How to Prompt AI for a Financial Model
A three-prompt chain, assumptions first, that keeps an AI-built financial model from filling gaps with invented numbers.
A single prompt that says "build me a financial model" produces a generic table with made-up growth rates. A financial model worth using comes from a chain of three prompts: state your real assumptions first, ask for the model built strictly from those assumptions, then ask for a sensitivity check on the one or two numbers you are least sure about. Skip the chain and you get a confident-looking spreadsheet built on numbers the model invented to fill gaps you left open.
Why the one-shot version fails
Ask an AI tool to build a financial model with no inputs and it will not say it lacks the information. It will assume a 10 to 15 percent monthly growth rate, a plausible-sounding churn number, and a cost structure that matches whatever similar company it has seen described online. None of that is your business. The output looks like a real model and is actually a template with your logo on it. The fix is not a better single prompt, it is refusing to let the model guess anything you can supply yourself.
The three-prompt chain
Prompt 1: State assumptions, ask nothing yet
List every number you actually know or have a defensible estimate for: current customers, current MRR, historical growth rate if you have 3+ months of data, gross margin, fixed monthly costs, and your best-guess customer acquisition cost. Explicitly tell the model which numbers are firm and which are guesses, and ask it to restate its understanding back to you before building anything. This step alone catches half of the errors that would otherwise show up in the output, because you see your own assumptions written back and often notice one is wrong or outdated.
Prompt 2: Build the model from those assumptions only
Now ask for a 12-month projection using only the numbers confirmed in the previous step, with an explicit instruction not to introduce any additional assumption without flagging it separately. Ask for the output as a table: month, customers, MRR, costs, net. This is where specifying the exact structure matters, an unstructured paragraph of numbers is much harder to sanity-check than a table.
Prompt 3: Sensitivity check on the shakiest number
Identify the one input you are least confident in, usually churn or acquisition cost for an early-stage product, and ask for the same 12-month table run at two alternate values: a pessimistic and an optimistic case. This turns a single point estimate that will definitely be wrong into a range you can actually plan against, and it is the step most one-shot prompts skip entirely.
A worked example
A two-person AI-built app studio had 40 customers at $29 a month, 6 months of data showing 12 percent monthly growth, and no reliable churn number yet. Rather than let the model guess churn, they ran the sensitivity step at 3 percent and 8 percent monthly churn, two plausible bounds for an early B2C product. The 12-month MRR projection ranged from about $3,100 to $5,400 depending on which churn rate held, a wide enough gap that it changed their runway math and pushed a hiring decision back a quarter. A one-shot prompt would have returned a single confident number with no visibility into how much of the outcome depended on a figure they did not actually have yet.
What to paste in and what to leave out
Include | Leave out |
|---|---|
Actual MRR, customer count, and cost figures from your own records | Numbers you are not sure about, stated as if they were certain |
A note on which inputs are confirmed versus estimated | Competitor figures you found in a pitch deck or news article, unverified |
The specific output structure you want (a month-by-month table) | Requests for the model to also suggest a valuation, a separate and much less reliable exercise |
Where this still needs a human
Any number that will go in front of an investor needs to be checked against your actual accounting records, not trusted from a chat output.
Tax treatment, especially for anything involving multiple jurisdictions, is not something to model this way at all.
A model is only as current as the assumptions you fed it. Re-run the chain when a real number changes, do not patch an old projection by hand.
Financial modeling is one specific business task among many that benefit from a structured prompt chain rather than a single ask; how to prompt AI to write a product spec and how to prompt AI to negotiate a lower price with a vendor follow a similar assumptions-first pattern for different tasks. For the broader discipline behind all of these, see the prompt engineering pillar guide. And once a model gives you a number worth acting on, how to forecast revenue for an AI subscription product covers the next step of turning a projection into a plan.
FAQ
Can AI build an accurate financial model from scratch?
No, and it should not try to. It can build an accurate model from your real numbers plus a clearly stated set of assumptions. The accuracy comes from what you provide, not from the model's own judgment about your business.
What is the biggest mistake people make prompting AI for financial models?
Asking for the model in one shot without first stating assumptions. The model fills every gap with a plausible-sounding guess, and a plausible guess dressed up as a formatted table is more dangerous than an obvious placeholder, because it does not look like a guess.
Should I trust an AI-generated model for a fundraising pitch?
Use it to build the structure and speed up the math, then verify every figure against your actual books before it goes in front of an investor. The chain in this guide reduces invented numbers, it does not eliminate the need for a final human check.
How did this land?
About the author
Developer Advocate
Steve builds something with Swarmz every week and writes up what worked, what broke, and what he'd do differently. Tutorials and hands-on guides are his lane.
