What happens when AI opens your model

You ask an AI to add a downside case to your model and flag where free cash flow turns negative. A minute later it hands you a number. The question you should be asking isn't is the number right? — it's how did it get to that number? Because the answer changes whether you can put it in front of an investment committee.

The way an AI touches your file is invisible to you, but it decides three things you care about deeply: whether the model stayed live, whether the edits were safe, and whether you can show your work afterward. There are three ways it can touch a model. Only one gets all three right.

Way 1: it opens your Excel file the old way#

This is the AI doing what a junior might do with a macro — prying open the raw .xlsx and editing cells directly. It feels familiar because it's your file. It's also the shakiest option, for one reason above all:

The formulas go dead. An Excel file, opened this way, is a photograph of the last time someone hit "recalculate." The AI can change an input, but it can't actually see the ripple through your model — the dependent cells still show yesterday's answer until you reopen the file in Excel. So the AI is editing blind, and the number it reports may be from before its own change. You'd never accept that from an analyst. It's worth noticing when it comes from a machine.

On top of that, poking at cells by position is fragile. The AI has no idea that C7 means "2027 revenue, base case" — it's guessing from where things sit on the grid. Insert a row and its map is wrong. And when it saves, formatting, links, and charts can quietly break, with no record of what changed.

Useful for one thing only: getting old numbers out of a legacy spreadsheet. Not for doing the actual work.

Way 2: it scribbles directly on the underlying file#

Some newer formats store the model as readable text instead of a sealed binary — a .deepcell is one. That makes the file easy for an AI to read, which is a real step up: it can understand the structure of your model without guessing.

But letting it write by editing that text directly is like letting someone edit a contract by retyping it from memory. All the safeguards that make the model trustworthy — the checks that a formula points at a real line item, that a value respects its own limits, that the scenarios still hang together — live in the engine, not in the raw text. Skip the engine and you skip the safeguards. The AI can produce a file that looks fine and is quietly self-contradictory, and nothing will have flagged it.

Fine for reading. Risky for building.

Way 3: it works through a proper interface#

This is the one that gets all three right, because the AI never touches the raw file at all — it works through the same controlled interface a disciplined analyst would.

The model stays alive. When the AI asks "what's free cash flow in 2027 under downside?", the answer is recalculated on the spot, reflecting the change it just made. No stale photographs. The number it reports is the number the model actually produces.

Every edit is checked. Each change runs through the model's own rules before it's allowed to stick. A formula pointing at a line item that doesn't exist, a value that breaks a constraint you set — these come back as rejected, not as a silent error you find three weeks later in a board deck.

You can show your work. Every change the AI makes is recorded with a short note on why — a running history you can read back line by line. The line between what you built and what the AI built stays visible, which is exactly what lets you sign your name to the result.

The same task, three ways to trust it#

None of this is about the AI being smart. A brilliant analyst working from a dead photocopy, with no one checking their edits and no record of what they did, still can't be trusted with a committee model — not because they're wrong, but because you can't see whether they're right.

The fix isn't a smarter AI. It's giving it a live model, a set of rules it can't sidestep, and a history that writes itself. That's the whole idea behind a model that's a document, not a spreadsheet: when the AI opens it, the model stays alive, the guardrails stay on, and the work stays yours to sign.