How to Analyze a Franchise Disclosure Document (FDD) with AI

You can analyze a Franchise Disclosure Document with AI in minutes: feed it the FDD (or connect it to a structured FDD database), then ask for the Item 7 investment range, the Item 19 earnings, the Item 20 closure counts, and any litigation in Items 3 and 4. AI handles the data extraction. You handle the judgment, and a lawyer handles the legal clauses.

That's the short version. Here's how to actually do it well, because most people either trust the AI too much or feed it the wrong thing.

Why FDDs are a perfect fit for AI

Every FDD follows the same structure. The FTC Franchise Rule (16 CFR Part 436) requires franchisors to disclose the same 23 items in the same order, every year. Item 7 is always the investment table. Item 19 is always financial performance. Item 20 is always the outlet counts.

That standardization is exactly what makes AI useful here. The model isn't guessing where the numbers live. It knows. The work that eats your weekend (reading 300 pages of legal boilerplate to find six numbers) is the work AI is best at.

The problem is volume and repetition. A single FDD runs 200 to 400 pages. If you're comparing five brands, that's well over a thousand pages, and the important figures are buried in tables across different sections. Reading one is tedious. Reading five and holding them in your head is close to impossible.

The five items to pull first

Don't ask the AI to "summarize the FDD." You'll get a bland overview that smooths over the parts that matter. Ask for specific items.

FDD item	What to ask for	What it tells you
Item 7	Low and high total investment	The real cost to open, including working capital
Items 5 & 6	Franchise fee, royalty %, ad fund %	What you pay upfront and forever
Item 19	Average and median revenue/profit	Closest thing to an earnings figure, if disclosed
Item 20	Units opened, closed, terminated, transferred	Whether franchisees are staying or leaving
Items 3 & 4	Litigation count, bankruptcy history	Legal and financial trouble

Pull those five and you've covered roughly 80% of a first-pass screen. If a brand looks weak on any of them, you've saved yourself the deeper read.

A prompt that actually works

Vague prompts get vague answers. Try something like this:

"From this FDD, give me a table with: total investment low and high (Item 7), the initial franchise fee and royalty rate (Items 5 and 6), whether Item 19 discloses financial performance and the median figure if so, and the number of franchised outlets that closed or were terminated in the most recent year (Item 20). Quote the page number for each."

Asking for page numbers does two things. It lets you verify fast, and it forces the model to find the actual figure instead of inventing a plausible one. If it can't cite a page, treat the number as a guess.

Where AI gets FDDs wrong

This is the part the hype skips. A general chatbot reading a raw PDF makes predictable mistakes:

• It misreads tables. Item 7 and Item 19 are dense grids. Models sometimes grab the wrong row or merge two columns.
• It confuses averages and medians. Item 19 often reports both, and they can differ wildly. An average gets dragged up by a few top performers; the median is what a typical franchisee actually sees. Ask for the median.
• It treats a missing Item 19 as zero. Roughly a third of franchises don't disclose financial performance at all. That's a finding, not a blank, and a chatbot will sometimes paper over it.
• It hallucinates specifics. Ask for a number that isn't in the document and a model will often produce one anyway. This is why page citations matter.

The fix isn't to stop using AI. It's to feed it clean data. A model querying a database where the Item 19 figures were already extracted and validated will beat a model squinting at a PDF every time.

Summaries vs. real numbers

Here's the trap. Ask AI a general question about a franchise and you'll get a confident paragraph that reads well and commits to nothing. That's worse than useless for due diligence, because it feels like an answer.

Specific data gets you a decision. "Average gross revenue of $1.2M with a median of $840K, and 14 of 210 units closed last year" tells you something. "This is a well-established brand with strong unit economics" tells you nothing.

So push for figures, ranges, and counts. If the answer doesn't contain a number, ask again.

Research franchises without copy-pasting PDFs

The cleanest way to do all of this is to skip the PDF entirely. FranDB has already pulled the data from 2,488 FDD filings across 1,786 franchises: Item 7 investment ranges, Item 19 earnings for 1,695 brands, Item 20 outlet growth, litigation history, plus 87,000+ SBA loan records and 577,000+ franchisee contacts.

Connect that database to ChatGPT, Claude, or Cursor over our MCP integration and you ask questions in plain language: "compare total investment for these three brands," or "which had the most closures last year." The AI calls the data directly, so the numbers are extracted and checked, not guessed from a scanned document.

Where to stop and call a lawyer

AI is a screening tool. It is not your franchise attorney.

The legal clauses carry the most risk and the least room for a model to be "mostly right." Item 17 covers renewal, termination, transfer, and dispute resolution, the terms that decide whether you can sell your business or get pushed out of it. Item 12 covers territory. Item 8 covers what you're forced to buy and from whom. Have a franchise lawyer read those before you sign anything.

Use AI to decide which franchises deserve that lawyer's time. That's the right division of labor: the machine does the reading, you do the thinking, and a professional handles the contract.