To ensure reliability and to use it responsibly, avoid using COPILOT for:
- Numerical calculations: Use native Excel formulas (e.g., SUM, AVERAGE, IF) for any task requiring accuracy or reproducibility.
- Responses that require context other than the ranges provided: The COPILOT function only has access to the prompt and context provided to or referenced by the function. It does not have access to other data from your workbook, data from other files or enterprise information.
- Lookups based on data in your workbook: Use XLOOKUP to look up data based on a table or range.
- Tasks with legal, regulatory or compliance implications: Avoid using AI-generated outputs for financial reporting, legal documents, or other high-stakes scenarios.
- Recent or real-time data: The function is non-deterministic and may return different results on recalculation. Currently, the model's knowledge is limited to information before June 2024."[0]
Look forward to trying it, but can't say I'm hopeful from the screenshot. That "do a full analysis & find me insights" is the excel equivalent of hollywood's "zoom in, make it sharper" detective work.
To be fair, looking at a lot of financial analysis, there's a lot of vibes going on. Lots of significant figures are shown in reports and graphs and so on, but it strongly feels like everyone is stacking 5-20% tolerances everywhere and presenting the analysis as precise.
Copilot is so disappointing. Its marketed as this integrated AI but it has zero integration with any of the microsoft apps. Years later we are only just getting to scratch the surface of integration. Why cant copilot in teams actually interact with the chats its such a missed opportunity.
Some 3D CAD programs (FreeCAD, Fusion 360, others I assume) have this sort of confusing tree view for history+dependencies+state all-in-one. Since the face of one feature could be the origin of another feature, you need to have a pretty rigid order of operations that's fully explorable. The tree-view works great but UI wise for average folk, it's a cryptic use of that tree-view pattern.
There is SOMETHING there though, I think if you could show the flow of data in the same space as the data itself, you'd have a worthy Excel-like thing that does well with an AI agent jumping in to either fix something or attempt to irreparably break everything. Rather than looking at a diff grid of cells with hidden expressions, or a spaghetti mess of nodes with lines, I think a structured and shallow-nested tree would be easier to understand... maybe. Rather than expressions, you'd have features like in CAD that act on some base table, and you could jump back to the table as it was at a state where that feature was applied, make a tweak, and jump forward again.
Eg. Adding an item-specific tax calculation is a single operation in between the items + sum feature, rather than two operations where I have to hook up items => itemsWithTaxes, then tweak sum to read from each itemsWithTaxes. If I'm ever confused about what sum is doing while setting this up, I can go "back" in time to the state the sum feature works on, and I only see the data it has to work with.
I wonder if this applies?
"When NOT to use the COPILOT function
COPILOT uses AI and can give incorrect responses.
To ensure reliability and to use it responsibly, avoid using COPILOT for:
[0] support.microsoft.com/en-us/office/copilot-function-5849821b-755d-4030-a38b-9e20be0cbf62Look forward to trying it, but can't say I'm hopeful from the screenshot. That "do a full analysis & find me insights" is the excel equivalent of hollywood's "zoom in, make it sharper" detective work.
... enhance.
https://www.youtube.com/watch?v=EMBkWtDAPBY
My mind immediately went here:
Manager: "What's the rate of return according to our financial model?"
Analyst: "Let me vibe the answer for you. Just a sec."
To be fair, looking at a lot of financial analysis, there's a lot of vibes going on. Lots of significant figures are shown in reports and graphs and so on, but it strongly feels like everyone is stacking 5-20% tolerances everywhere and presenting the analysis as precise.
Lot of these sheets are easily very well manipulated.
> Microsoft says its Agent Mode in Excel has an accuracy rate of 57.2 percent in SpreadsheetBench
Amazing
Copilot is so disappointing. Its marketed as this integrated AI but it has zero integration with any of the microsoft apps. Years later we are only just getting to scratch the surface of integration. Why cant copilot in teams actually interact with the chats its such a missed opportunity.
The brave new world will be humans communicating through AI generated documents, which are never read and only are consumed as AI generated summaries.
What really is lacking here and what AI does not solve, is the problem of communication and the problem of work as time wasting anti-productivity.
Interesting feature, but I'm afraid of using it on Excel.
It's so easy to mess up on Excel, there are no crystal clear diffs like on Git. And even if that on Git we make mistakes.
Very often you have data that affects other data etc.
Also, we have DBs and other resources (even MS Access...) because we know how error-prone are Excel sheets.
If using AI adds another layer of potential problems with my Excel sheets, I'm not sure how using it would improve my overall work.
I believe they'd need to change the Excel quite a lot in order to make using AI 'excel'.
Maybe the new Excel is a new Excel-esque application with a new layer of tools to enable it to work properly with AI.
Like something that allows you to validate against a set of values described by prompt etc.
There are so many possibilities now with LLMs and I feel those integrations are very shallow, more of the same thing.
I don't need a Clippy on steroids. I want a new experience.
Some 3D CAD programs (FreeCAD, Fusion 360, others I assume) have this sort of confusing tree view for history+dependencies+state all-in-one. Since the face of one feature could be the origin of another feature, you need to have a pretty rigid order of operations that's fully explorable. The tree-view works great but UI wise for average folk, it's a cryptic use of that tree-view pattern.
There is SOMETHING there though, I think if you could show the flow of data in the same space as the data itself, you'd have a worthy Excel-like thing that does well with an AI agent jumping in to either fix something or attempt to irreparably break everything. Rather than looking at a diff grid of cells with hidden expressions, or a spaghetti mess of nodes with lines, I think a structured and shallow-nested tree would be easier to understand... maybe. Rather than expressions, you'd have features like in CAD that act on some base table, and you could jump back to the table as it was at a state where that feature was applied, make a tweak, and jump forward again.
Eg. Adding an item-specific tax calculation is a single operation in between the items + sum feature, rather than two operations where I have to hook up items => itemsWithTaxes, then tweak sum to read from each itemsWithTaxes. If I'm ever confused about what sum is doing while setting this up, I can go "back" in time to the state the sum feature works on, and I only see the data it has to work with.
As if Excel weren't bad by itself by failing to keep genomic data untouched...
They added an option to avoid gene conversions two years ago:
https://www.theverge.com/2023/10/21/23926585/microsoft-excel...
Source: https://www.microsoft.com/en-us/microsoft-365/blog/2025/09/2...
there goes middle management