How To Use Copilot Finance Skills In Excel

Microsoft shipped Copilot finance skills into Excel in June 2026, and almost nobody is using them.

The skills sit behind a settings button most people have never clicked. Unless someone tells you they are there, you keep using Copilot the way you always have and never notice the difference. That is the normal fate of AI in Excel features, which is a shame given that the spreadsheet is still where most planning and reporting happens.

I spent a week with them on a real dataset: the P&L for all 7 locations of a coffee company, with a general ledger and manager commentary alongside it. Some of what Microsoft built is good. Some of it is built for a job most of us do not have.

Here is the split, and then the two skills I wrote myself to make the feature useful for a monthly close.

A skill is a markdown file

That is the whole thing, and it is worth sitting with, because the name suggests something much heavier. Markdown is a lightweight text format that strips out styling and leaves plain, structured text. You can open one in Notepad. AI tools read it easily, which is why it caught on as the format for this kind of instruction.

The file has three parts: a purpose, a set of rules, and an output format. No code, and no configuration screen to fight with.

Here is a working one, complete:

---
name: monthly-variance-review
description: Use when I ask for a variance review, a budget-to-actual, or a
  month-end variance summary of the P&L.
---

# Monthly variance review

## Rules
1. Compare Version = Actuals against Version = Budget for the month I name.
2. Report variance in dollars and as a percent of budget. Dollars first.
3. Flag anything over 10 percent OR over $25,000, whichever hits first.
4. Sort worst variance to best. Never sort alphabetically.
5. Ignore any location with no rows in the month. A location with no data
   had not opened yet. That is not a miss.
6. Never guess a cause. If nothing in the file explains a variance, write
   "cause not stated."

## Output
A table: Location, Account, Actual, Budget, Variance $, Variance %, Flag.
Then three sentences of commentary naming the largest single driver.

If you can write a handover note for whoever covers your close while you are out, you can write one of these.

Skill versus prompt

A prompt is a request. It says what you want done.

A skill carries the tactics. It says what you want done, how you want it done, and what the finished thing should look like. Steps, order of work, and the edge cases you already know about.

That difference is why the results stop drifting. When you retype a prompt from memory every month, you write it slightly differently every month, and the output moves with it. A skill runs the same way every time, on any machine, for anyone on your team who has the file.

Rules 5 and 6 in the file above are the ones I would point at. Every finance team has a location, a region or a product line that opened partway through a period, and every one of us has been handed a budget versus actual variance analysis that flags that opening month as a catastrophic miss against a full-month budget. You explain it every close, forever. Write it down once inside the file and you never explain it again.

Where the skills live in Copilot for Excel

Open Copilot in Excel from the ribbon, then open the settings menu inside the Copilot pane and choose Manage skills.

There are three categories.

Category	Who made it	What is in it
Custom skills	You	Anything you write, stored in OneDrive
Finance	Microsoft	Deal, valuation and modeling workflows
Visualization	Microsoft	Chart building and formatting

To call a skill in a prompt, type @ and pick it from the list. You can also toggle skills on and off in Manage skills, which controls whether Copilot reaches for them on its own. If you have already been using Agent Mode in Excel, skills sit alongside it rather than replacing it.

The Copilot finance skills Microsoft built

Open the finance category and you get a list of slugs rather than friendly names. On my machine in late August 2026 it read 3-statements, buyer-list, catalyst-calendar, check-model, comps-analysis, datapack-builder, dd-checklist, dd-meeting-prep, deal-screening, deal-sourcing, and more below the fold. Microsoft keeps adding, so treat any list you read as a snapshot.

Read that list again with your own job in mind.

Three-statement models, sure. Check model, fine. The rest is investment banking and corporate development. The two dd- entries are due diligence artifacts. Buyer lists and deal sourcing do not come up in a monthly close, and a catalyst calendar is an equity research thing.

I want to be fair about the one that travels, because comps analysis is not only a banker’s tool. Every finance pro I know has been asked how our margins stack up against comparable operators, and building sensitivities is ordinary FP&A work. So that is the one I ran.

Running comps analysis on a real P&L

I turned on comps-analysis, plus chart-design from the visualization category, and pointed Copilot at a P&L covering 7 locations with monthly revenue and cost of sales.

@comps-analysis We're a seven store specialty coffee chain. Using the actuals
in this workbook, build a sensitivity model on our revenue growth and cost of
sales ratio, then benchmark that against comparable operators.

Two things happened worth reporting.

First, it asked questions before it started. Which base period should anchor the sensitivity model, which peer set to use for the margin benchmark, and what the sensitivity grid should output. A tool that stops to ask three questions before running is a tool that has been told the work matters.

Second, it went and found real comparables on its own: Dutch Bros, CAVA, Sweetgreen and Shake Shack. It came back with a base case, key drivers, a sensitivity analysis, peer operating statistics, peer valuation and an implied valuation. Every figure carried a citation back to where it came from.

Cited output is checkable output, and checkable output is the only kind you can put in front of a CFO.

One caveat before you plan around it. Several of Microsoft’s finance skills lean on third-party market data through federated connectors, and those providers may want their own subscription. Microsoft’s launch post names LSEG, Moody’s, CB Insights, Daloopa, FactSet, Morningstar, PitchBook and S&P Global. If your company pays for none of them, expect some skills to reach a wall.

If you want the manual version of what this produced, sensitivity analysis in Excel is the same idea without the peer set.

The model picker

Inside the Copilot pane you choose which model runs the work. When I ran the comps analysis the list held Auto, GPT-5.6, Claude Opus 4.8, Claude Opus 5 and Claude Sonnet 5. I picked Opus 5.

For comparison work and multi-step analysis, the larger models earn their keep. For routine formatting or a quick lookup, a smaller one is fine and faster.

Two things worth knowing before you go looking for it. The picker needs a commercial Microsoft 365 Copilot subscription or Microsoft 365 Premium. And on an enterprise tenant, an admin has to enable Anthropic as a Microsoft sub-processor before any Claude model shows up at all, which is the most common reason the option is missing for people who otherwise have the license. If you are weighing the two engines against each other, I put Claude for Excel against Copilot on the same files.

What the list leaves out

Variance and close. Neither has a skill in the finance category.

Microsoft’s own launch post makes that stranger rather than clearer. Its sample prompts show @variance-analysis, @model-update and @portfolio-monitoring being called by name. None of those three appeared in my Manage skills dialog in late August 2026. The post also names closing the books and preparing a variance analysis as the kind of work skills are for, and then ships a list built around deals.

Maybe they are staged by tenant, maybe they land next month. Either way, the two processes my month runs on were not in the box, so I built them.

Building a custom skill, step by step

What you need before you start: a Microsoft 365 Copilot license, OneDrive, and a text editor. That is the full list.

Create the skills folder. In the Copilot pane, go to settings, then Manage skills, then Custom skills, then Create skills folder. Copilot builds it in OneDrive and hands you a button to open it. Everything you write lives there.

Create the skill. Each skill is a folder holding a file called SKILL.md. In OneDrive, click Create or upload, then scroll down to Markdown document. Most people have never noticed that option because they stop reading at Word and Excel.

Write the three parts. Name and description in the frontmatter at the top, rules in the middle, output format at the bottom. Keep the same shape across every skill you write, because consistency is what makes them easy to maintain 6 months later.

Match the name to the folder. The agent skills specification says the name in the frontmatter has to match the parent folder, lowercase letters, numbers and hyphens only. Get that wrong and the skill quietly does not load.

Turn it on. Back in Excel, go to Manage skills, then Custom skills, then hit Refresh. Your skill appears in the list. Toggle it on.

Call it. Type @ in the Copilot box, pick your skill, and give it a short instruction. Because the file already carries the rules, the prompt can be almost nothing:

@monthly-variance-review Run this for March 2025.

The output came back in seconds. Every location and account in one table, unfavorable variances sorted to the top, favourable at the bottom, and a short commentary naming the largest driver. Exactly the shape described in the file.

That formatting was specified rather than lucky, which is the part that matters. It will be the same next month.

Giving a skill context with a references folder

A skill that only carries rules is a well-organized prompt. The next step is where this turns into something you would hand a new analyst.

Alongside SKILL.md you can add a references folder and put documents inside it. The skill can then be told to follow the methodology in those documents instead of inventing its own.

For a rolling forecast I built a skill called forecast-roll with two files in references.

A company overview, so the tool understands the business it is forecasting: what the company does, how many locations, when each one opened.

A forecast guidance file, holding the methodology. How revenue is driven. What happens to a location that opened partway through a quarter. What the rule is when a manager files no commentary.

That second document is the real artifact of this whole exercise. Every finance team has those answers and almost none of them are written down anywhere except in one person’s head. 90 seconds of typing turns institutional knowledge into something a tool can follow and a new hire can read.

The prompt stayed short:

@forecast-roll Roll the forecast forward three months using the actuals in the
P&L and the manager commentary.

What came back covered revenue drivers, cost drivers with the manager comments folded in, a forecast by location for all 7 stores, an intercompany elimination section, and a driver, source and confidence rating for every line. The intercompany section was never asked for. It came from a rule in the file about excluding internal revenue from consolidated totals.

If you have done driver-based forecasting by hand, this is that, with the methodology written down somewhere the tool can read it.

Common issues

My skill does not appear after I add it

Hit Refresh in the Custom skills dialog. Copilot does not watch the OneDrive folder continuously, so a newly added or renamed skill stays invisible until you refresh. Microsoft’s own documentation calls this out, and it is still the first thing to try.

The skill ignores my reference documents

Check the folder name against what your skill file points at, character for character. I named a folder reference while the skill file said references, and the skill quietly ignored the whole folder. Nothing errored. It just did not use them, and that is the single easiest way to lose twenty minutes here.

Also worth knowing: a document in a references folder is available to the model, not obeyed by it. If the output looks like it used generic assumptions, ask the tool directly which methodology it applied.

Copilot will not pick up my skill on its own

Toggling a skill off in Manage skills stops Copilot reaching for it unprompted, but an @ mention still works. If you want a skill in play without naming it every time, leave it toggled on.

What skills do not do

They do not remove the review step. The forecast that came back was structurally right and I would still read every line before it went anywhere.

They do not fix a bad job choice. If you hand over work whose only check is redoing it yourself, a skill makes you wrong faster and more consistently.

And they do not carry your data. A skill carries instructions and reference documents. The numbers still come from the workbook in front of you.

Cost and access

Microsoft pitched the whole finance release at Microsoft 365 Copilot customers, and the Copilot pane on my machine labeled itself M365 Copilot (Premium) while the skills ran. The pre-built finance skills went generally available across Excel for web, Windows and Mac at launch. Custom skills started on the Insiders channel and rolled out afterwards, and partner-built skills were slated for Q3 2026, so what you see in your own dialog depends on your tenant and your update channel. CFO Dive covered the launch if you want the independent version.

Third-party data connectors are separate money, charged by the provider rather than Microsoft. Worth checking before you build a process around a skill that needs one. For the wider license question, what Copilot for Excel includes at each tier is the fuller answer.

Why the same file runs in Claude and ChatGPT

Microsoft did not invent a format for this. Skills use the open agent skills specification, the same format Claude skills use.

So the file you write for Excel opens in Claude with no edits. Same forecast methodology, same variance rules, same reference documents.

That changes the math on learning this. The ten minutes you spend writing your close process down is not a bet on Microsoft’s product plans. It is yours, and it moves with you when your company changes its mind about which AI it is buying this year.

The most common bottleneck in finance is that the knowledge lives inside one person’s head. Skills are the cheapest way anyone has had to get it out.

Pick the process you re-explain most often, open a markdown file, and write the rules down. Then go look at what else in your close can be automated once the rules are on paper.