Strip away the ceremony and most FP&A work is one thing: answering questions. Why did opex jump? What’s driving the margin miss? What did we assume for volumes in the March case?
So it’s worth being precise about what answering a question actually involves. Take the bread-and-butter example. A variance lands and someone asks why. The work is:
- Find the appropriate spreadsheet(s). Which file has this cut? Which vintage? Is the answer split across the cost model and the board pack?
- Open, understand, and navigate. Twelve tabs, merged headers, a summary sheet that almost-but-not-quite ties to the detail.
- Find the section and make sense of it. Locate the line, check the version, confirm the sign convention, trace the total back to its parts.
- Write the analysis. What moved, why, what it means, what to do about it.
Step 4 is analysis. Steps 1–3 are retrieval, and they routinely take longer than step 4.
flowchart TB
Q[A variance lands. Someone asks why.] --> S1[1. Find the right files]
S1 --> S2[2. Open and navigate them]
S2 --> S3[3. Locate the figure, check the version]
S3 --> S4[4. Write the analysis]
S1 -.-> R[Retrieval: most of the week]
S2 -.-> R
S3 -.-> R
S4 -.-> J[Judgment: the job they were hired for]We pay analysts to do step 4. They spend their week in steps 1–3.
Nobody hires an FP&A analyst for their gift at locating numbers. They’re hired for judgment, the thing that happens in step 4. But because the first three steps are unavoidable and time-boxed by deadlines, they crowd out the fourth. The analysis that gets written is the analysis there was time left to write.
This is also why “just work faster” and “better file naming” never fix it. The retrieval tax isn’t a discipline problem. It’s structural. The answers live across hundreds of workbooks and decks, produced by different people, in formats optimised for presentation, not for being asked questions later.
Which steps should a machine do?
Steps 1–3 are exactly the shape of work machines are now good at, if the machine is built for it. That qualifier is doing a lot of work, and it’s where most “AI for finance” attempts fall over. A spreadsheet’s meaning lives in its structure. The row, the column header, the tab, the version. A tool that reads it as flattened text will fetch you something plausible rather than something correct. (That failure mode deserves its own post.)
But get retrieval right, meaning position-aware, version-aware, with the source one click away, and the economics of the team change. The question that cost an afternoon costs a minute, plus ten seconds to verify the citation. Step 4 gets the time it deserved all along.
The test to apply
If you’re evaluating anything that claims to speed up analysis, ignore the demo gloss and ask one question: which of the four steps does it actually do?
- If it only finds documents, it’s a search box. You’re still doing steps 2 and 3.
- If it answers but can’t show the exact cell it came from, it’s added a fifth step: checking whether the machine is right, usually by doing steps 1–3 yourself anyway.
- If it does steps 1–3 verifiably and hands you a cited number, it’s giving your analysts their week back for the one step that was ever really theirs.
This is part of a series explaining AI and the systems around it for finance people, in their own language. I build AI systems for finance teams; the series is what I’ve learned doing it.