Can AI Build a Financial Model? What It Gets Wrong
AI tools can draft an LBO or DCF in seconds. Here is what they handle well, where models break, and why checking the output is now the analyst's job.
By Victor D’Errico, Founder & CEO of SheetRank
Finance & Business Analytics at Sacred Heart University · NY licensed real estate salesperson
IB · 4 min read · Updated
Reading is half of it. Build a real model and get graded cell by cell.Try it free →Short answer: yes, AI can draft a financial model, and it's getting better fast. No, you can't send that draft to anyone without checking it. And the person who can check it is worth more than the person who can only prompt it.
That second part is the one that matters for your career, so let's look at it honestly.
What AI is genuinely good at
Give a current AI model a prompt like "build a five-year LBO with a term loan, a revolver and a cash sweep" and you'll usually get something that looks right:
- Structure. Sources and uses, an operating build, a debt schedule, returns. The sections are in a sensible order.
- Standard formulas. Revenue growth, margins, interest on a beginning balance, MoIC, IRR.
- Speed. A first draft in seconds instead of an hour.
- Explaining. Ask it why the sweep comes after mandatory amortization and you'll get a clear answer.
That's real value. Laying out a model is tedious, and a good draft gets you to the interesting part faster.
Try it
Work out enterprise value and EV/EBITDA
Equity value plus debt, minus cash, is enterprise value. Divide by EBITDA for the multiple.
| A | B | |
|---|---|---|
| 1 | Share price | |
| 2 | Shares outstanding | |
| 3 | Debt | |
| 4 | Cash | |
| 5 | EBITDA | |
| 6 | Equity value | |
| 7 | Enterprise value | |
| 8 | EV / EBITDA |
Tap cells to put them in the formula. On a laptop you can type the numbers and operators.
Where it breaks
The problems aren't usually in the big ideas. They're in the details that make a model tie out, and they look exactly like the mistakes junior analysts make:
| Failure | What it looks like |
|---|---|
| Wrong reference | A formula reads the row above or below the one it should |
| Broken roll-forward | Every year's beginning balance points at Year 1 |
| Timing | The exit uses Year 4 EBITDA, or a discount factor is a year off |
| Sign | Change in working capital added instead of subtracted |
| Hardcodes | A plug number typed in to make something balance |
| Missing claims | Preferred equity or PIK interest never reaches the exit waterfall |
| Circularity | Interest and the cash sweep feeding each other with no plan for it |
Each of these produces a model that runs, balances and returns a believable number. In SheetRank's institutional LBO, a beginning debt balance that stopped rolling forward takes the IRR from 16.2% to 12.2%. Forgetting to redeem the preferred takes it to 20.9%. Neither looks wrong unless you know what the right answer should roughly be.
Benchmarks point the same way. On the Vals AI Finance Agent benchmark, which tests models on realistic financial analyst research tasks, the best model scored about 69% as of November 2026. That's impressive, and it also means roughly one in three tasks still needs a person to catch what went wrong.
Why "just check the output" is harder than it sounds
To catch a broken cash sweep, you have to know how a cash sweep is built. You have to know that cash available for debt paydown starts from cash flow after mandatory amortization, that it respects a minimum cash balance, and that the revolver usually gets repaid first. If you don't know that, the AI's version and the right version look the same to you.
This is the uncomfortable part for anyone hoping AI means they can skip learning to model. The skill didn't go away. It moved from typing formulas to reviewing them, and reviewing is harder, because you're looking for something you didn't write.
How to use AI on a model without getting burned
Know the answer's shape first
Do a paper estimate of the key outputs before you read the draft. If they disagree, start digging there.
Trace from the output back
Start at the IRR or share price and follow precedents to the assumptions. Check each step against how you'd build it.
Check every roll-forward
Beginning balance equals prior ending balance, in every schedule, in every year.
Flex the inputs
Change growth, exit multiple and rates. Anything that doesn't move is hardcoded or disconnected.
The full version of this process is in how to check a financial model for errors.
What this means for recruiting
Modeling tests aren't going away. If anything, they matter more now, because a written application tells a recruiter less than it used to. A Robert Half survey from March 2026 found 67% of HR leaders say AI-generated applications are slowing hiring. When every cover letter reads the same, a timed test on a blank sheet is one of the few signals left that's hard to fake.
So the advice for students hasn't changed much, it's just more urgent:
- Learn to build the core models from scratch: three statements, DCF, LBO, and an acquisition model if you're going into real estate.
- Practice finding mistakes in models you didn't build, under time.
- Be able to explain every line out loud.
Try it
The weekly AI Audit is built for exactly this. You get a complete institutional model, every schedule built, with five realistic mistakes hidden in it: wrong rows, broken links, flipped signs, the wrong year, a typed-in number. Fix them all and the clock stops. It's free with an account, and the same model goes to everyone each week, so you can see how your time compares.
Practice it
Now build one from a blank grid.
Project Sunbelt is a full 300-unit multifamily underwriting. Every formula you write is graded the moment you leave the cell. Free, no account, no card.