Legal work is expensive. A single contract review can run 40-80 hours at AUD 300-500/hour. That’s why law firms and in-house teams are looking hard at AI. But most legal AI products oversell what they can do, and most founders building in this space miss where the real money actually lives.
The efficiency gains aren’t where you think they are. They’re not in replacing lawyers. They’re in cutting the repetitive work that burns junior staff and delays deals. Let’s be specific about where AI wins in legal, where it fails, and what that means if you’re building here.
Document Assembly and Template Generation: Real and Repeatable
This is where legal AI actually works well. Document assembly-taking a set of inputs (party names, dates, jurisdiction, deal structure) and producing a first-draft contract-is a real efficiency win. A 30-page commercial lease that normally takes 6 hours to draft from a template can be generated in 10 minutes. The output isn’t perfect. But it’s 80% there, and a junior lawyer can review and adjust in 45 minutes instead of starting from a blank page.
The ROI is straightforward: you’re replacing routine cognitive labour with pattern matching, which is what large language models do well. The catch is that your documents need to follow predictable structures. Hyper-custom deals or edge-case jurisdictions break the model. But for high-volume, repeatable document types (NDAs, employment agreements, lease schedules, purchase order terms), this generates real AUD value.
If you’re building here, the moat is in your template library, your integration with document management systems, and your ability to handle local variation. An Australian legal AI product that understands NSW, VIC, and QLD variance and integrates with Xero and Zapier has something worth money.
Contract Review and Risk Flagging: Where AI Breaks
Here’s where the hype overextends. Every legal AI vendor claims their system can do contract review. Most can’t, not really.
The problem: contract review isn’t a single task. It involves spotting missing clauses, identifying risk in clause language, catching deviations from company-standard terms, understanding implied obligations, and contextualising risk against deal size and business intent. That last part-context-is where AI stumbles. A force majeure clause that’s fine for a AUD 50k service contract is dangerous in a AUD 5m capital works agreement. The AI doesn’t know the context without being told explicitly.
Where AI does useful work in review:
- Clause extraction. Pull out all price terms, liability caps, termination triggers, and renewal dates into a structured format. Fast and accurate.
- Precedent comparison. Flag deviations from your company’s standard template version. Good for catching sneaky changes.
- Completeness checking. Highlight missing clauses (e.g., “this contract has no force majeure clause”). Useful but not sufficient.
- Red-flag keywords. Find language that statistically correlates with litigation (e.g., “indemnify”, “sole discretion”, “gross negligence”). Adds noise but can catch things humans miss under time pressure.
What AI can’t do: tell you whether a liability cap is commercially reasonable in your industry, whether a non-compete is enforceable in your state, or whether a counterparty is hiding something in the structure. That’s judgement. AI can’t replace it.
If you’re building contract review, be honest about what you’re selling. You’re selling labour reduction on low-value tasks (reading and pulling data). You’re not replacing the lawyer; you’re replacing the paralegal doing first-pass work.
Due Diligence and Disclosure Analysis: Narrow but Valuable
M&A and fundraising involve mountains of disclosure documents. A Series A data room can have 500+ documents. A mid-market acquisition due diligence room has thousands. Reading through all of it is how lawyers bill 200+ hours per deal.
AI can help here, but only on specific questions. Good use cases:
- Identify all material contracts (anything above a revenue or spend threshold).
- Extract key terms from cap tables, shareholder agreements, and convertible notes.
- Find regulatory filings and summarise compliance status.
- Cross-reference information (e.g., find inconsistencies between representations in disclosure schedules and actual contracts).
- Produce a structured summary of intellectual property ownership claims and registration status.
None of this is complex. It’s tedious. AI is better at tedious than humans. A team of two lawyers + an AI system can do the document intake and categorisation work that used to take a paralegal two weeks. That’s real money saved, especially in boutique M&A firms or corporate teams doing multiple deals per year.
The technical bar here isn’t high, but the execution details matter enormously. You need good OCR for scanned documents, stable PDF parsing, and a reliable way to chunk and embed hundreds of documents so searches actually work. You also need to handle metadata-dates, party names, document type-so context isn’t lost.
Legal Research and Case Law: Surprisingly Weak for AI
You’d think AI would dominate legal research. Feeding it case law and having it find precedents seems like an obvious win. In practice, it’s disappointing.
The challenge is that legal reasoning is highly contextual and precedent-dependent. A case that looks similar on the surface can have opposite application depending on five lines buried in the judgment. LLMs are bad at this kind of careful, narrow reading. They hallucinate citations. They miss the actual holding because they get distracted by dicta. Traditional legal research databases (LexisNexis, Westlaw) are still better because they’re built on structured metadata, indexed by actual legal principles, and curated by humans who understand the law.
Where AI adds value: it can speed up the initial trawl, generate a reading list faster than keyword search, and help a lawyer explore “what if” scenarios. But for anything requiring certainty, you still want a human lawyer and a proper database. This market is defensible by incumbents with good data and isn’t where startups should focus right now.
Building Legal AI: The Realistic Roadmap
If you’re a founder thinking about building legal software, here’s what actually works:
Start narrow. Pick one document type (employment contracts, NDAs, leases) or one process (first-pass document intake in M&A). Build something that works perfectly for that small scope instead of something mediocre across five.
Integrate with existing tools. Lawyers use Word, email, Dropbox, and specialist software like LawSoft or file management systems. Your product needs to work within that workflow. API-first design. Plug-in architecture. Export to Word without corruption.
Handle Australia specifically. State-level variance in contract law, corporate law, and employment law is real. Templates that work nationally are compromises. Build for the Australian market explicitly, and you have defensibility against overseas products that don’t understand NSW or Queensland nuance.
Combine AI with people for quality. The best legal tech products layer human review into the workflow. AI does the work; a lawyer (on your team or in a partner network) validates it. This costs money, but it’s how you ship something people actually trust.
Price on value, not usage. If your product saves a firm 200 hours per year on a specific process, that’s worth AUD 60-100k per year to them (at blended rates). Charge for that value. Don’t charge per document or per review; lawyers hate usage-based pricing.
If you’re thinking about building something in this space, talk to Amora about your build. We’ve shipped several SaaS and AI products, and we know what it takes to get legal software into the hands of users who actually need it.
The Bottom Line
Legal AI works where it replaces routine, repeatable cognitive work-document generation, data extraction, categorisation. It fails where it tries to replace judgment or expertise. The founders who’ll win in legal tech aren’t the ones claiming AI will replace lawyers. They’re the ones building tools that make good lawyers faster and cheaper, one specific process at a time.
Got something you want built?
Amora Digital is an Australian software and AI agency. We scope it, build it, and ship it – live in 28 days. No offshore teams. No surprises.