Copilot in Excel for Finance: What It Does and What Juniors Still Check

Copilot in Excel is Microsoft's AI assistant inside the spreadsheet. Describe a task in plain English and it writes the formula, builds the chart or the PivotTable, or plans and builds a whole workbook on its own, and since June 2026 Microsoft has shipped it with ready-made skills for finance work such as a DCF or a variance analysis.

For a junior in finance, that changes the order of the job more than the job itself. The first draft of a model can come from a prompt. What stays with the analyst is saying precisely what the model should do, then checking every cell of what comes back, because Copilot is still wrong often enough that its work cannot be trusted unread.

There is a catch that has nothing to do with the AI. At a bank, Copilot in Excel arrives with a paid Microsoft 365 Copilot seat, and it is a different product from GitHub Copilot, the coding assistant the same bank gives its engineers.

What is Copilot in Excel?

At work, Copilot in Excel is part of Microsoft 365 Copilot; at home it comes with Microsoft's consumer plans. It opens from the Copilot icon in the lower-right corner of the workbook, runs in Excel on Windows, Mac and the web, and works in three modes: chat answers questions without touching the sheet, edit changes the workbook directly, and plan sets out its changes before making any.

The name causes more confusion than the product. Microsoft sells several things called Copilot, and a finance junior meets one of them at a desk.

What Copilot in Excel changes for a junior analyst

Of all those Copilots, this is the one that changes a junior's day, and the change is in where the hours go. Building the first version of a workbook, the formulas, the layout and the chart, used to be most of the work. With Copilot the first version can come from a prompt, and the hours move to the two ends of the task: specifying what the workbook must do, and checking what comes back.

Microsoft's own example of the new work is a quarter-end review. Its launch post for the finance features suggests asking Copilot to "compare last quarter's actuals to plan," then to "identify the five largest revenue, expense, margin, and cash flow variances, explain likely drivers, and draft an executive-ready business review summary."

Picture what that leaves for a first-year analyst. Copilot drafts the variance table and the commentary in one pass. The analyst then confirms three things no prompt can: that each variance pulls from the right actuals and the right version of the plan, that a cost line coming in under budget is marked favorable rather than flagged like a revenue miss, and that the drivers Copilot names match what the business actually did that quarter.

Two habits separate an analyst who uses Copilot well from one who is used by it:

• Specify it the way a reviewer would. Name the ranges, the assumptions, the conventions and the shape of the output, instead of asking for "an analysis."
• Check the numbers, not the story. Trace each figure back to its source rather than trusting the summary written around it.

It is the same move from making the first draft to checking it that banks describe for their in-house tools, and it is why the analyst job is changing rather than disappearing.

What can Copilot in Excel do in a finance workflow?

What Copilot can take off an analyst's plate has grown in three layers, each built on the last.

Formulas, clean-up and charts

The first layer is the one most people have seen. At general availability in September 2024, Microsoft added support for formulas such as XLOOKUP and SUMIF, conditional formatting, charts and PivotTables, and let Copilot work with text as well as numbers, which matters to anyone handed a column of company names spelled five ways.

Copilot in Excel with Python, previewed at the same time, lets a user run forecasting or risk analysis in plain English, with no coding required. Copilot can also fetch information from the web and search a user's work files for context, so a question about a company can draw on recent news or an internal document.

Agent Mode: whole workbooks from a prompt

The second layer is Agent Mode. Instead of answering one request, it plans a sequence of steps, carries them out in the workbook, checks the result and repeats until the outcome is verified. It can also run what-if scenarios on revenue, budgets or forecasts.

Microsoft's own demonstration prompts were ordinary finance work: a monthly close report for a small business with product-line and year-over-year breakdowns, and a loan calculator with a full amortization schedule.

Finance skills and data connectors

The third layer is aimed squarely at finance. Skills are saved instructions that walk Copilot through a repeatable process the same way every time. Microsoft's examples are building a DCF, closing the books, refreshing a monthly reporting model and preparing a variance analysis, and teams can write their own. Workbook rules store a firm's naming and formula conventions inside the file, so they travel with it.

Connectors bring data from outside providers straight into the workbook, from FactSet, PitchBook, S&P Global and LSEG among others. Microsoft has also partnered with the Financial Modeling Institute, which credentials professional modelers, and says the institute's library of real modeling cases is now "a foundational part of how we evaluate Copilot in Excel for finance work."

Here is how those layers map onto the work a junior is actually handed:

Where does Copilot in Excel go wrong?

Every row of that table ends in a check because Copilot still gets things wrong, and Microsoft is unusually plain about it. The product's help page says Copilot "can sometimes make mistakes, misinterpret information, or produce inaccurate results," and adds a line worth reading closely in finance: "Avoid using Copilot for decisions in sensitive areas such as finance, legal, or medical topics." The operative word is decisions. Copilot drafts; a person decides.

The same page warns that Copilot explains its formulas "in fluent, grammatically correct language, but the content can be inaccurate."

What the benchmark shows

The yardstick Microsoft itself reports against is SpreadsheetBench, a set of 912 real questions collected from online Excel forums. When its authors first ran it in 2024, Copilot in Excel solved 20% of a 50-question sample, and a panel of four Excel experts solved about 71% of a 50-question subset. In 2025 Microsoft ran Agent Mode on all 912 questions and reported 57.2%.

Copilot in Excel, 2024 (50-task sample)

20%

Agent Mode, 2025 (all 912 tasks)

57.2%

Excel experts, 2024 (50-task sample)

71.3%ref

Three separate tests, not one race. The 2024 bars are the benchmark authors' own runs on 50-task samples, with the experts drawn as a reference point; the 2025 bar is Microsoft's run of Agent Mode on all 912 tasks, graded with the authors' script.

Read the 57.2% for what it is. Microsoft's engineers say SpreadsheetBench does not capture much of what makes Excel powerful, including PivotTables, charts and formatting, and that Agent Mode still has "plenty of room for improvement, particularly around things like formatting and presentation-worthy layouts." A 57% score on forum questions is not a 43% error rate on a bank's model. It does say that on a hard, unfamiliar task, a wrong first draft is common rather than rare.

The benchmark's authors found a second weakness that matters more in finance than on a forum. They also graded each answer against versions of the spreadsheet holding different values, and the models' scores fell: formulas that worked on one set of numbers did not always survive another. Even the experts slipped, from about 71% to 62%. A model gets rolled forward every quarter, so a formula that only works on this quarter's data is a liability.

Built to be checked

Microsoft's response is to make the work auditable rather than to promise it is right. Agent Mode writes small tests before it acts and keeps every calculation as a live formula on the grid instead of pasting in hardcoded values, so a reviewer can trace each number.

Plan with Copilot lists the ranges, worksheets, formulas and assumptions Copilot intends to change before it touches them, and each edit it makes is attributed to Copilot in the Show Changes pane, beside colleagues' edits.

Two practical limits apply. Copilot edits only when calculation is set to automatic, and once its changes are saved, anyone with access to the workbook, co-authors included, can see them.

Checking a Copilot-built DCF: a worked example

Here is what that checking looks like on a valuation a junior is likely to be handed. Say you ask Copilot for a five-year DCF of a company whose free cash flow runs $80m, $85m, $90m, $95m and $100m, with a 9% discount rate and 2.5% terminal growth. The model comes back clean: the five years are worth about $346.7m today and the terminal value about $1,024.9m, for an enterprise value near $1,371.6m.

Three checks, with the arithmetic:

1. The terminal value formula. Gordon growth needs next year's cash flow: 100 × 1.025 ÷ (0.09 − 0.025) = 102.5 ÷ 0.065 = $1,576.9m before discounting. A formula that divides the final year's $100m instead gives $1,538.5m, $38.5m or 2.4% short, which takes about $25m off the enterprise value. Nothing on the sheet looks broken.
2. The growth input. Type 3.5% instead of 2.5% and the terminal value becomes 103.5 ÷ 0.055 = $1,881.8m, 19.3% higher, from one keystroke. Every hardcoded assumption needs checking against the source it came from.
3. The discounting convention. Discounting that terminal value over five full years (1.09 to the fifth power is 1.5386) gives $1,024.9m; over four and a half years (1.4737) it gives $1,070.0m. Neither is wrong in principle, but that $45m gap has to match the convention your team uses.

The same discipline applies to the sensitivity table Copilot builds beside the model: change one input by hand and confirm the table moves with it. A table that stays still has been pasted in as values.

The audit trail Microsoft built in is the place to start all three: read the plan before Copilot edits, then walk the Show Changes pane cell by cell.

Who uses Copilot in Excel, and at what scale?

All of that assumes the analyst has Copilot at all, and that is a purchasing decision the firm made long before the analyst arrived.

What an employer has to buy

Inside a company, Copilot in Excel depends on a Microsoft 365 Copilot license the employer buys seat by seat on top of its Microsoft 365 plan. Computerworld put the price at $30 per user per month for enterprises in March 2026.

Since the spring of 2026, staff at organizations with more than 2,000 users have needed that seat to use Copilot inside Excel at all; smaller organizations keep a more limited standard access without it. Any bank of real size sits far above that line, so some banks have bought seats by the tens of thousands and others have built assistants of their own.

Which banks have bought it

Microsoft counts Copilot in paid seats of Microsoft 365 Copilot, and the count has grown fast: by July 2026, Microsoft said, net additions were more than doubling from one quarter to the next.

January 2026

15m

April 2026

20m+

July 2026

30m+

From Microsoft's quarterly earnings calls. One seat covers every Microsoft 365 app, not only Excel.

Banks are among the largest buyers on record, most of them named by Microsoft itself:

Read the fine print before repeating any of these as Excel users. Microsoft does not break out how many seats are used in Excel, and its announcement of the Barclays rollout talked about booking travel, checking policy and answering HR questions rather than building models. The finance team Microsoft describes running Copilot in Excel day to day is its own, across FP&A, accounting, tax, compliance and treasury.

What else Bank of America's bankers use alongside Copilot is covered on the Citi and Bank of America page.

Where Copilot in Excel came from, and where it is going

Those seats buy a product that is only a few years old and has already been rebuilt several times.

From a sidebar to an agent

The pattern in those dates is the story. In under two years of general availability, Microsoft moved from a sidebar that suggested formulas, to an agent that builds the workbook, to a version tuned for finance work.

What happened to the COPILOT function?

One experiment along the way did not last. In August 2025 Microsoft put a worksheet function, =COPILOT(), into beta: it took a prompt and a cell range and returned an AI answer inside the cell. Microsoft retired it on September 14, 2026, while it was still a preview feature, and points users to the Copilot pane, which covers many of the same tasks.

Whose model is doing the work

Copilot in Excel is no longer tied to one AI lab. It launched on OpenAI's models, and Agent Mode was built on OpenAI's reasoning models. A switcher now lets users pick Anthropic's Claude instead, or leave the choice on Auto, and the choice lasts only for the session: close Excel and Copilot goes back to its default model.

Anthropic's models are hosted outside Microsoft-managed environments, and at a company an administrator has to enable Anthropic as a Microsoft sub-processor before anyone can use Claude.

By July 2026 Microsoft had its own model in the mix. Chief executive Satya Nadella told investors that in Excel, MAI-Code-1-Flash was "delivering comparable quality to GPT-5.6 for the most common tasks, while operating at significantly lower cost."

Anthropic also sells around Microsoft. It previewed its own Claude add-in for Excel in October 2025 and made it generally available in May 2026, so the AI inside a firm's spreadsheets need not be Microsoft's. Banks that built their own assistants, such as JPMorgan's LLM Suite and Goldman Sachs's GS AI Assistant, are a third route.

What to know about Copilot in Excel for an interview

Knowing which of those routes a bank has taken helps, but Copilot in Excel is not a topic to recite. It is a concrete way to show judgment about AI in the work the job actually involves, and a strong answer draws on four things:

1. The right product. Mixing up Copilot in Excel with GitHub Copilot or the free Copilot Chat is the fastest way to sound secondhand.
2. What it does in a finance task. Formulas and clean-up, agent-built workbooks and the finance layer each touch a different part of a junior's week; know which is which.
3. How it goes wrong. Its errors are fluent and often silent, and Microsoft's own benchmark run solved a little over half of a hard set of real forum problems. Saying that plainly shows you have thought about it seriously.
4. Where it is heading. Microsoft is teaching it a team's own conventions, grading it against professional modeling cases and running it on models from more than one lab. The part of the job that keeps its value is judging whether the output is right.

How to practice with Copilot before your first day

Unlike a bank's in-house assistant, Copilot in Excel is something a student can use at home. Microsoft 365 Personal and Family subscribers can get it with an AI credits plan, and Microsoft 365 Premium includes it. Three habits turn that access into something worth talking about:

1. Build a small model by hand first, then ask Copilot to build the same thing from one prompt, and list every difference between the two.
2. Use plan mode before edit mode, and read the plan the way a reviewer would.
3. Keep a log of every error you catch: the prompt, what came back, what was wrong and how you found it.

Once those habits are in place, the interview readiness tool is a structured way to rehearse the answers, and AI in investment banking covers the bank-built assistants a candidate may meet instead of Copilot.

The bottom line

Copilot in Excel does not change what a finance junior is answerable for; it changes where the hours go. The formulas, the chart and the first pass at a model can now come from a prompt.

What cannot come from a prompt is the judgment that a number is right. Microsoft built the product to be audited and tells its users to check what it makes. Know which Copilot your employer pays for, know what it does and where it slips, and practice catching it, and you are ready for the version of the job that already exists.