How to use AI to prep accurate financial projections for…

Most small business owners skip financial projections because they can't afford a CFO and don't know where to start. This post walks you through using AI to create financial projections for your small business — a complete 12-month forecast covering revenue, expenses, and cash flow — using ChatGPT{:target="_blank"} or Claude{:target="_blank"} with your real business numbers. AI can execute the mechanical parts of financial modeling accurately when you give it structured inputs and clear assumptions.

What You Need Before You Start

ChatGPT (GPT-4o or o3) or Claude 3.7 Sonnet: The core tool for generating projections from your data. ChatGPT requires a Plus subscription{:target="_blank"} ($20/month) for GPT-4o file uploads. Claude 3.7 Sonnet is available via Anthropic's Pro plan{:target="_blank"} ($20/month). Both handle CSV uploads and multi-step financial calculations.

Your accounting data export: Pull 3–6 months of actuals from QuickBooks{:target="_blank"}, Xero{:target="_blank"}, Wave{:target="_blank"}, or FreshBooks{:target="_blank"} as a CSV. Monthly totals by category is sufficient — you don't need transaction-level detail.

A notes document: You'll record your assumptions before prompting. This step separates usable projections from generic noise.

Time required: 2–4 hours for a complete first draft, including data prep and review.

Skill level: No coding required. Basic comfort with spreadsheets and exporting files from your accounting software.

Why AI Can Replace an Accountant for This Task (and Where It Can't)

AI is good at one specific part of financial forecasting: taking historical numbers and a set of stated assumptions and calculating forward projections across multiple scenarios. That's the mechanical work. According to SCORE{:target="_blank"}, small businesses with formal projections are 30% more likely to secure financing — and that document previously cost $500–$2,500 to produce with a professional.

What AI cannot do: predict external shocks, assess whether your assumptions are realistic for your market, or access your accounts directly. You export the data. You set the assumptions. The AI runs the numbers.

Step 1: Gather and Clean Your Historical Data

Open your accounting software and navigate to the reports section.
Export a profit and loss report covering the last 3–6 months as a CSV. You should see columns for revenue categories, expense line items, and monthly totals.
Open the CSV and remove any one-time items that won't repeat — a large contract, an insurance payout, equipment you already bought. Flag these in a separate column labeled "Non-recurring." You should see a cleaner monthly run rate once these are isolated.
Save the cleaned file. Name it something like actuals_jan_jun_2025.csv so you can reference it precisely in your prompt.

Cleaning your data before you prompt is the most important step. AI will project anomalies forward unless you explicitly remove them.

Step 2: Write Your Assumptions Document Before You Touch the AI

This is the step most people skip. It's also why most AI-generated projections are useless.

Open a plain text document or Google Doc. Write answers to each of these:

Revenue growth rate: What monthly or annual growth do you expect? If you don't know, write "flat" — don't guess.
Seasonality: Do you have slow or busy months? Name them specifically.
Planned changes: New hires, price increases, new services launching, lease renewals, equipment purchases — list them with the month they happen.
Revenue model: Are you subscription-based, project-based, or retail? This changes how AI models your cash flow timing.
Payment terms: Do clients pay immediately or on net-30 or net-60 terms? This is critical for cash flow accuracy.
One-time future items: Tax payments, annual insurance premiums, equipment you plan to buy.

You should now have a one-page assumptions document. This is the brief you give the AI before it builds anything.

Step 3: Run the Assumptions Conversation First

Do not paste your data and ask for a projection in one message. Separate the conversation into two parts.

Open a new chat in ChatGPT or Claude.
Paste your assumptions document and send this message first:

"I'm building a 12-month financial projection for my small business. Before I share my historical data, I want to align on assumptions. Here are my stated assumptions:

Revenue model: [subscription / project-based / retail]

Expected monthly growth rate: [X%]

Seasonality: [describe slow and busy months]

Planned changes: [list with months]

Payment terms: [net-30 / immediate / mixed]

Known large expenses: [list with months and amounts]

Based on these, what clarifying questions do you have before I share the data? Flag any assumptions that seem inconsistent or that would significantly affect the projections."

Review what the AI flags. You should see specific questions about your revenue model, seasonality logic, or expense timing — not generic responses. If it produces a projection at this stage without asking questions, prompt it to pause and ask first.

This conversation catches logical gaps before they corrupt your numbers.

Step 4: Use AI to Build Your Small Business Financial Projections

Once your assumptions are confirmed, upload your cleaned CSV and send this prompt:

"Here is my historical financial data for [month range]. Using the assumptions we confirmed, build three documents:

12-month revenue forecast — monthly totals by revenue category, with growth rate applied and seasonality adjustments noted.

12-month expense budget — monthly totals by expense category, including the one-time items we identified at the months they occur.

12-month cash flow statement — month-by-month opening balance, cash in, cash out, and closing balance. Apply my payment terms to adjust when revenue actually hits the account.

Format each document as a table. Label all projections as estimates. List the assumptions used at the top of each table."

You should see three formatted tables with monthly columns, category rows, and a stated assumptions block at the top. If the output is a narrative paragraph instead of tables, reply: "Format this as structured tables only, one per document."

Step 5: Build Best, Base, and Worst Case Scenarios

One scenario is not a plan. It's a guess.

After you have your base projection, send this follow-up:

"Now create two additional versions of the cash flow statement:

Best case: Revenue comes in 15% above base projections each month.

Worst case: Revenue comes in 20% below base projections and one large unexpected expense of [choose an amount] hits in month 6.

Keep expenses the same in the best case. In the worst case, show what month the closing cash balance goes negative."

You should see three side-by-side cash flow tables. The worst case scenario tells you your actual risk exposure. That number is the one to pay attention to.

Step 6: Sanity-Check the Output Before You Use It

Check that monthly revenue figures are within 20% of your actual recent months. Wild swings mean your assumptions or data have a problem.
Confirm that one-time items appear in the correct month — not spread or missing.
Verify that the cash flow timing reflects your payment terms. If you invoice net-30, month 1 revenue should not appear as month 1 cash in.
Cross-check that the expense categories match your real cost structure. AI sometimes invents categories.
Check the IRS and lender requirement: projections must be labeled as estimates with stated assumptions. Confirm both are present in the output.

If any of these fail, correct the specific issue and reprompt. Do not patch numbers manually — fix the source and regenerate.

Dedicated Tools for Small Business Financial Projections

General-purpose AI handles the above well. But three dedicated tools offer more structure if you want a spreadsheet-native workflow.

Rows{:target="_blank"} is a spreadsheet with built-in AI analysis. You import data directly and run AI queries inside the sheet. Good for teams already working in spreadsheets who don't want to manage prompts manually.

Causal{:target="_blank"} is a financial modeling tool built specifically for scenario planning. It handles multi-variable models with less manual setup than prompting a general LLM. Better for businesses with complex revenue structures.

Fathom{:target="_blank"} connects directly to QuickBooks, Xero, and MYOB. It generates reports and dashboards with AI-assisted commentary. Best for businesses already on one of those platforms who want less manual data export work.

Gemini in Google Sheets lets you generate projection formulas and flag anomalies directly inside a sheet. Available via Google Workspace{:target="_blank"}. Free with a Google Workspace subscription. Limited compared to Causal but zero additional cost beyond what you may already pay.

When Something Goes Wrong

The projections show revenue growing every month in a straight line, ignoring your seasonality notes. You described seasonality in your assumptions but the AI ignored it. Add this to your prompt: "Apply the following seasonal multipliers by month: [list each month with a multiplier, e.g., December: 1.4, February: 0.7]." Quantified inputs override narrative descriptions.

The cash flow statement shows the same number as your profit and loss — timing is not adjusted. The AI treated cash flow as equivalent to profit. Explicitly state in your prompt: "This is a cash flow statement, not a profit and loss. Apply my net-30 payment terms by shifting revenue receipt one month forward from invoice date."

The AI produces a projection but refuses to upload your CSV, saying it can't read the file. You're on a free tier or a model version without file analysis enabled. Paste the data directly into the chat as plain text instead of uploading. Copy the CSV content, include it in the message body, and specify that it's monthly financial data formatted as comma-separated values.

What to Do Next

Export your three projection tables to Google Sheets and set up a monthly actuals-vs-projections comparison. Updating your projections against real numbers each month is where the tool becomes a decision-making system, not just a document.

If you want to automate the monthly data pull so you're not manually exporting every time, learn how to connect your accounting software to a Make.com workflow to save time on data prep each month.

FAQ

Can I use AI financial projections to apply for a bank loan or SBA financing? Yes, with conditions. The IRS and most lenders require projections to be clearly labeled as estimates and to include documented assumptions. AI-generated projections meet that standard when the output includes both — which is why the prompt in Step 4 specifies labeling and assumption blocks explicitly. Some lenders may also want a CPA to review or sign off, so confirm requirements with your specific lender before submitting.

How accurate are AI-generated financial projections compared to an accountant's? The mechanical accuracy — applying a growth rate, calculating monthly totals, adjusting for payment terms — is equivalent. The gap is in judgment: an accountant who knows your industry can challenge your assumptions. AI will use whatever assumptions you give it without pushback unless you specifically ask it to flag inconsistencies, which is why Step 3 (the assumptions conversation) exists.

What if I only have one or two months of financial data? Two months is thin. AI will have no basis to identify trends or seasonality. In that case, skip trend-based projection and instead build a budget-forward model: start with your known fixed costs, estimate revenue based on your sales pipeline or capacity, and treat the output as a planning document rather than a statistical forecast. Be explicit with any lender that the projection is based on estimates, not historical trends.

Is OpenAI's o3 model better than GPT-4o for financial projections? For multi-step financial calculations and catching logical gaps in assumptions, o3 produces noticeably more reliable outputs than GPT-4o. If you're on a ChatGPT Plus plan, select o3 when it's available in the model menu and use it for the Steps 3 and 4 prompts. For simpler month-over-month calculations, GPT-4o is sufficient.

Do I need to share sensitive financial data with AI to use it for projections? You share what you export. Strip out client names, bank account numbers, and any personally identifiable information from your CSV before uploading. Monthly revenue and expense totals by category contain no sensitive identifiers. Review OpenAI's data usage policy{:target="_blank"} and Anthropic's privacy policy{:target="_blank"} if you have concerns about how submitted data is handled.

How to use AI to prep accurate financial projections for the next 12 months without an accountant