How to research a company with AI
Last updated August 2026
Short answer
The difference between useful and dangerous here is whether the model is reading something in front of it or recalling something from training.
Start with the source
Get the latest 10-K from EDGAR or the company's investor relations page, along with the most recent 10-Q.
Give the model the document rather than asking about the company in the abstract.
Everything it then says is checkable against a text you also have, which is what makes the output trustworthy rather than plausible.
The questions worth asking
How does this business actually earn money, broken down by segment and geography.
What are the largest customer or supplier dependencies, and are any of them disclosed as concentrations.
What did the risk factors say last year that they do not say now, and what has been added.
Reading for change
A first pass on an unfamiliar company establishes what it does. Every pass after that is about difference.
Language that quietly softens around a product line, a debt maturity moving closer, a segment that stopped being broken out separately.
None of that appears in a headline, and all of it is public on the day the filing lands.
Try it in Walnut
Walnut connects to your brokerage and can answer questions about companies you actually hold, against their own filings rather than a summary of them.
Verifying as you go
Ask where each figure came from and open the page it names.
Cross-check anything numerical against the financial statements rather than the discussion, since the discussion is management's framing.
Treat an unsourced claim as a prompt to look rather than as information, which costs seconds and removes the main failure mode.
What it will not tell you
Whether the price is reasonable, which depends on expectations rather than on anything in the filing.
What competitors are about to do, which is the risk least visible in any company's own disclosure.
What management privately believes, as distinct from what they were willing to write down and sign.
A worked sequence
Ask for a plain summary of how the business earns money, then check it against Item 1 of the filing.
Ask what the risk factors say now that they did not say last year, then read the two versions side by side.
Ask for the debt maturity schedule and the gap between net income and operating cash flow, then verify both in the statements and the notes.
Prompts that produce checkable answers
Name the section and page for every claim you make.
Quote the sentence you are relying on rather than paraphrasing it.
Tell me what this filing does not say about the thing I asked, which is the question that catches confident invention.
Sources
Filings are published by the SEC through EDGAR full-text search, with guidance on researching investments at investor.gov. Walnut is informational and is not an investment adviser. This guide is educational and not personalized investment advice.
FAQ
What is AI actually good for when researching a company?
+
Reading. Working through a 10-K, summarising how the business earns money, comparing this year's language against last year's, and answering specific questions against the document. All of it is checkable against the filing.
What should I not use it for?
+
Price targets, buy or sell conclusions, and any figure you have not verified. A confident valuation with no stated assumptions is a number rather than an argument, and the assumptions are where the disagreement lives.
What is the most useful single question?
+
What changed since the last annual report. New or reworded risk factors, softened language about a product line, a customer concentration that grew: those are signals no headline carries and no ratio shows.
How do I avoid being told something wrong?
+
Ask for the source of every claim and open it. A model reading a document you can also read is a very different situation from one recalling something it absorbed in training, and the difference is entirely in whether you check.
Where do I get the filings?
+
EDGAR full-text search on the SEC's website, free, covering every filing a US public company makes. The company's investor relations page carries the same documents alongside presentations and transcripts.
Can it read the earnings call?
+
Transcripts are text, so yes. The question and answer section is the valuable part, because analysts press on what the prepared remarks avoided, and summarising the evasions is a reasonable task to delegate.
Does this replace an analyst?
+
It replaces the reading an analyst does, not the judgment. What it gives an individual investor is the ability to work through material that would otherwise take an evening, which is where the gap with institutions was always largest.
How long should this take?
+
An hour or two for a company you do not know, and considerably less for one you follow, because subsequent years are read for what changed rather than from nothing.
What prompts work best?
+
Ones that force verification: name the section and page for every claim, quote the sentence you are relying on rather than paraphrasing, and tell me what this filing does not say about the thing I asked. The last one catches confident invention better than anything else.
Should I trust a summary without opening the filing?
+
No, and the discipline is what makes the whole method work. A model reading a document you can also read is checkable; one recalling something from training is not, and the two produce output that looks identical.