My three-pass rule for analyzing financial data with Claude Cowork

The spreadsheet lands at 4:50pm. Twelve tabs, forty thousand rows of actuals, and a question from leadership that needs an answer by tomorrow morning. The old version of this evening involves pivot tables, VLOOKUP archaeology, and a growing suspicion that tab seven doesn’t tie to tab two.

The new version, the one I teach now, involves dropping the file into Claude Cowork and running a session with actual discipline. And that second part is the whole game. The difference between “an AI summarized my spreadsheet” and “an AI produced analysis I can stand behind” isn’t the model. It’s the process you wrap around it.

Treat it like a brilliant junior analyst

Here’s the mental model that works for me. Claude Cowork is a fast, tireless junior colleague. Brief it properly, give it the context a junior would need, and check its work the way a reviewer would. Skip the briefing or skip the review and you’ll get exactly what you’d get from an unbriefed, unreviewed junior: something plausible, fast, and occasionally wrong in ways that hurt.

Every one of my data sessions runs the same three passes. Not because I love process for its own sake, but because each pass produces something I can check before I build on it.

Pass one: prove it read the file

First instruction, every single time: load the workbook, list the tabs and their structure, and confirm the totals tie to the consolidation tab. Row counts, date ranges, control totals.

This feels bureaucratic and takes about a minute. It also eliminates the most dangerous failure mode in AI data work: confident analysis of misread data. If the load is wrong, everything downstream is wrong, and you won’t find out until the CFO does. I never let a session proceed past an unverified load, and Cowork never complains about being checked. Humans at 9pm do.

Pass two: decompose, and make it sum

Now the real question. Say EMEA margin is down versus budget and everyone wants to know why. My ask: bridge the variance into price, volume, mix, input costs, and FX, and show it as a table that sums exactly to the total.

That last clause matters more than anything else in the prompt. A decomposition that sums is auditable. You can spot-check any component against the source, and the arithmetic is on the page instead of behind a curtain. A bridge that ties is a bridge you can present. If you’ve read my PVM work, you know I’m strict about this: additive effects or it doesn’t ship.

Pass three: narrate, with receipts

Last ask: draft five sentences of commentary for the pack, where every number comes from the bridge, and flag anything you’re unsure about. Then I read it like a reviewer. Does each claim trace to a figure? Are the caveats real or decorative?

Total elapsed time on a real file of this shape: about twenty minutes, and most of that is my review. The same work by hand is an evening. And honestly, the output is better, not just faster, because the twenty minutes I spent were all judgment, and because Cowork checked ties I would have skipped at 9pm.

Keep a record, build a habit

Two habits close the loop. Save the brief and the key outputs next to the analysis, so when someone asks “where did this number come from?” three weeks later, the answer is a file and not a memory. And when the same brief works three months running, stop typing it. Package it, refine it, hand it to the team. That’s how a session becomes a workflow, which is the progression I teach in the AI FP&A Accelerator.

I filmed what that progression looks like in practice. I handed Cowork a full monthly close for a fake company, with errors planted on purpose: outstanding checks, a bank fee missing from the GL, a duplicate payment to the same vendor one day apart. Out of the box it caught every single one, which honestly impressed me. But the output was improvised. Functional, not professional. So I installed a plugin with my own close process, three skills that tell Claude exactly how I want the bank reconciliation, the AR and AP review, and the three-statement build done, and ran the same prompt again.

Same intelligence, completely different output: my format, my checks, every time. That’s what packaging your process buys you. It takes away the probabilistic character of AI and makes it closer to deterministic. You train it once, the way you’d train a new analyst, and it follows your steps from then on.

Try it on a file where you can check everything

The download below is my Cowork data-analysis starter pack: the three-pass session structure as ready-to-use briefs, the verification checklist, and a practice dataset with a known answer key. Build the habit on a file where every number can be checked. Then the 4:50pm spreadsheet stops being a threat.