Your AI contract drafting tool just generated a 48-month payment schedule at ₹12,000 per month for a SKU your company doesn't sell. Your sales team caught it before sending. Your competitor's didn't. Large language models are precise at structure and persuasive at tone. They are terrible at facts. They invent pricing, SKU definitions, tax percentages, and payment terms with the same confidence they use for boilerplate language. The problem isn't that AI hallucinates—it's that contract pricing is too expensive to hallucinate on. This guide maps seven contract clauses that must stay locked to manual human review, shows real examples of LLM pricing failures, and explains how to design your contract workflow so AI helps without killing your deal math. Why Contract Hallucinations Cost More Than You Think When an LLM invents contract language about indemnification or confidentiality confidentiality obligations, the worst that happens is your legal review catches it and you rewrite. When it invents a payment schedule, SKU, tax rate, or discount, the worst that happens is your customer signs it—and your accounting and fulfillment teams discover the gap weeks later. Three patterns emerge in real hallucinations: Pricing fabrication: The model sees "monthly retainer" in the template and drafts a price it never saw in the deal context. It picks a round number or extrapolates from thin data. SKU invention: When product names are abbreviations (e.g., "SKU-2847"), the AI generates SKU numbers that don't exist in your catalog. Term confabulation: Tax rates, currency assumptions, payment day-of-month, and billing frequency get wrong—often internally consistent but factually absent from the source data. In each case, the customer signature makes it a binding agreement. Your CFO enforces it. You ship at a loss or lose the customer contesting the terms. The Seven Clauses That Must Stay Manual Not all contract sections carry equal financial risk. AI can safely draft boilerplate. These seven sections must pass human review before any customer sees them: 1. Payment Terms and Schedule This includes amount, currency, due date, payment method, late fees, and any installment breakdown. LLMs frequently invent payment amounts when they're not explicitly stated in the deal context. A common failure: the model sees "quarterly" and drafts a quarterly amount it constructed from surrounding numbers. Lock strategy: Create a template variable for [PAYMENT_SCHEDULE] that your sales system populates from your deal record, not from AI inference. Mark this block as "human mandatory review" in your approval workflow. 2. Product/Service Description and SKUs The model cannot reliably map service names to product codes or verify that a SKU actually exists in your catalog. It will invent plausible-sounding SKUs. Example: Your contract template says "Standard Support Package." The AI model drafts "SKU-SP-4892" which your product database has never seen. Your fulfillment team marks the order unshippable. Lock strategy: Require SKUs to be selected from a dropdown or imported from your product management system. Do not allow AI to draft or infer product codes. Use template variables bound to your CRM or product database. 3. Discount, Promotional, or Volume-Based Pricing Adjustments Discounts require authorization, often from a manager or approval chain. LLMs do not have access to your discount authority matrix. They will invent discounts that don't exist or exceed your policy. Lock strategy: Disable AI editing of discount fields. Require manual entry and manager approval before the contract is generated. Treat discounts as a separate approval step before contract drafting. 4. Tax Rate and Calculation Basis Tax rules vary by jurisdiction, product category, and customer type. An LLM cannot reliably infer the correct rate. It will often default to a hallucinated round number ("10%", "15%") rather than research the applicable rule. Lock strategy: Integrate your invoicing platform's tax engine into the contract workflow. Let the contract pull the tax rate from your accounting system, not from AI drafting. Verify jurisdiction before contract generation. 5. Liability Cap and Insurance Requirements While boilerplate liability language is safe for AI, the dollar amounts and thresholds must match your company's insurance and financial policy. LLMs often invent caps that sound legal but exceed your authority to commit. A real case: an AI model drafted "liability capped at 12 months of fees" where the deal was a 36-month contract worth ₹60 lakhs. The firm's insurance only covered 6-month caps. Lock strategy: Create a cap matrix in your template: bind liability caps to contract value and term using your policy rules, not AI logic. Do not let the model calculate or propose caps. 6. Service Level Agreements (SLA) and Performance Metrics SLAs must match your operational capability. An LLM drafting "99.99% uptime" or "4-hour response time" sounds credible but may exceed y