safe. Why AI rewrites the terms you need unchanged Language models are trained to make text 'better'—clearer, shorter, more legally coherent. When an LLM reads a contract clause, it does not know which numbers and conditions are business-critical and which are boilerplate. It optimizes everything equally. A liability cap of ₹5 lakh looks 'unbalanced' next to a 15-year non-compete, so the model adjusts both toward what its training data says is 'standard.' The result: payment terms drift, liability limits vanish or explode, renewal conditions flip, exclusions disappear. You end up with a legally coherent contract that contradicts your actual deal. The solution is architectural. You do not ask the LLM to 'draft a contract.' You give it a skeleton—a template with hard boundaries—and ask it only to fill the gaps the skeleton expects. Seven clauses that must never be auto-generated These are the terms that sink deals or leak liability if the AI changes them: Payment amount and schedule. If your contract says ₹50,000 on day 1 and ₹50,000 on day 90, lock both numbers and both dates. Do not let the AI merge them into a single payment. Do not let it add 'or net 30.' The AI will optimize the payment term as 'cleaner' without understanding that you've already negotiated two tranches with a client for cash flow reasons. Liability cap. If you've agreed to cap liability at ₹10 lakh, the LLM will rewrite it as 'not to exceed the fees paid in the preceding 12 months' to make it 'more standard.' Now your cap is dynamic and you've lost control of your downside. Renewal terms and termination notice. Auto-renewal at 30 days notice looks 'fair' to an LLM. Your business model might require 90 days notice or manual renewal only. The AI will 'improve' the notice period without asking. Exclusions and carve-outs. If your contract excludes third-party data or limits your obligation to systems you control, the AI will soften the language to sound less defensive. Your exclusion becomes aspirational instead of binding. Pricing adjustment triggers. If pricing adjusts only on annual renewal or only if costs rise 10%+, lock both the timing and the threshold. The LLM will treat these as negotiable and make them 'more flexible.' Confidentiality duration. If secrets stay confidential for 5 years after the contract ends, do not let the AI make it 'for the duration of the relationship.' That's half the protection. Indemnification scope. Which party indemnifies which? For what? The AI will broaden mutual indemnification to seem 'fair' and you'll end up defending the client against their own negligence. Every one of these has been negotiated with your client or set by your risk team. The AI does not know that. Mark them as frozen from the start. Building the template structure with no-edit zones Create your contract template in a tool that supports field-level locking or versioning. If you're using contract e-signature software , check whether it allows you to mark certain sections as 'read-only for AI' or to version-control template changes. If it does not, use a three-layer approach: Layer 1: The frozen skeleton. All seven critical clauses above, plus your company name, governing law, signature block, and insurance requirements. These go into a master template that the AI never edits. They are your baseline. Layer 2: The draft zones. Description of services, deliverables, timelines, specific client names, and project scope. These are where the AI adds detail. They are important but not legally binding in the same way—they describe what you do, not what you're liable for. Layer 3: The audit checklist. After the AI drafts, you run the output against a checklist: Does the payment amount match the deal? Is the liability cap at ₹10 lakh? Is the renewal notice 90 days? Did any exclusion get reworded? This takes 5 minutes and catches 90% of hallucinations. The principle: If a clause affects liability, pricing, or duration, it goes in the frozen skeleton. If it describes work or scope, it goes in the draft zone. The AI never touches the skeleton. How to mark sections as untouchable in the AI workflow If your contract software does not have native read-only zones, use these workarounds: Approach 1: Separate documents. Keep the frozen clauses in a separate 'Master Terms' document that never gets regenerated. The AI dr