Here's the most common AI investing prompt I see. "What stock should I buy?" It's also the worst one you can type. Let me show you why, and what to do instead.
When you ask a model what to buy, you're asking it to do the one thing it cannot actually do. It doesn't know your mandate. It doesn't know your risk limits. It doesn't know what's already in the book. And it doesn't have today's filing in front of it unless you put it there. So it gives you a confident answer that sounds like research and is actually just a sentence shaped like advice. That's the failure mode. You get conviction with no source, no date, and no chain you can check. On the buy-side, that's not a shortcut. That's a liability.
Here's the reframe. AI doesn't pick the name. AI structures the process around the name. You're still the analyst. The model is the junior who reads fast and never gets tired. So stop asking it for a verdict and start asking it for a workflow.
Let me walk it. Say you're picking up coverage on a company. The bad prompt is "is this a buy." Delete that. The first better prompt is, "What should I read first to understand this business, and in what order?" Now the model isn't guessing a price. It's building you a reading list. The latest 10-K, the last four earnings call transcripts, the most recent 8-K, the proxy. You get a stack, not an opinion.
Second prompt. "What changed since the last filing?" This is the one that earns its keep. Paste in the new 10-Q next to the prior one and ask it to flag every material difference. Language that changed in the risk factors. A new line item in the segment breakout. A shift in how they describe the competitive set. Guidance that moved. The model surfaces the deltas in about thirty seconds. You still read them. But you're reading the ten things that moved instead of re-reading two hundred pages to find them.
Third prompt. "What are the three numbers that would break this thesis, and where do I find each one in the filings?" Now you're not asking it to agree with you. You're asking it to build your disconfirming checklist. Gross margin trend. Receivables growing faster than revenue. Stock comp as a share of operating cash flow. It tells you where to look. You go look.
Notice what every one of those prompts has in common. The model never tells you what to do. It tells you what to read, what moved, and what to test. That's the whole game. You're using it to compress the boring part of the work so you have more time for the judgment part, which is the part that's actually yours.
Now the risk control, because this is where people get sloppy. Four things have to be attached to every output before it goes anywhere near a memo. Source. Date. Assumption. Human review. Source means the model cites the exact filing and page, not "studies show." If it can't point to the document, you treat the claim as unverified, full stop. Date means you know whether you're looking at the current quarter or something stale, because a model will happily hand you a number from two years ago with total confidence. Assumption means every projection states what it's assuming out loud, so you can argue with it. And human review means you, the analyst, read the primary source before a single line of it lands in an IC memo. The model drafts. You verify. That order never flips.
Here's why this matters beyond just being safer. When you ask for a stock pick, you get one sentence you can't defend in a meeting. When you ask for a structured read of the filings, you get a comp table you can source, a delta list you can walk your PM through, and a disconfirming checklist that shows you did the work. One of those survives a room full of people poking holes in it. The other one doesn't.
And a heads-up on the thing that trips people up. The output always sounds finished. It's fluent, it's organized, it reads like a finished note. That fluency is exactly what makes it dangerous if you skip the verification step. Confidence is not accuracy. The model will state a wrong number in the same tone it states a right one. So the workflow isn't there to slow you down for its own sake. It's there so the polish never substitutes for the check.
So, to put it plainly. Don't ask the model what to buy. Ask it what to read first. Ask it what changed since the last filing. Ask it what would break your thesis and where to find it. Then attach source, date, assumption, and human review to everything before it moves forward. You stay the decision-maker. The AI just makes you faster at the research.
This is educational only. It's a research-process workflow, not investment advice. No buy, sell, or hold calls here, no price targets, no promised returns. Markets carry risk, the data needs verifying, and you should talk to a qualified professional before you act on anything. The model structures the work. You make the call.