You prompt ChatGPT: Draft me a service agreement for a digital marketing retainer in Malaysia. Thirty seconds later, you have prose that reads like a contract—headings, payment terms, termination. But your lawyer flagged seven things that AI missed, each one carrying real cost if a dispute lands in court or arbitration. This is not about AI being bad at language. It's about AI having no incentive to ask the hard questions that lawyers ask. Why AI contracts feel complete but aren't Large language models are trained on public contract templates, case law summaries, and legal writing—but not on the lived experience of fighting a contract dispute. A lawyer has sat across from unhappy clients, watched deals collapse over ambiguous language, and seen what happens when a clause gets read in the worst possible way. AI has not. AI excels at structure and tone. It produces contracts that look professional and that cover the obvious: scope, fees, confidentiality. But seven specific liability gaps appear in AI drafts with striking consistency. Each gap represents a scenario where ambiguous language will be read against you. AI drafts the skeleton. A lawyer stress-tests it against real failure modes. Gap 1: Indemnification scope is too narrow or circular AI typically writes indemnification as a simple swap: Vendor indemnifies Client against claims arising from Vendor's breach. What's missing: Third-party IP claims — If you build a campaign using a stock image, AI rarely specifies who indemnifies whom if the image owner sues. A lawyer forces you to specify: Does the vendor indemnify the client for vendor-sourced assets? Contributory negligence — AI almost never address the scenario where both parties share blame. Courts read silence as strict indemnity (you pay even if they were negligent). A lawyer adds language carving out client negligence and limiting your exposure. Procedure and defense control — AI drafts indemnity but forgets to say who controls the lawsuit. Can the client settle without your consent and then invoice you? A lawyer requires you to defend the claim and limits settlement exposure. SE Asia twist: Malaysian and Singapore courts interpret indemnification clauses narrowly. Thai courts even more so. AI doesn't know this. A lawyer adds explicit language acknowledging the jurisdiction and citing local precedent. Gap 2: Limitation of liability is absent or unilaterally protective AI often writes: Neither party shall be liable for indirect, consequential, or punitive damages. This sounds balanced. It is not, because: Cap on direct damages is missing — You limit type of damages but not amount . Client's campaign tanks and causes ₹50 lakhs in lost revenue. You capped indirect damages, but direct damages (breach, failure to deliver) have no cap. A lawyer forces you to add: Total liability capped at 12 months of fees. Carve-outs swallow the cap — AI rarely specifies what pierces the cap. Most contracts exempt IP infringement, data breach, or willful misconduct from the limitation. But AI drafts these exceptions so broadly that nearly every claim ends up exceeding the cap. A lawyer narrows the exemptions ruthlessly. No mutual cap — AI may cap vendor liability but not client liability. In a retainer dispute, client breaches by non-payment; you should have equal protection. A lawyer enforces symmetry or documents why asymmetry is necessary. SE Asia reality: Indonesia and Malaysia courts often strike or narrowly read liability caps, especially when they shield large platforms from SME claims. A lawyer knows this and drafts the cap to survive local scrutiny—usually by anchoring it to actual damages, not arbitrary thresholds. Gap 3: IP ownership is vague or assumes vendor retains all rights This is where AI's training data creates the most dangerous assumption. Most public templates favor the vendor (agency, freelancer) retaining IP. AI, mimicking this pattern, drafts: Vendor retains all intellectual property rights in deliverables unless otherwise agreed. But your client expects to own the marketing assets they paid for. Ambiguity here costs deals: What counts as deliverable? — AI rarely defines it. Does it include methodology, templates, custom code, or only final designs? Client assumes all of it. You assume only the final output. Six months later, they want to reuse your templates with another agency and you have a lawsuit. Pre-existing IP is not ring-fenced — AI drafts: Client owns deliverables, Vendor retains pre-existing tools and methodologies. But what is pre-existing? AI doesn't define it. A lawyer specifies: Vendor retains ownership of [list specific tools/frameworks]; all custom work, modifications, and compilations become Client property upon full payment. License vs. ownership is conflated — AI may give the client a license to use deliverables but not own them. Sounds minor. It means client can't modify, sublicense, or transfer the asset. If they change agencies, they start over. A lawyer clarifies whethe