AI financial modeling: the parts it helps with, and the parts it must not touch
In financial modeling, AI belongs to the scaffolding and the narrative, never to the calculation chain. Structuring a model, drafting assumption documentation, explaining what a sensitivity implies and reviewing another analyst's work for common errors are all tasks it does well. Producing the projected numbers themselves is not — a model whose figures were generated rather than computed cannot be traced from output back to assumption, which is the only property that makes a model defensible.
A model is an argument, not a spreadsheet
The output of a financial model is a number, but the purpose is a chain of reasoning: if these assumptions hold, this follows. Its value lies entirely in that chain being traceable — someone should be able to disagree with an assumption and see precisely which outputs move.
This is why generated figures are worthless in a model even when they are approximately right. A plausible revenue projection that cannot be traced to a growth assumption is not a projection. It is a guess wearing a projection's formatting.
The practical rule follows directly: every number in a model should be either an input someone owns, or a formula deriving from inputs. Nothing in between.
Where it genuinely accelerates the work
Once calculation is off the table, a good deal of modeling work remains — and much of it is drafting, structuring and checking, which is precisely the territory where language models are strong.
| Task | Suitable | Notes |
|---|---|---|
| Deciding model structure and schedules | Yes | Good at proposing a conventional layout to work from |
| Drafting assumption documentation | Yes | The task everyone postpones; a draft is better than nothing |
| Explaining what a sensitivity shows | Yes | Reading a computed result, not producing one |
| Reviewing for common structural errors | Yes | Sign conventions, circularity, hardcodes inside formulas |
| Writing the investment narrative | Yes | With every figure quoted from the model, not restated from memory |
| Producing the projected figures | No | Must derive from assumptions through formulas |
| Setting the assumptions themselves | No | This is the judgement being paid for |
| Computing a valuation | No | Deterministic arithmetic with an auditable chain |
Reviewing a model you did not build
Most model errors are structural rather than arithmetical, and they recur. A review that checks the same list every time catches more than a careful read-through, because a read-through follows the model's own logic and therefore inherits its blind spots.
- Trace one output back to inputs, end to end. If any cell in the chain is a typed number that should have been a formula, treat the whole chain as suspect.
- Check sign conventions at every schedule boundary. Cash flow sign errors are the most common material mistake in practice.
- Flex one assumption to an extreme and watch what moves. Anything that does not move when it should is disconnected; anything that explodes reveals an unintended circularity.
- Compare the terminal period against the historic base. A model projecting margins no comparable business has sustained is making an argument nobody wrote down.
- Check that the balance sheet balances in every period, not just the first. A model that balances only at period one has a plug hiding in it.
This checklist is worth running whether the model was built by a person or drafted with assistance. The failure modes are identical.
The honest limitation
No system removes the part of modeling that is actually hard. Choosing a growth rate, judging whether a margin is defensible, deciding what a discount rate should be — these are the work, and they rest on knowledge of a specific business and market.
What can be removed is the eight hours of building the structure, formatting the schedules and writing the documentation around those judgements. That is a substantial share of the elapsed time and almost none of the value, which makes it exactly the right thing to automate.
Common questions
Can AI build a financial model?
It can propose a structure, draft the documentation and review the result, but the projected figures must be computed from assumptions through formulas rather than generated. A model whose numbers cannot be traced back to an assumption is not usable for a decision.
Is AI reliable for DCF valuation?
The discounting arithmetic should be performed in code, where it is exact and can be re-performed. AI is reliable for explaining what the valuation implies and for checking the model structure, not for producing the valuation figure itself.
How do you check a financial model built with AI?
The same way you check any model: trace an output back to its inputs, verify sign conventions at schedule boundaries, flex assumptions to extremes, and confirm the balance sheet balances in every period. Assisted models fail in the same ways as hand-built ones.
What parts of financial modeling cannot be automated?
Setting the assumptions. Deciding a growth rate, a defensible margin or a discount rate depends on judgement about a specific business, and that judgement is the substance of the model rather than its packaging.