How Were You Taught to Check a Model?

If you’re in finance, someone taught you how to check a model with a consistent sequence of steps: make sure the sums tie across tabs, make sure the statements agree with each other, make sure an assumption used in one place is the same one used everywhere else, etc.

Then, ask whether the numbers make sense for this business.

Several of the practitioners we’ve spoken with over the past few months were trained in exactly this way; it’s a real body of professional knowledge, and it’s why a good reviewer can move through a workbook so quickly. They know what order to look in, and they know what a problem usually looks like when they find one.

What were you actually taught?

The techniques break into roughly two groups, one mechanical (how things add up) and one semantic (what things actually mean).

The mechanical checks include questions like:

• Do the totals tie from tab to tab?
• Does the P&L flow through to the balance sheet and the cash flow statement intact?
• Does the growth rate on the assumptions sheet match the one in the revenue build?

Most teams will have already built these in as standing practice.

The semantic checks include questions like:

• Is that line really revenue, or is it annual recurring revenue with a generic “revenue” label?
• Does this margin look right?

On its own, no amount of raw arithmetic catches issues of the second kind, meaning that this part of the check has always come down to personal judgment.

What did the model-checking discipline assume?

Each of these groups assumed you knew how the model was built, usually because you built it yourself. If you didn’t, you could walk over to the person who did and ask, and by the time you reviewed everything, you already had the map you needed.

AI-assisted modeling takes that away. The financial model arrives finished, often in more versions than you can keep straight, and there’s nobody to ask about the thinking behind a given cell.

What this means is that the checks themselves might still be valid, but they’re sitting on top of assumptions that have broken.

What does checking cost you now?

The mechanical checks can still pass; an AI model, for example, will often hardcode an assumption in a way that makes the final figures work out even when the logic underneath is broken. This means the checks that count things come back clean, while the errors that matter would’ve only been caught by someone who knows the build.

So you rebuild the map yourself, cell by cell, starting from the detail and working up. That’s the inverse of how you were trained, and it’s a big part of why review takes as long as it does now.

Can the missing step be put back?

If what’s gone is knowing (in detail) how the model was built, the most relevant question is whether that knowledge can actually be written down in the first place.

We’re working through this ourselves. If you were taught to check a model differently or are experiencing other problems, we’d like to hear about it. Reach out via our Contact page, or join our Discord server to compare notes with our growing community of AI-first builders.

Frequently asked questions