A spreadsheet that's been running for years is a legacy codebase with no docs

Start by admitting it’s code

That cost-allocation sheet finance has been running for six years has 14 sheets, 200-odd columns, IF statements nested seven deep, and one cell where +0.03 is there because of a reason only the colleague who left in 2023 ever knew.

That isn’t a document. It’s a codebase with no documentation, no tests, no version control — running in production. And it’s worse than most legacy code, because code at least has syntactic structure. Excel flattens logic and data onto one plane, so you can’t tell at a glance which cell is an input, which is an intermediate, and which is the answer.

Which makes “put this manual Excel process on AI rails” fundamentally a reverse-engineering job. And reverse engineering always starts the same way: see clearly what you’re looking at.

That’s the step excel-ai-analyst (anp.asia) does.

The first obstacle isn’t compute, it’s trust

The first wall wasn’t technical.

You drag a real payroll sheet into a web page, it says “uploading for analysis” — and you hesitate for one second before pressing enter. That hesitation is correct, and it’s fatal: a tool that requires you to override your instincts before using it doesn’t get used.

So step 0 is deliberately pure front-end:

• everything runs in this page’s memory — no server ever receives your data
• it works with the network disconnected
• results vanish on refresh; download if you want to keep them, and the report stays on your device too
• the page makes no external requests at all — the engine is a function-by-function port of excel_ai.py, and the file-format layer is same-origin SheetJS

“Your file is never uploaded” only carries weight when you can pull the network cable and check.

(The site still recommends redacting first — names, ID numbers, phone numbers. The file doesn’t leave your browser, but it’s a good habit.)

What detection actually finds

Drop a sheet in and it answers four questions:

1. What shape the sheet is, and how many header rows
2. Which columns are inputs and which are results — the most useful one, because it finally separates data from logic
3. What the real formulas look like — the expressions, not the displayed values
4. Which cell has been edited by hand — a position that should hold a formula and now holds a hard-coded constant

Number 4 is my favourite. The dangerous thing in a business spreadsheet isn’t the complicated formula; it’s that at some point somebody deleted a formula and typed a number in to hit a deadline. It throws no error, affects no neighbouring cell, recomputes to the same wrong value every time, and just sits there.

Supports .xlsx / .xls / .csv, up to 20MB, multi-sheet workbooks analysed one sheet at a time. Reports download as Markdown, HTML or JSON.

After detection: why AI is still needed

One thing has to be stated plainly, and the site states it too: detection only tells you roughly what the sheet is.

Confirming that an AI genuinely understood the calculation logic requires recomputing every row against the full real dataset — 100% row-level pass rate and zero sheet-level anomalies, or it doesn’t count. That’s step 4 of the six-step method, it needs AI in the loop (writing the field ontology, formula chains and spec.json), a plain web page can’t do it, and the site doesn’t pretend otherwise.

The full six steps:

0 detect  ← what the web page does
1 structure to Markdown
2 field ontology
3 formula chains and lineage
4 full-data validation  ★ the actual test of "understood"
4b ontology graph
5 delivery and what-if
6 web dashboard

Pulling step 0 out into a page that needs no signup, no upload, and works offline has a simple rationale: it’s the only step that can be completed while you still don’t trust me at all. Use it to confirm the method can read your sheet, then decide whether the other five steps are worth it.

Known limits

Straight from the site:

• date columns are treated as numbers
• formula extraction from .xls is best-effort — save as .xlsx first for better results

No suitable sheet to hand? There’s a built-in demo file.