Financial Modelling with AI: What It Does and What Stays Human

AI is genuinely useful in financial modelling for the mechanical work: gathering inputs from filings, spreading several years of history into a clean sheet, checking that the statements tie, and drafting a first version you can edit. It should not set your assumptions or hand you a forecast to trust blindly, because those are judgement calls that decide the answer, and a model is only as good as the thinking behind its inputs.

Put plainly, AI is fast at the parts of modelling that are tedious and slow at nothing that matters most. The trick is knowing exactly where the line sits, so you get the speed without handing over the decision. This piece walks that line task by task, honestly, with no hype about a machine that models for you while you sleep.

What a financial model actually is

A financial model is a structured view of how a business turns activity into revenue, profit, and cash, built so you can change an assumption and see the result flow through. Most models rest on three linked statements: the profit and loss, the balance sheet, and the cash flow, wired together so they always reconcile. If you are new to that structure, our guide on how to build a three-statement model covers the mechanics; this piece is about where AI fits into that work.

The job splits into two very different kinds of effort. There is the plumbing: finding the numbers, typing them in, laying out the history, making sure the balance sheet balances. And there is the thinking: deciding what will drive the business next year, how fast a segment grows, what margin is sustainable, and what could break the whole picture. AI is transformative on the first and dangerous on the second. Keeping those two straight is the entire discipline.

Where AI genuinely helps

Start with the honest wins. These are the tasks where AI saves real hours and rarely costs you anything, provided you keep a source check in the loop.

Gathering inputs from primary documents. A model needs numbers, and those numbers live in annual reports, quarterly results, and notes to the accounts, often as PDFs that resist copy-paste. Pulling line items, segment splits, and disclosures into a usable table is exactly the kind of high-volume, low-judgement work AI does well. The discipline that makes this trustworthy is primary sourcing: build from the filings themselves, not a vendor’s pre-chewed sheet, so you can trace every figure back. We go deeper on that in building a financial model from primary sources.

Spreading the history. Once the inputs exist, laying out five or ten years of statements in a consistent format is mechanical and error-prone by hand. This is a natural fit for automation. It compresses an afternoon of retyping into minutes and frees you to look at the shape of the history instead of transcribing it.

Checking consistency. A model has dozens of internal ties: the closing cash on the cash flow should match the balance sheet, retained earnings should roll forward correctly, segment revenues should sum to the total. AI is good at flagging where these do not reconcile, which is often the first sign of a typo, a units mistake, or a genuine reporting quirk worth understanding.

Drafting the first version. A blank sheet is slow to start. AI can produce a skeleton (the statement layout, the obvious historical ratios, a first pass at the structure) that you then correct and sharpen. A draft you edit is faster than a page you build from nothing.

The common thread is that AI is strongest when it stays close to a source document and to arithmetic that can be verified. The further it drifts from a citable filing toward opinion, the less you should lean on it. For the broader version of this argument across research tasks, see our practical guide to AI for equity research.

Where the human must stay in control

Now the harder half. These are the parts that decide whether the model is any good, and they are exactly where a language model’s confidence is most misleading.

The assumptions. Every forecast is a stack of judgement calls: how fast does volume grow, does pricing hold, where does margin settle, how much does the company reinvest. These are not facts to retrieve; they are views to form and defend. A machine can offer a starting number, but if you cannot explain why you chose 8% rather than 12%, the model is not yours and you should not stand behind it.

The structure of the forecast. How you drive a model matters as much as the numbers in it. A conglomerate is best modelled by segment, because a single consolidated topline hides how differently its parts behave. Reliance Industries is the standard illustration: in the quarter ended September 2025 its oil-to-chemicals segment did about ₹1,60,600 crore of revenue at roughly a 9% segment EBIT margin, while its digital services segment did about ₹42,700 crore at roughly a 52% margin. The biggest revenue segment is one of the thinnest by margin, and the fattest margins sit in smaller segments. A model that forecasts one blended growth rate would miss all of it. Deciding to build by driver and by segment is a human choice about how to see the business.

The forecast itself. This is the line most worth defending. A forecast that feeds a valuation should come from an explicit, repeatable method: named drivers, stated assumptions, and arithmetic you can reproduce and audit. A number a language model produces in free-form prose fails that test, because you cannot re-run it, trace it, or explain how it was reached. A forecast you cannot reproduce is not usable for putting capital at risk. We make the full case in why deterministic forecasting beats LLM guesses.

Verifying every extracted number. AI’s most expensive failure is a wrong number delivered with total confidence. A single fabricated revenue figure can travel straight into a valuation. So treat every AI-extracted input as a claim, not a fact, and check it against the filing before it earns a place in the model.

A workflow that keeps you fast and safe

The practical answer is not to pick a side but to split the work cleanly. Here is a division that holds up.

Modelling task	Lean on AI	Keep human
Pulling inputs from filings	Yes, with a source check	Deciding which disclosures matter
Spreading historical statements	Yes	Spotting what the history implies
Checking internal consistency	Yes	Investigating why something does not tie
Setting growth and margin assumptions	No	Yes, and you own the reasoning
Building the forecast that feeds valuation	No	Yes, explicit and reproducible
Signing off on the output	No	Yes, always

Read the table the right way. The left column is where speed comes from, and there is a lot of it. The right column is where the model’s worth is decided, and none of it should be outsourced.

One more habit. When you anchor a forecast to something external, like a management guidance band, record the exact words and grade them later rather than taking them as given. A promise of “high single-digit volume growth” is an input to weigh, not a fact to accept. The judgement about how much to trust it is yours.

The honest bottom line

AI changes the economics of the boring parts of modelling and leaves the important parts exactly where they were. It reads faster than you, spreads history without complaint, and never tires of checking arithmetic. It does not know what a sustainable margin is for a business it has never thought hard about, and it cannot be accountable for a number that moves capital.

Used with that boundary in mind, AI makes a good modeller faster and a careful one more thorough. Used without it, the same speed just lets you build a wrong model more quickly. The tool is worth having. The judgement is still the job.