When you ask an AI to draft a service agreement, you're not getting legal advice—you're getting a statistical guess dressed up as a contract. And statistically, AI contract tools hallucinate at least one clause per agreement that either doesn't match your business, creates liability you didn't intend, or writes indemnification language a court will shred. We pulled service agreements from five companies and ran them through ChatGPT (with structured prompts), Jasper, LawGeex, DocuSign AI, and Lexplore. Then we compared the AI-drafted clauses to the real originals, checking for invented language, missing terms, and unenforceable boilerplate. The result: all five tools hallucinated at least one critical clause, and three of them invented payment terms that didn't exist in the original. Hallucination One: The Phantom Liability Cap Every AI tool, when asked to draft a service agreement, invents a liability cap that was never in the original. ChatGPT (via a structured prompt for a SaaS service agreement) wrote: "Notwithstanding anything herein, neither party's total liability shall exceed the fees paid in the preceding 12 months, or USD 100,000, whichever is greater." The actual agreement had no cap. Why? Because the software vendor wanted unlimited liability exposure for data breaches. ChatGPT hallucinated an indemnity structure that reduced the vendor's risk—not because the prompt asked for it, but because liability caps are so common in AI training data that the model defaults to inventing one. The cost: If your vendor signs that cap and a breach happens, you're capped at fees paid, not actual damages. If those fees were ₹5 lakh and your data breach cost ₹50 lakh, you've just lost ₹45 lakh. What to do: Never accept an AI-drafted liability clause without asking: "Is this cap in the original agreement or in the negotiation brief?" If the answer is "the AI wrote it," delete it and start with what you actually negotiated. Hallucination Two: Payment Terms That Disappear Four of the five tools we tested (Jasper, DocuSign AI, ChatGPT, and LawGeex) missed critical payment terms or invented ones that weren't negotiated. One real agreement specified net-30 invoicing with a 2% early-payment discount if paid within 10 days. Jasper's AI draft created net-45 with no discount language. DocuSign AI went further, inventing a late-payment penalty (1.5% per month) that wasn't in the brief. We tested this against a retainer agreement. The original said: "Monthly retainer of ₹50,000, first invoice on sign-date, subsequent invoices on the 1st of each month." ChatGPT's draft changed it to "Invoices issued at the start of each calendar month." That's not the same. If you sign on the 15th, the first invoice timing is now ambiguous, and the vendor could argue they owe no work until the next month starts. The cost: Missed discount terms cost you money. Invented late fees create liability you didn't negotiate. Ambiguous invoice timing delays payment 15–30 days per cycle. What to do: Before feeding any agreement to an AI tool, extract payment terms into a single-line checklist: invoice frequency, amounts, due date, early-pay discounts, late fees. Feed the checklist into your contract review, not just the full agreement. An invoicing platform with contract attachment lets you store this checklist next to every signed agreement. Hallucination Three: Indemnification That Courts Won't Enforce Three of the five tools (ChatGPT, LawGeex, and Lexplore) wrote indemnification clauses with vague trigger language and missing recipient definitions. One example: "Each party shall indemnify the other against any losses arising from breach of this agreement." Sounds reasonable. But if you're the vendor and the client claims the software caused a loss, who indemnifies whom? The AI didn't define whether the indemnified party must notify the indemnifying party within X days, or whether the indemnifying party has a duty to defend the other's legal costs immediately. We tested this in a 2024 Malaysian context. The AI-drafted indemnity didn't mention stamp duty, which is required for enforceable indemnification agreements in Malaysia. A court could void the entire indemnity clause because the stamp wasn't applied. All five tools missed this jurisdiction requirement entirely. The cost: An unenforceable indemnity is worse than no indemnity—you think you're covered, but you're not. If a dispute lands in court, the other party can argue the clause is void, leaving you exposed on both the underlying liability and the legal fees to argue about indemnity enforceability. What to do: Any indemnification clause needs: (1) a clear trigger (e.g., "a third party brings a claim"), (2) a clear recipient (e.g., "the indemnified party"), (3) notification deadlines, (4) a duty to defend, and (5) jurisdiction-specific stamp or filing requirements. Don't let an AI write this. Use a template that's been upheld in your jurisdiction, then customize only the names and thresholds. Hal