An agency owner sits down at 2pm on a Friday with a Statement of Work (SOW) half-drafted and a client waiting for a signed contract by Monday morning. They feed the client name, project scope, and budget into an AI contract generator. Ninety seconds later, a 15-paragraph document appears. The owner glances through it, clicks 'sign', and sends it over. Monday morning arrives. The client's legal team flags three problems: liability is capped at nothing (meaning unlimited exposure), intellectual property ownership is vague enough to create a dispute later, and termination terms give the client a unilateral exit with no notice. The contract is now a liability. This is not a hypothetical. AI contract generators are fast and plausible-looking, but they trade specificity for speed. For agencies billing time, taking on uncontrolled liability, or transferring IP that should stay proprietary, a 90-second draft is a 90-day problem waiting to happen. What AI does well: templates with zero customization AI excels at documents that are truly boilerplate—mutual legal promises with no asymmetry and no buried traps. NDAs (Non-Disclosure Agreements). These are symmetric: both parties agree not to share secrets. The substance is nearly identical across deals. An AI NDA template typically covers definition of confidential information, permitted disclosures, term length, and return of materials. Agencies can generate these in seconds and sign them same-day. Service Order templates. A simple "Client authorizes us to perform X service for $Y, due on Z date" with standard payment terms. No complexity, no hidden risk. Basic SOW shells. The structure—deliverables, timeline, payment schedule, approval process—is repeatable. AI can fill in the outline. You add the specifics. AI saves time on symmetrical, repetitive language. It fails catastrophically on anything that shifts risk or control to one party. The pattern is clear: if the contract would be identical across 100 clients, AI can draft it. If it varies by client risk profile, budget, or what you're agreeing to do, AI is a first draft at best. Where AI drafts create exposure: the three fatal gaps Services contracts for agencies contain three categories of clause that AI routinely gets wrong because they require judgment, not pattern matching. Liability caps and indemnification An AI contract might read: "Neither party shall be liable for indirect, incidental, or consequential damages." This sounds protective. But it doesn't say what the cap is . If you cause a client data breach worth $500k, and your contract has no dollar limit, that liability is yours. A human-drafted contract typically reads: "Service Provider's total liability under this Agreement shall not exceed the fees paid in the preceding 12 months." This is specific, directional, and enforceable. AI rarely generates this specificity unprompted. Worse, AI often skews toward the template, not the negotiation. If your standard is "liability capped at 1.5x annual fees", but the client is enterprise and pays $300k/year, capping at $450k exposure matters. An AI generator doesn't know to flag this. Intellectual property ownership AI frequently produces vague IP clauses like: "Client owns deliverables. Service Provider retains rights to tools and methodologies." This sounds fair. In reality, it creates disputes: Does "tools and methodologies" include custom code built for the client? Can you reuse the design system you built for Client A when you take on Client B? Who owns pre-existing components you used? An agency needs IP clarity defined upfront: what you can reuse (your templates, frameworks, libraries) versus what transfers to the client (custom deliverables). AI drafts these as generalities. A human review catches the gaps and fills them in with specifics. Termination and change order process An AI termination clause often reads: "Either party may terminate this Agreement with 30 days' notice." This kills scope-managed work: If you're three weeks into a four-week project and the client terminates, you lose the final payment but you're also stuck mid-delivery. There's no process for change requests, so scope creep has no brake. If the client terminates after you've started but before deliverables, you have no provision for payment for work completed. A functional termination clause specifies termination-for-cause (breach with 15 days to cure), termination-for-convenience (only after milestones complete or with payment for work done), and a change order process. AI doesn't generate this level of operating detail. The real workflow: AI draft + human checkpoint The fastest safe path for agencies is to use AI as a first draft generator, then apply a manual review checkpoint before any client sees it. Phase 1: AI generates the structure (5 minutes). Feed the AI your project type, scope, budget, and timeline. Ask it to generate a SOW with sections for deliverables, timeline, payment schedule, and acceptance criteria. This gives