FDD Update Engine: AI + Deterministic Drafting for Franchise Disclosure Documents

I was hunting for an untapped opportunity where LLM document synthesis and analysis could do real work that nobody had automated yet. Franchise Disclosure Documents (FDDs) was the biggest wedge I found. Every U.S. franchisor — roughly 9,000 brands — has to update its 23-item FDD every year, it’s done by hand by attorneys at $4,000–$15,000 a filing, and no one is automating the franchisor side. I also have a government angle to lean on (Secret Clearance through ANG), which matters for the larger adjacent wedge (defense / CUI compliance documents). The full market reasoning lives in the venture assessment.

And So I built a demo to figure it out, and while building it changed the idea.

The original plan was 23 extraction prompts: read each item out of the prior FDD and re-generate the text from scratch. What I realized is that’s not how the work actually happens. An attorney updating an FDD doesn’t draft 23 items from a blank page — they put last year’s FDD next to this year’s source documents and compare them section by section, fixing what changed and keeping it compliant.

So I pivoted. The real product is simpler and matches the actual workflow: upload last year’s FDD, upload the new inputs (amended franchise agreement, outlet roster, updated data), and get a side-by-side, tracked-changes comparison the attorney can review and finalize — instead of writing from nothing. The deterministic items (9, 17, 20, 23) are generated in pure Python; the narrative items are diffed by an LLM with the prior text routed per item.

See it

• ▶ Try the live demo → — it diffs two real consecutive filings: Alloy Personal Training’s 2025 FDD against its 2026 FDD (both public, from the Minnesota CARDS registry). Hit run and the engine walks all 23 items live, surfacing every change an attorney must make each one quoted verbatim from the actual filing. Because the 2026 FDD is the answer key, you can check the engine against what the franchisor’s lawyers really filed: it catches the outlet count jumping 78 → 129 (+51), fee changes, and the advisory-council that formed mid-year. The narrative items run on Gemini 2.5 Flash-Lite behind a verbatim-substring gate that cannot invent a value; items 9/17/20/23 are computed in pure Python with no model at all.
• The venture assessment behind it → — the market, the incumbents, the unauthorized-practice-of-law risk, and why I led with FDD.
• Code on GitHub →