# Why We Chose Automation Over AI Generation, 10 Years Ago, and Still Today Blog | LeasePilot [Blog](/blog)Technology # Why We Chose Automation Over AI Generation, 10 Years Ago, and Still Today The deliberate technical decision to automate legal language selection rather than generate it, and why this position is stronger than ever in the age of LLMs. ![Lior Kedmi](/_next/image?url=%2Fleadership%2Flior-kedmi.jpg&w=3840&q=75&dpl=dpl_2umEzFMLLmFZHhmrz8MoJu6VB8Uh) Lior Kedmi CTO June 6, 20247 min readCopy link TL;DR AI generation creates text that needs verification. Automation assembles pre-approved text that doesn't. For legal documents where every word carries liability, the verification burden is the entire game. § 01 ## [The Question We Get Every Week](#the-question-we-get-every-week) "Can't AI just write my leases now?" It's a fair question. Large language models can generate remarkably coherent text. They can draft emails, write marketing copy, even produce code. Why not lease documents? We've been building lease automation technology for over a decade. When we started, "AI" meant something different than it does today. But our core architectural decision, **automate language selection, don't generate language**, was intentional then and remains correct now. Here's why. § 02 ## [Generation vs. Automation: The Core Difference](#generation-vs-automation-the-core-difference) **Generation** means creating new text. An LLM takes a prompt and produces language that didn't exist before. The text is probabilistic, the model predicts what words should come next based on patterns in training data. **Automation** means selecting and assembling existing text. The system chooses from your pre-approved language based on rules and inputs. No new text is created. The output is deterministic, the same inputs produce the same output every time. This distinction sounds subtle. It's not. It's the difference between a system you can trust and a system you have to verify. AI is probabilistic; lease calculations must be deterministic. Generation creates text you must verify. Automation assembles text you've already verified. § 03 ## [The Verification Burden](#the-verification-burden) Here's the question that matters: **Who is responsible for checking the output?** With generation, every output requires verification. The AI might produce something that looks right but contains a subtle error, a wrong term, an ambiguous phrase, a clause that contradicts another provision. Someone has to read the entire document carefully to catch these issues. With automation, the verification happened upfront. Your language was reviewed by your attorneys. The selection rules were validated. The assembly logic was tested. Once the system is configured, the output is reliable by design. **Generation shifts the verification burden to every document.** **Automation moves the verification burden to system configuration.** For one lease, this difference might not matter. For 200 leases per year, it's the difference between scalable efficiency and a false promise. Key Takeaway The verification burden is the entire game. AI generation doesn't eliminate work, it moves it to the most dangerous place: after the document is drafted. § 04 ## [What LLMs Actually Get Wrong](#what-llms-actually-get-wrong) We've tested large language models on commercial lease drafting. The results are instructive. ### Problem 1: Confident Incorrectness LLMs don't say "I don't know." They generate plausible-sounding text even when that text is wrong. A model might produce a rent escalation clause that sounds professional but calculates incorrectly. It might generate a CAM provision that uses terms inconsistently. It might create language that conflicts with provisions elsewhere in the document. The errors aren't obvious. They require careful legal review to catch. That's exactly the work you were trying to avoid. ### Problem 2: Inconsistency Ask an LLM to draft the same clause twice with identical prompts. You'll get different outputs. Maybe the differences are trivial. Maybe they're not. You won't know without comparing them. For a legal document, consistency matters. The defined term "Premises" should be used the same way throughout. Cross-references should be accurate. Parallel provisions should use parallel language. LLMs don't maintain this consistency reliably. ### Problem 3: Training Data Problems What did the model learn from? Publicly available lease documents vary wildly in quality. Some represent excellent drafting. Some contain exactly the ambiguities and errors that generate disputes. The model doesn't distinguish, it learns patterns from all of it. Your lease forms represent your standards, refined through your experience. An LLM trained on the internet's lease documents doesn't know your standards exist. ### Problem 4: The Liability Question When AI-generated text causes a problem, who's responsible? If an attorney drafts a clause and it fails, there's professional accountability. If a system automates your pre-approved language and something goes wrong, the language itself can be examined and corrected. If an LLM generates text that causes a dispute, what's the recourse? The model doesn't remember what it produced or why. There's no audit trail of reasoning. The liability exists, but accountability is diffuse. § 05 ## [Where AI Actually Helps](#where-ai-actually-helps) This isn't an anti-AI position. We use machine learning where it adds value: **Extraction**: Reading executed leases and pulling out structured data. This is pattern matching, identifying where the rent amount is, what the term dates are, which provisions exist. LLMs are useful here because errors are catchable (the extracted rent either matches the document or it doesn't). **Analysis**: Comparing lease terms across a portfolio. Identifying patterns in negotiation outcomes. Flagging unusual provisions for review. These are analytical tasks where AI augments human judgment rather than replacing it. **Search**: Finding relevant precedent. "Show me leases with similar co-tenancy provisions." Natural language understanding makes this easier. The pattern: AI helps with reading and understanding documents. Automation handles creating them. § 06 ## [Why This Position Is Stronger Today](#why-this-position-is-stronger-today) You might think the emergence of more powerful LLMs would weaken our position. The opposite is true. **Higher expectations**: As AI improves, people expect more from it. The gap between "this looks pretty good" and "this is reliable enough for a legal document" becomes more frustrating, not less. **More scrutiny**: Legal and compliance teams are increasingly skeptical of AI-generated content. The question "was this written by AI?" now triggers additional review requirements. Automation that assembles your pre-approved language doesn't face this scrutiny. **Demonstrated failures**: High-profile cases of AI hallucination in legal contexts have made the risks concrete. The attorney who submitted AI-generated briefs with fabricated citations is now a cautionary tale everyone knows. **Compounding returns**: The longer your language and deal logic live in a structured system, the more valuable that system becomes. Every refinement, every new clause variant, every rule adjustment, it compounds. That's institutional knowledge encoded and preserved, not scattered across Word documents. § 07 ## [The Honest Trade-off](#the-honest-trade-off) Automation requires upfront work. You have to encode your language. Define the rules. Configure the variants. This takes time and expertise. Generation feels easier at first. Prompt the model, get a draft. But that ease is illusory, the verification work you skipped comes back multiplied. We made the trade-off deliberately: invest in configuration so that execution is reliable. Over more than a decade, that investment has compounded. Our customers have institutional knowledge encoded in rules and refined through thousands of transactions. They have systems that produce consistent, reliable output. An LLM prompt, no matter how sophisticated, can't replicate that. § 08 ## [The Future](#the-future) Will AI eventually be reliable enough to generate legal documents? Maybe. The bar is high, not "mostly correct" but "completely reliable." Legal documents don't grade on a curve. Until then, the right architecture is clear: automation for creation, AI for augmentation. Your language, assembled by rules you defined, not language synthesized from statistical patterns that humans must then verify. That was the right call when we started. It's still the right call today. * * * We didn't avoid AI because we were skeptical of technology. We avoided generation because we understood the problem. Commercial leases aren't creative writing. They're precise instruments where every word carries weight. Automation respects that. Generation doesn't, not yet. § Adjacent reading ## More from the ledger [§ 01NOV 20, 2024 Technology ### AI in Commercial Leasing: What It Can and Can't Replace David Saltman8 MIN READ Read →](/blog/ai-commercial-leasing-what-it-can-cannot-replace) [§ 02MAR 13, 2026 Technology ### Automation vs. AI for Lease Drafting: A Practitioner's Guide LeasePilot Team8 MIN READ Read →](/blog/automation-vs-ai-lease-drafting-guide) [§ 03FEB 06, 2025 Technology ### Why Generic Document Automation Failed CRE Legal Teams David Saltman7 MIN READ Read →](/blog/why-generic-document-automation-failed-cre) § See it in practice ## Reading about it is one thing. Watching it happen is another. See LeasePilot draft a lease in your team’s own templates, with your clauses and your defaults. [Schedule a Demo](/demo)