ChatGPT, Claude, and purpose-built AI contract generators are fast. They're also consistently incomplete in ways that create real legal exposure. Over the past six months, we've analyzed contracts drafted by three leading LLMs and Orin's contract module, then had them reviewed by three in-house counsel and two external practitioners. The pattern is stark: eight out of ten AI-drafted contracts contain at least one material gap that a junior lawyer catches in minutes. The gaps aren't in grammar or boilerplate comfort. They're in the architecture of risk allocation—the clauses that decide who pays when something goes wrong. A missing indemnity chain, a vague IP ownership clause, or jurisdiction language that contradicts your actual service location can cost tens of thousands in litigation or leave you holding liability you expected to shift. Here are seven gaps we've found across all three LLM platforms, what they cost, and a checklist to catch them before signature. Gap 1: Indemnity is one-directional or silent on causation LLMs default to symmetric indemnities: "Each party indemnifies the other for losses arising from its breach." That sounds balanced. In practice, it's dangerous. What lawyers found: indemnity clauses that don't explicitly chain causation. They say "losses arising from breach" but omit the requirement that the indemnifying party's breach must be the direct cause of loss, not merely a contributing factor. This creates a situation where, if a third party sues you and your vendor claims you were partly at fault, your indemnity obligation triggers even if the vendor was 90% responsible. Real example from our review: A vendor's AI-drafted services agreement said Vendor would indemnify Client for "third-party claims arising from Vendor's breach of this Agreement." Six months later, Vendor's integration broke, Client lost data, and a client of Client's sued. The Vendor was partially at fault but so was Client's infrastructure. Indemnity language was too broad; Client ended up in a three-year dispute. The fix: Indemnity must say "directly caused by" or "resulting solely from." And it must carve out the indemnifying party's liability for the indemnified party's own negligence. Lawyers call this a "sole negligence" carve-out, and LLMs almost always miss it. Gap 2: Limitation of liability has no cap on indirect damages Every AI contract generator caps direct damages. None we tested explicitly carved out what's typically uncapped: lost revenue, lost profits, lost data, and loss of business opportunity. These are called "indirect" or "consequential" damages in the law, and they're where the real exposure lives. What happens: Contract says "Neither party's liability shall exceed the fees paid in the preceding 12 months." That's a reasonable cap on direct damages. But without an explicit carve-out, a court might allow the other party to claim unlimited lost profits when a service fails. You capped direct damages, but you didn't cap the thing that actually hurts. The fix: Add: "Neither party shall be liable for indirect, incidental, consequential, or punitive damages, including lost revenue, lost profits, or loss of business opportunity, even if advised of the possibility of such damages." Then separately cap direct damages. LLMs know to cap damages; they just don't know to carve out the categories that matter. Gap 3: IP ownership defaults to the customer, but work-product language is silent If you're a service provider, this is critical. AI contract generators typically say something like: "Client owns all work product created under this Agreement." That's often what you want. But the language doesn't address what happens to pre-existing IP, templates, code libraries, or methodologies you brought to the engagement. Scenario: You're a consultant. You draft a client's pricing model using a framework you've used for ten years. The contract says "Client owns all work product." Client claims that includes your framework. You disagree. No lawyer would have written a contract that ambiguous—but LLMs do, every time. The fix: Split IP ownership explicitly: Custom work product: "All documents, code, and materials created specifically for Client's use shall be owned by Client." Pre-existing IP: "All tools, methodologies, code libraries, and templates existing prior to this Agreement shall remain the property of [Service Provider] and may be reused in future engagements." Improvements to pre-existing IP: "Any improvements or modifications to [Service Provider]'s pre-existing IP shall be owned by [Service Provider], provided Client retains a non-exclusive, perpetual license to use such improvements for Client's internal purposes." LLMs will write "Client owns work product," but they won't think to carve out your intellectual property. You have to. Gap 4: Jurisdiction and venue are silent on dispute escalation LLMs are good at stating jurisdiction. "This Agreement shall be governed by the laws of [State/Country] and dis