What Changed That You Didn’t Ask For?

You asked for one thing: update the Q3 revenue build with new pricing information. The assistant does it, the numbers look right, the totals tie, so you happily move on.

That reviewing process was the right work, because you checked the changes you asked for.

But the trouble lies hidden somewhere else in the file. Sometimes a row you never mentioned moves too. Sometimes a settings tab you said not to touch comes back empty. In our conversations with AI-first financial professionals, the people who work in these files every day describe it the same way: the edit they wanted was fine, and something else changed alongside it.

Why is this its own kind of error?

Most worries about an AI agent’s output are centered on wrong answers, such as a formula that computes correctly but means the wrong thing or a figure with a hallucinated source document. Because those errors come from the request you made to the model, a careful review of the request can usually find them.

This kind of error is different; it’s less “I’m getting a bad answer to a question” and more “the AI agent is changing something nobody asked about in the first place.”

Why doesn’t review catch it?

Generally, a review follows the work. Most financial professionals will look where the edit was supposed to happen, because (naturally) that’s where mistakes usually are. For work you did yourself that’s a perfectly sensible approach, since you know what you touched.

But you’re less likely to check a cell that wasn’t supposed to be touched at all.

To illustrate, here’s a constructed example of what that can look like in a workbook:

• You ask for the opex forecast in rows 40 to 60 to follow the new headcount plan, and the assistant updates them.
• Trying to be helpful, it then concludes that a freight assumption in row 112 needs to be amended as well. Row 112 feeds into gross margin figures two tabs over.
• Every total still ties, because the change flowed through correctly, and your cross-tab checks pass.
• The forecast now rests on a freight number nobody chose.

The checks you were taught still work. They confirm the file agrees with itself. They can’t tell you whether it agrees with what you meant.

What does catching it cost?

The reliable answer is to look at absolutely everything: snapshot the file before the run, then compare it after. Plenty of people already do this, and it does work.

It also costs something every run, and the cost grows with the size of the file. A full comparison of a large model can return hundreds of differences, most of them the expected ripple of the change you asked for. This puts you back in the position you were in before: a list of 100 changes, with no guidance on where to direct scarce human attention.

Then, there’s frequency. If the assistant stays in scope most of the time, the habit of looking outside scope fades. That’s reasonable, given a clean record, but it’s also exactly when an out-of-scope change sneaks through.

What would tell you nothing else moved?

None of this is a problem unique to AI. A human colleague can paste over the wrong range and leave the file looking just as tidy. What’s changed today is how many edits you now hand off to an AI agent, and how little time there is to look past the edge of each one.

If this sounds familiar and frustrating, we’ve been building a solution called Lecxe. You can give it a try at that link, or schedule a time to talk to our team if you want to go deeper. We love hearing from people who are living this reality day-to-day, so don’t hesitate to reach out!

Frequently asked questions