Your CRM's AI assistant just drafted a contract in 90 seconds. It looks plausible. You tweaked a few names, added your logo, and sent it over. Six months later, the other party invokes a liability cap you never intended, an indemnity clause that reads backwards, and a dispute resolution clause that points to a court three continents away. The AI didn't hallucinate those words—it borrowed them from training data that had nothing to do with your business, your jurisdiction, or your actual risk profile. Courts don't care where the language came from. They will enforce it. This isn't theoretical. Contract disputes involving AI-drafted terms are already landing in discovery, and judges are treating machine-generated boilerplate the same way they treat any other written agreement: as binding. The three liability gaps below are the ones LLMs repeat most often—and the ones that cause the most damage when enforced. Gap One: Liability Caps Set Too Low for Your Exposure When an LLM generates a contract, it tends to pull liability cap language from templates in its training data. Those templates often reflect small SaaS deals, non-critical services, or consumer-grade risk. A ₹50,000 cap might have made sense for a newsletter signup. It does not make sense for a ₹20 million implementation. The problem: courts enforce the cap as written. If your contract says liability is limited to the fees paid in the last 12 months, and your service fails in month 13, you are capped at fees from month 1. If those fees were ₹10,000 and the failure cost the other party ₹500,000, you are paying the difference out of pocket. AI-drafted contracts rarely adjust liability caps by contract value, service criticality, or data sensitivity. A human reviewing a contract will often catch this and negotiate. But if you're using AI to draft and then auto-approve through e-signature without counsel , the cap locks in. What to audit: Is the liability cap a fixed amount or a percentage of annual contract value? Fixed amounts almost always end up too low. Does the cap apply to direct damages only, or does it exclude indirect damages, data loss, and business interruption? (Asymmetry here often favors the other party.) Is there a separate, higher cap for indemnification, IP infringement, or gross negligence? If not, you've capped your own indemnity to something absurdly low. Does the cap survive contract termination, or does it reset? Some LLM drafts create 18-month caps that restart on renewal—which means the other party can claim damage on year 1.5 and still be within the renewed cap. Courts enforce liability caps as written. If your AI-drafted contract says damages are capped at ₹50,000, that's what you pay—even if the actual loss is ₹500,000. Judges don't ask whether the AI understood your business. Gap Two: Indemnification Clauses That Read Backwards Indemnity language is dense. An LLM will copy it accurately—but often doesn't reverse it correctly when switching parties. This leads to contracts where you indemnify the other party for negligence they caused. Example: An AI drafts "Company A indemnifies Company B against all claims arising from Company B's use of Company A's software." That's a one-way gate. You (Company A) are promising to pay for their losses even when they misuse your software or fail to apply security patches you provided. The clause should read "against claims arising from Company A's breach of this agreement" or "from defects in Company A's software." But AI templates often blur that line. Worse: asymmetric indemnity rarely cuts both ways. You'll see clauses where you indemnify them for negligence, intellectual property issues, and regulatory violations—but they indemnify you only for breach of confidentiality. Courts have enforced this lopsidedness. They read the contract as written, not as you intended. LLMs also miss jurisdiction-specific indemnity limits. In many common law jurisdictions, you cannot indemnify someone for their own willful misconduct. Some AI drafts still try. In Singapore and Malaysia, indemnity clauses for breach of statutory duty are often unenforceable if they fail to meet prescribed language. An LLM trained on generic US templates will miss both. What to audit: Who indemnifies whom? Draw an arrow. If both arrows point toward you, that's asymmetric. What events trigger indemnity? "Claims arising from Company B's use of our software" is too broad. "Claims arising from our breach of Section 5" is tight. Is there a carve-out for the indemnified party's negligence? Most well-drafted contracts exclude that. AI often forgets it. Does your indemnity include defense costs? If yes, the other party can rack up ₹100,000 in legal fees and bill you in real time—even if they ultimately lose. Is there a notice requirement? (You should have to be told within 30 days or lose indemnity.) Does the clause survive termination? Many AI drafts leave indemnity hanging for 3–5 years post-termination, creating open-ended exposure.