Building a Financial Model in Excel With Copilot You Can Actually Trust
A financial model earns trust one verified cell at a time, and loses it the first time someone finds a number that's wrong. That makes Excel a genuinely strange place to use an AI tool that can write a formula fast and read a range wrong with the exact same confidence. Copilot in Excel is worth using for a model, but the way you use it has to account for that trade-off directly, not hope it doesn't come up. If you haven't seen the general spreadsheet workflow yet, Analyzing a Real Spreadsheet With Copilot in Excel covers the verification habits this article builds on for the specific case of a model, not just a one-off analysis. The Beginner's Guide to Copilot covers the fundamentals if you're new to the tool entirely.
One setup note before starting: Copilot in Excel tends to work best on clean, consistently structured data rather than a sheet full of merged headers and inline subtotals, and it has its own current file and storage requirements. Check Excel's own Copilot help before you start if anything about your setup seems to be blocking it, since the exact requirements are the kind of detail Microsoft updates from time to time.
What Copilot is actually good at in a model
Copilot's real strength in a financial model isn't calculating the number, it's structuring the sheet so the calculation is correct once and reusable everywhere. Given a clear description of what a model needs to do, it can build the skeleton, the labeled sections, the formulas that reference each other correctly, and the formatting that makes a model readable by someone other than the person who built it, faster than doing that scaffolding by hand.
Build a simple three-year revenue model for a subscription business. Rows for: starting customers, new customers added per month, churned customers per month, ending customers, average revenue per customer per month, and total monthly revenue. Twelve columns per year, three years, with a running total column. Use formulas that reference the prior month's ending customer count, don't hardcode the customer numbers month to month.
”The instruction not to hardcode numbers month to month is the important part. A model where every cell is a live formula referencing the assumptions above it is one you can stress-test by changing a single input. A model where Copilot typed in plausible-looking numbers for each month is a document that looks like a model but doesn't behave like one, because changing your churn assumption in January does nothing to February through December.
Here's what the first four months of that structure look like, with entirely hypothetical sample numbers standing in for whatever a real business would actually have:
A hypothetical subscription model, Year 1, months 1 through 4, illustrated with sample numbers only
| A | B | C | D | E | F | G | |
|---|---|---|---|---|---|---|---|
| 1 | Month | Starting Customers | New Adds | Churned | Ending Customers | ARPU | Monthly Revenue |
| 2 | Jan | 120 | 18 | -6 | 132 | $49.00 | $6,468.00 |
| 3 | Feb | 132 | 20 | -7 | 145 | $49.00 | $7,105.00 |
| 4 | Mar | 145 | 17 | -8 | 154 | $49.00 | $7,546.00 |
| 5 | Apr | 154 | 19 | -9 | 164 | $49.00 | $8,036.00 |
The detail that makes this a model and not just a table is invisible in the grid above: February's "Starting Customers" cell isn't the number 132 typed in, it's a formula pointing at January's "Ending Customers" cell. Change January's churn assumption and every month after it recalculates on its own. That's the entire point of asking for live references instead of letting Copilot fill in numbers that merely look consistent.
Where the arithmetic itself can't be trusted blind
Copilot in Excel is generally accurate, and it's also handling a lot of structure at once: cell references, ranges, filters, and formula logic across dozens or hundreds of cells in a single agent-mode action. That's exactly the kind of task where a single off-by-one range or a formula copied down incorrectly produces a wrong number that still looks completely plausible, because it's close to the right shape.
A wrong number and a right number look identical in a model
Unlike a chat answer, a wrong formula in a financial model doesn't announce itself. It sits in a cell, gets referenced by five other cells, and shows up as a confident, specific figure in a board deck three weeks later. Treat every formula-heavy section Copilot builds as a first draft that needs the same scrutiny you'd give a junior analyst's first pass, not a finished calculation.
- 1
Read the formula, not just the output number
Click into a handful of cells, especially ones that reference several other cells or ranges, and check that the formula does what you'd expect. A SUMIFS pulling from the wrong column, or a growth rate applied to the wrong base month, produces a number that's wrong but not obviously wrong.
- 2
Rebuild one section by hand as a check
Pick one row, like month-twelve ending customers, and calculate it manually from the assumptions: starting customers plus new, minus churned, carried forward month by month. If your hand calculation matches the model's, that's real evidence the formula logic holds across the sheet.
- 3
Stress-test with an extreme input
Change churn to an unrealistic number, like 50% monthly, and confirm the model's ending customer count actually drops the way it should. A model that doesn't visibly react to an extreme input usually has a broken reference somewhere in the chain.
A worked example: catching a real error
Say the model above comes back and month-twelve ending customers for year one look flat compared to month eleven, even though the assumptions call for continued net growth. That flatness is worth investigating before moving on, not explaining away as "probably a slow month." Ask Copilot directly:
Show me the exact formula in the ending customers cell for month 12, and the formula for month 11. Explain what each one is referencing.
”A common root cause here is a formula that references a hardcoded month-11 value instead of the cell containing month eleven's calculated result, so the chain of live references silently breaks at exactly that point. That's the kind of error that's easy to miss by eye and easy to catch once you specifically ask Copilot to show its work cell by cell. Laid out side by side, hypothetically, the break is obvious once you see the two formulas rather than the two numbers:
Month 11 and month 12's ending customer formulas, illustrated with a hypothetical broken reference
| A | B | C | |
|---|---|---|---|
| 1 | Cell | Formula | What it actually does |
| 2 | Month 11 Ending Customers (K14) | =K13+K11-K12 | Correctly adds new adds and subtracts churn from month 11's own starting count |
| 3 | Month 12 Ending Customers (L14) | =164+L11-L12 | Uses the number 164 typed in directly instead of referencing K14, so it ignores whatever month 11 actually calculated |
If month 11's starting assumptions ever change, K14 recalculates correctly and L14 doesn't, because L14 was never actually connected to it. The two cells look the same at a glance, a number next to a label, and only reading the formula bar tells you one of them stopped being a model somewhere around month 11.
Tip
Ask Copilot to explain a formula in plain language even when you can read it yourself. Having it state, in a sentence, what it thinks a formula does is a fast way to catch a mismatch between what it intended and what the formula actually calculates, since those two things aren't guaranteed to agree.
What to keep doing by hand
The assumptions themselves, growth rate, churn, pricing, are judgment calls that belong to you, not Copilot. It has no basis for guessing whether 5% monthly churn is realistic for your specific business; it can only build correctly on whatever number you give it. The same goes for any model going to investors, a bank, or a board: run the finished model past a second set of human eyes before it leaves your hands, the same discipline you'd apply to a model built entirely without AI. Copilot changes how fast you build the scaffolding. It doesn't change what a serious financial model requires before anyone else sees it.
Describe the model's logic and structure precisely, and insist on live formulas, not hardcoded values
Read a sample of the actual formulas, not just the numbers they produce
Rebuild at least one row or section by hand as an independent check
Stress-test with an extreme input to confirm the model reacts correctly
Keep your own assumptions as your judgment call, and get a human review before anything goes external
Official sources
Checked on September 21, 2026. Features, plans and names change often, so the vendor's own pages are the final word.