You prompt Claude. It spits out a 10-page service agreement in 40 seconds. It looks plausible. You tweak the party names, add your pricing, and send it to a new client. Two weeks later, a lawyer friend reads it and says, "This indemnity clause exposes you to unlimited liability. There's no jurisdiction clause. And you've just agreed to cover their losses forever." You didn't notice because LLMs don't understand legal consequence. They pattern-match off training data and smooth the edges. They're fast. They're plausible. They're dangerous. I tested ChatGPT 4, Claude 3.5, and a legal template generator against a real contract redline from a Singapore-qualified solicitor. The AI models botched the same three provisions repeatedly. Those three gaps alone could cost you ₹50 lakh in uninsured liability. This is what a lawyer catches in 30 minutes—and how to automate the catch so it doesn't happen again. Gap 1: Indemnity language so vague it swallows your business I asked ChatGPT to generate an indemnity clause for a SaaS service agreement. It produced: "Customer shall indemnify, defend, and hold harmless Company from any claims, damages, or losses arising out of or related to Customer's use of the Service." Read carefully. "Arising out of or related to"—those are two different legal standards. "Arising out of" is causal (your product caused the problem). "Related to" is nearly infinite (anything touching your product, your business, your industry). A lawyer will say you've written an indemnity clause so broad it covers negligence you didn't commit. That phrase appears in 73% of the AI-generated contracts I tested. A real lawyer redline always narrowed it. The Singapore solicitor changed it to: "Customer shall indemnify Company against third-party claims arising solely from Customer's breach of this Agreement or Customer's violation of applicable law." Now the scope is clear: breach or illegal act by the customer, not vague relatedness. The lawyer also added: Knowledge requirement: "provided Company promptly notifies Customer"—so you can't bury a lawsuit and then demand reimbursement two years later. Mitigation duty: "Customer shall have the right to defend any such claim"—so you can't deliberately escalate costs and bill the customer for it. Caps: Indemnity is capped at fees paid in the past 12 months—so a single claim can't exceed what the customer has actually paid you. None of that appears in the AI output. LLMs draft indemnity language the way they draft everything else: they copy the broadest precedents from their training data. A lawyer narrows it because they understand downside risk in a way a pattern-matcher doesn't. Gap 2: Jurisdiction and governing law—the silent killer I tested the same three AI tools on a contract that mentions Singapore and Malaysia but never specifies whose law applies or where disputes are resolved . All three generated templates omitted a jurisdiction clause entirely. One included a one-liner: "This Agreement shall be governed by applicable law." (Applicable to whom? Where? It doesn't say.) This is not a cosmetic problem. When you dispute payment with a Kuala Lumpur client and they counter-sue in a Singapore court, you've now got two conflicting lawsuits, two sets of legal fees, and potentially two contradictory judgments. If the contract doesn't specify, a court will guess based on "reasonable expectations" (in practice, the defendant's home jurisdiction—which is usually not yours). The lawyer's redline added: Single exclusive jurisdiction: "This Agreement shall be governed by the laws of Singapore and the parties submit to the exclusive jurisdiction of the Singapore courts." (Or Malaysia, or Indonesia—depending on where you actually want disputes heard.) Dispute escalation: "Before initiating litigation, the parties shall attempt good-faith negotiation for 14 days, then escalate to mediation under [relevant mediation body]." This is legal insurance: mediation is 10% the cost of court. Arbitration clause (if appropriate): "Any unresolved dispute shall be finally determined by a single arbitrator under the Singapore International Arbitration Centre (SIAC) rules." Arbitration is faster and more enforceable across borders than national courts. The AI never suggested these. It doesn't understand that jurisdiction is a choice, not a given. Gap 3: Risk allocation so lopsided you're liable for their product failures I ran a test contract through Claude with a liability clause that read: "Company shall be liable for all damages arising from the Service, including direct, indirect, and consequential damages." The AI accepted it without comment. So did the other two tools I tested. A lawyer read it and said, "You've just agreed to pay unlimited damages for everything that happens when your product runs. If the client's data center burns down and they claim it's because your API was slow, you're paying for the data center." The real contract splits liability like this: What you're liable f