Here's a mistake I see analysts make all the time. They open a company's investor presentation, they read it front to back, and they let the deck set the agenda. That's the trap. An investor deck is not research. It's a sales document. Every chart is cropped to flatter. Every metric is the one that looks best this quarter. The bridge charts always go up and to the right. The deck is engineered to persuade you, and if you read it the way it wants to be read, you end up underwriting management's narrative instead of testing it.
So let me show you how I flip that. I take the same PDF and I turn it into a red-flag matrix. Not a summary. A matrix that forces every claim in the deck to defend itself.
Here's the workflow. Step one. Drop the investor presentation PDF straight into the model. Forty, fifty slides, doesn't matter. No coding required. You're just uploading the file.
Step two is where the structure comes in. The prompt does the work. I tell it: go slide by slide. Pull out every quantitative or forward-looking claim management makes. For each claim, give me four columns. Column one, the exact claim, quoted, with the slide number. Column two, what evidence the deck itself provides for that claim. Column three, the source and date behind that evidence. Column four, the unstated assumption the claim depends on. Then flag it. Green if the claim is backed by a cited, dated, third-party source. Yellow if it's company-reported with no external verification. Red if there's no evidence at all, just an assertion.
Paste the prompt, wait about thirty seconds, and you get a table. That table is the red-flag matrix.
Now watch what it surfaces. The deck says, quote, "industry-leading retention." Slide twelve. Evidence column. Empty. Source column. None. Assumption column. That "industry-leading" is defined however management wants to define it. That's a red row. The deck says "record adjusted EBITDA." Evidence, a number. Source, company-reported, non-GAAP, with three add-backs in the footnote. That's a yellow row, and the matrix tells you exactly which add-backs to go check against the 10-K. The deck cites a third-party market sizing report by name, with a date. That's a green row. You can verify it.
In ten minutes you've separated what's verifiable from what's vibes. You haven't formed a view yet. You've just built a worklist.
Here's the part that matters most, and it's the part people skip. The matrix is the starting point, not the answer. Every red and yellow row is now a question you take somewhere real. The retention claim, you check the actual churn disclosure in the 10-K and the last four earnings call transcripts. The EBITDA add-backs, you pull the reconciliation table and decide for yourself whether those add-backs are recurring. The market sizing number, you read the cited report and check whether the company's definition of its addressable market matches the report's definition. The AI did not verify anything. It told you where to look. You verify.
And you put a date on it. The deck is from a specific quarter. The 10-K is from a specific filing date. The transcript is from a specific call. When a claim and its evidence come from different periods, that gap goes in the matrix too. A retention number from two years ago next to a "current" headline is a flag, not a fact.
A few heads-up notes so you run this cleanly. The model will sometimes paraphrase a claim instead of quoting it. Force exact quotes with slide numbers so you can always trace back to the source page. If the model can't find evidence in the deck, it should write "none stated," not invent a plausible source. That's a real failure mode and you control it with the prompt. And never let the green rows lull you. Green means "the deck cited something," not "the something is true." You still open the cited document.
Why this works. You've taken a document designed to lead you and you've turned it into a document that lists its own weak points. The persuasion is stripped out. What's left is a structured set of claims, each tagged by how well it's supported, each pointing you to the primary source where you'll actually do the analysis. The deck stops driving. You drive.
You can run this on any deck. Earnings presentations, IPO roadshow decks, activist short reports, the bull case a colleague forwarded you. Same four columns. Claim, evidence, source and date, assumption. Same color logic. It scales because it's not about the company, it's about how claims are supported.
One thing to be clear about. This is a research process, not a recommendation engine. The matrix doesn't tell you to do anything with the stock. It doesn't pick winners. It structures your skepticism so you spend your time on the claims that actually need testing, instead of reading the deck cover to cover and absorbing the story. The judgment is still yours. The model just organizes the interrogation.
So next time an investor presentation lands in your inbox, don't read it like the company wants you to. Drop the PDF in, run the matrix prompt, and let every claim defend itself. Quote, evidence, source, date, assumption, color. Then go verify the reds and yellows against the filings and the transcripts. Thirty seconds to build the worklist. The real work is yours.
This is educational only. It's not investment advice, it's not a recommendation, and AI structures the research process. It does not make the decision for you.