Copilot in Excel is finally good, but you need to be more careful than ever.
A few months ago, I had largely written off Copilot in Excel. I remember asking it to add a formula with a hyperlink and it talked me through some lovely options, (even referring to some research on Diabetes!) before simply hard coding some text. In the end I just did it myself.
For a long time, that seemed to be the reality of it; a far cry from the AI teammate Microsoft seemed to be promoting.
That has changed, and I think the risk has changed with it. A tool nobody uses can't cause much damage. A tool that works (or at least looks like it works) poses a much greater risk.
The updates over the past few months have been substantial.
Copilot can now work through edits step by step and show its reasoning as it goes, rather than presenting you with a naive formula and a hopeful expression. You can switch between Chat Only, Plan, and Allow Editing to give it different levels of autonomy over making changes to your workbook. And it can clean your data, create a change log, and create an executive dashboard from a single prompt.
The results look promising. It can even tap into templates or existing documents to use your preferred styles. Messy data to presentation ready from a short conversation.
Personalisation for Copilot in Excel has started rolling out, too, as well as Excel-specific skills you can write and load into your model. Rules for your formatting and business logic.
Copilot can now produce workbooks with your data that look the part. And that can be dangerous.
If you’ve ever audited a model you’ll know how hard it is to spot an error.
As finance professionals, we’re trained to do this, but even so, we miss small details that can have material impacts. A simple mistake in signage, error management functions that hide a structural issue, a value that’s been hardcoded rather than calculated. Things we might spot when we’re building the model ourselves but miss as a reviewer.
Sometimes a well formatted model can trick us into a false sense of security. Copilot super charges this issue.
I think this is particularly pertinent for your average Excel user. People who aren’t finance professionals, but who begrudgingly use the tool because it’s the best one for the job. Forget XLOOKUP vs VLOOKUP, these are the colleagues who slowly and manually type corresponding labels between two lists while wishing there were a faster way. They don’t have the time to care about functions. And now there’s an invasive Clippy-esque colourful cloud promising to solve their problems.
“Hey Copilot, can you add a column that shows whether it’s a large, medium, or small sales client?”
A lot of the problems this brings will be like the problems we already face. Colleagues using unofficial definitions. Hard coded values. Sprawling workbooks with no auditable logic. But this time it comes with a robotic promise that they’ve done the right thing.
I wish this were purely hypothetical, but I’ve seen it in action. One of my clients recently, and proudly, showed me that they had completely replaced their VLOOKUP column with a Copilot prompt instead. A probabilistic solution for a deterministic problem!
Here's the shift I'd encourage every modeller to think about. The old issue was that Copilot was uselessly wrong. The new issue is Copilot is plausibly right.
A step-by-step edit that looks sensible can still rest on a misread of your data. A generated formula can be syntactically perfect and logically off. Copilot doesn't know that column F is stale, that the March figures were restated, or that your revenue definition changed at year end. It sees the grid. You know your business.
Even Microsoft is candid about the limits. The support page for the new =COPILOT function tells users not to rely on it for any task requiring accuracy or reproducibility, and to keep calculations in native functions like SUM and XLOOKUP. If that's the guidance for one function, it's a reasonable mindset for the whole feature.
So should you keep Copilot away from anything that matters? No. You should treat its output the way you'd treat work from a capable new junior: useful, but needing review.
You should keep this in mind when prompting Copilot, too. Tell it to use modern functions so it doesn’t use VLOOKUP. Ask it to avoid hardcoded values and to keep things dynamic. If using it to clean data, ask it to keep a log of any changes. Keep it traceable and auditable.
If you’re on the insider track, you have the opportunity to give Copilot in Excel its own personalisation. This is a great place to mention hardcoded values and the like, as you won’t need to include that instruction with every prompt.
With these guardrails in place, spinning up a report or model from raw data is quicker with Copilot. But don’t let its fancy formatting trick you into forgetting to review its work.
Some of the builld work is disappearing, and honestly, good riddance to a lot of it.
But the judgement hasn't gone anywhere. Knowing when a confident answer is wrong, structuring a workbook so AI edits can be reviewed, deciding which tasks are safe to delegate: that is modelling skill, applied one layer up.
Copilot in Excel is finally good enough to be dangerous. The modellers who build a review habit around it will get the speed without the risk. Those that learn to review well will see some serious upsides. But along with that, I suspect we’ll all be reviewing a lot of confidently wrong work, too.