Your operations manager fed a vendor agreement into an AI contract generator and got a first draft back in 90 seconds. It looks professional. It has indemnity clauses, payment terms, dispute resolution. Then your lawyer opens it and marks it in red: the liability cap is a vague percentage, the payment cycle doesn't match your cash flow, and the governing law is set to a state you don't operate in. This is not a rare edge case. Language models are trained on millions of contract templates, and they interpolate seamlessly—but interpolation is not legal reasoning. An LLM does not understand your jurisdiction, your risk tolerance, or your standard operating procedures. It invents plausible-sounding legal language that reads like a contract but is often misaligned with your actual needs. Before you sign—or before your lawyer wastes billable hours rewriting—run three mandatory checks. They take 15 minutes and will catch 80% of the gotchas. Check 1: Liability Cap Must Be a Variable Slot, Not a Fixed Percentage AI-generated contracts routinely hard-code liability caps as percentages of the contract value—typically something like "neither party shall be liable for damages exceeding 50% of fees paid in the preceding 12 months." The problem: that number almost never matches your actual risk exposure. If you're a SaaS company with $2M in annual contract value (ACV), a 50% cap means your vendor is liable for a maximum of $1M in damages. But if their service goes down for a week and you lose $500K in customer churn, you're capped at $1M and you still absorb a loss. Worse, if you're the vendor being sued, a 50% liability cap might expose you to claims that dwarf your profit margin on that contract. The fix: Replace the percentage with a formula or a tiered variable. For example: "Liability is capped at 12 months of fees, except for indemnification of third-party IP claims, which is uncapped." Or: "Cap is 100% of fees for Service A, 25% of fees for Service B." The key is that the cap must be defensible in your jurisdiction and matched to the actual revenue or cost impact of failure. Critically, AI will invent a cap that sounds balanced but is not calibrated to your business model. A 50% cap is reasonable for a low-risk SaaS subscription. It's terrible for a high-stakes integration where downtime ripples across your entire revenue stream. Your lawyer will ask: "What happens if they fail? How much do we actually lose?" The AI doesn't ask that question. Check 2: Payment Terms Must Reference Your Standard Billing Cycle, Not a Generic 30-Day Period AI drafts payment terms like this: "Invoices due net 30 days from invoice date." But your standard cycle might be net 45 for vendors, net 15 for your own customers, or monthly arrears with a 10-day remittance window. An LLM does not know this. It picks a conventionally safe middle ground and moves on. The consequence: you now have a contract where you pay in 30 days (matching the AI's draft) but your customers pay in 45 days (your standard). That's a 15-day working capital gap. Multiply that by 10 vendors and you've artificially tightened your cash flow by weeks. Worse, if the contract has automatic late fees—"1.5% monthly interest on overdue amounts"—and your team misses a payment by 3 days due to processing delays, you're liable for accrued interest that a human negotiator would have waived. The fix: Before signing, replace the generic payment term with your actual cycle. If you're on net 45 with most vendors, write it explicitly: "Invoices due net 45 days from invoice date. Payment cycles align with [your company name]'s standard billing calendar, month-end processing." If there's a late-payment clause, tie it to a written notice requirement —not automatic accrual. For example: "Late fees apply only if payment remains unpaid 10 calendar days after written notice of non-payment." This is not pedantry. A 15-day swing in payment timing can be the difference between making payroll and not, especially in the first year when you're growing fast and cash is tight. Check 3: Governing Law Must Be Locked to Your Jurisdiction, Not a Vague Default AI often drafts governing law clauses like: "This agreement shall be governed by and construed in accordance with the laws of the state of [blank] without regard to conflicts of law principles." The [blank] is often left unfilled, or the LLM fills it with a U.S. state at random—or worse, the default jurisdiction in its training data (often New York or Delaware, neither of which may apply to you). If you're a Malaysian company selling to a Singapore vendor, you do not want the contract governed by New York law. You want it governed by Malaysian law or, by negotiation, Singapore law. Why? Because if a dispute arises, you'll litigate in your own jurisdiction under your own legal system. Anything else is a forced journey to an unfamiliar court, unfamiliar judges, and unfamiliar precedent—all of which cost more to defend. Similarly, if you're in Indonesi