Here's the mistake. A colleague quits and drops a spreadsheet on your desk. Zero context. No notes. Just columns. Right now your understanding of that data is zero percent. And most of us were never taught how to fix that in a structured way. So we either grind through it by hand for hours and still miss things, or we paste it into AI and ask what it means. Speed without a method is just faster confusion.
Here's how to do it right. Three steps. Describe the data. Interrogate the data. Then point it at a goal. I'm going to move that loop onto a research desk. Earnings files. Segment data. Comps tabs. Raw screens. I'm not telling you what to conclude. I'm giving you a method, and a discipline that keeps it honest.
Step one. Describe. Start by making the data explain itself. Upload the file and ask the AI to list every column and show a sample from each one. That forces it to actually look at the whole file. And it gives you a readable overview instead of a wall of cells. For us, that's the first pass on an unfamiliar 10-K exhibit or a vendor's data dump. You're not analyzing yet. You're orienting.
Then run a data-quality pass before you trust anything. Ask it to flag missing values, bad formats, and outliers. In the demo, one column was missing 99.7% of its values. Nearly the entire field was blank. That single check killed a whole branch of analysis before it wasted a day. On a research desk, this is the blank segment column. Or the comps tab where half the multiples never populated. Find the holes first. They decide what you can and cannot say.
Step two. Interrogate. Have the AI brainstorm the questions this data could actually answer. Good questions mean it understands the file. Weak questions mean it doesn't, and you fix that before going further. Then ask the harder one. What would someone want to ask that this data cannot answer, because the fields simply aren't there. That list of gaps is the honest part. It's how you manage expectations before your boss asks for the impossible.
Let me make this concrete. You're handed a company's segment file ahead of earnings. Describe step. The AI maps every column and shows you samples. Quality step. It flags that the regional breakout is mostly empty. So geographic analysis is off the table. You know that on minute one, not hour three. Interrogate step. It tells you the file can show segment mix over time. But it cannot show margin by segment, because cost data is missing. Now you know exactly what you can defend in a memo.
Step three. This is the one people skip. Set the goal. Imagine building twenty beautiful slides, and your boss says, I just wanted to know one thing. That's what happens when you analyze without a goal. Technically correct. Ultimately useless. So brief the AI like you'd brief a junior analyst. Tell it the single decision you're trying to support. Then ask which parts of the data matter for that goal, and which to ignore. It will prioritize. It'll even sketch a sequence. Clean the data. Score the categories. Rank them. Then stress-test the outliers.
Here's the line I want you to hold onto. The creator was clear the AI was not doing all of the work. It was making the human analyst's job easier. That distinction is everything in finance. The AI can structure, sample, and surface. It cannot own the assumption that a number is right. And it can be confidently wrong. So treat every output as a draft.
So we add one step the tutorial implies, and we make it explicit. Before any figure leaves your desk, trace it to the primary filing. The AI says revenue grew. You open the actual statement and confirm it. The AI joins two datasets on an ID field. You spot-check the matched rows yourself. The merge is convenient. The responsibility is still yours. A pretty merged table that nobody verified is the dangerous one.
So here's the honest takeaway. Describe the data so you understand it. Interrogate it so you know its limits. Set a goal so the work answers a real question. And verify every number against the source before you rely on it. The loop turns hours of blind exploration into minutes of structured reading. But it does not move the judgment off your shoulders. The AI builds the analysis. You own the assumptions and the read.
If you want the prompt sequence for this loop, comment the word DIG. I'll send the analyst version of the describe, interrogate, goal workflow. Subscribe if you want one AI research workflow each week. And read the full disclaimer below before you apply any of this. This is educational only. Not financial, legal, tax, or investment advice. AI can hallucinate, so verify against primary sources. The analysis is fast. Your verification is the standard.