Financial Model with AI in One Evening: Brief In, Spreadsheet Out

Most small businesses run finances on instinct: revenue is roughly this, costs get covered, the owner takes what's left. That works until a real decision shows up. Hire a second salesperson or wait? Raise prices or hold? Can the business carry a second location? Instinct has no answer to those; a spreadsheet does.

Hiring a finance person for one spreadsheet is overkill. Building it alone in Excel eats a weekend, most of it spent fighting formulas. There is a third route: you bring Claude the numbers and the logic of your business, and it assembles the model structure, the formulas, and the scenarios. One evening gets you a working table. Here is the step-by-step, including the parts where you should not take AI's word for anything.

What to gather before the first prompt

A model doesn't materialize from thin air. Collect four groups of numbers; every operating business already has them:

• Revenue and its sources. How much comes in monthly and from where: channels, average deal size, number of sales.
• Variable costs. Cost of goods or delivery, payment processor and marketplace fees, commission-based pay.
• Fixed costs. Rent, salaries, software subscriptions, accounting: everything you pay whether or not you sell.
• Hypotheses to test. The reason you're building this: a hire, an ad budget, a price change, a second location.

Precision to the cent is unnecessary. Ballpark figures from the last three months are enough for a model you make decisions with.

How to build a financial model with AI: five steps

Step 1. Write a brief, not "build me a financial model." The more specific the input, the less gets invented on the way out. Example prompt:

Build a monthly financial model for 12 months for [your niche, e.g. a six-person design studio]. Inputs: current revenue around [X] per month from two channels: [channel 1 and its share], [channel 2 and its share]. Average deal [Y]. Variable costs: [list them with percentages or amounts]. Fixed costs: [rent, payroll, software with amounts]. Before building anything, ask me clarifying questions if inputs are missing. Then propose the model structure: rows, assumptions, formulas written out explicitly. Put every assumption on its own line so I can change it.

Step 2. Answer the questions and approve the structure. Good sign: Claude asks about seasonality, repeat purchases, payment delays. Bad sign: it hands you a table with smooth growth for years ahead. If that happens, send it back a step: "ask me what's missing first."

Step 3. Ask for two scenarios. Base: things continue as they are. Pessimistic: revenue drops by a third, the main channel dies, supply costs rise. Skip the optimistic one; language models are optimistic enough on their own, and real decisions get made on the base and worst cases.

Step 4. Get the table as an artifact. Ask Claude to deliver the model as an artifact: a table right in the chat showing month-by-month dynamics, or a calculation with explicit formulas you move into Excel or Google Sheets. From there the file is yours: change an assumption and watch what happens to profit and cash.

Step 5. Interrogate the model. "At what revenue drop do we go negative?" "What does hiring a manager at [Z] salary with a quota of [N] deals change?" "How many months until a second location pays back in the base case?" These answers are the whole point of the exercise.

Where AI gets it wrong and what to check by hand

This section separates a working model from a pretty table with a bug inside.

• Optimism in forecasts. Language models love growth: say, ten percent month over month appears in projections on its own, backed by nothing. The rule: every growth rate must trace to your own historical numbers or be explicitly flagged as an assumption.
• Arithmetic and formulas. Verify three or four checkpoints by hand: add up one month's costs on a calculator, check the totals row. One broken formula quietly skews the entire model.
• Missing costs. Taxes, card fees, refunds, paid leave, loan payments: whatever wasn't in the brief won't be in the model. Walk through one month of your bank statement and compare it against the cost lines.
• Profit confused with cash. Paper profit and money in the account diverge, especially with delayed payments and prepayments. If you offer payment terms, ask for a separate cash flow line.
• Identical months. If December equals July in the model and your business is seasonal, add the seasonality yourself. The AI won't know about it until you say so.

Privacy: what stays out of the chat

The model needs aggregates, not documents:

• Don't paste bank statements with client names, account numbers, or contract details. "Channel A brings in about a hundred thousand a month" works just as well and reveals nothing.
• Rename counterparties: "supplier 1," "key client." The model's logic doesn't care.
• Check the privacy settings on your Claude account and choose how your data may be used.
• If you work with a team, keep finance in a separate project and don't drag extra documents into it. We covered how projects and context work separately.

Honest limits: who this model is not for

An evening model answers an owner's questions: hire or wait, raise the price or hold, will the cash last through the quarter. For that, it is enough.

It is not for banks, investors, or anyone who will audit the methodology. That territory requires accounting standards, verifiable data, and a finance professional who stands behind the numbers. Showing an investor a spreadsheet nobody has checked is worse than showing nothing at all.

And the division of labor stays fixed: the AI does the arithmetic, you make the call. The model shows the consequences of each option; choosing between them is your job.

The short version

Gather the numbers, write the brief, approve the structure, ask for a base and a pessimistic scenario, check the formulas by hand, and strip out unfounded growth. What you end up with is a table you can decide with instead of guessing.

Nearby reading: how to merge several Excel files into one and how to analyze business data with Claude.

If you'd rather learn this by doing: AGINE Academy teaches Claude through game-style missions you complete in the real tool, and every lesson ends with a working artifact. An AI mentor helps with the current step. The starter block of four lessons is free and open without signup.