AI Prompts for a Business Plan That Adds Up: Numbers First, Story Second

Most AI business plan prompts produce fluent prose and fictional numbers. The plan reads well until someone asks where the 30% growth in month four came from - and the answer is nowhere.

That is not a writing problem, it is an order-of-work problem. The sequence that holds up is the reverse of the one most people use: read the figures you already have, build an assumptions table you own, and only then let the model write the plan on top of it.

Why an AI Business Plan Fails on Numbers, Not on Prose

A language model cannot know your revenue, your churn or your supplier terms. It knows the shape of the sentence those numbers usually appear in, which is why the output is fluent, well organised and quietly fictional.

What the Survival Data Actually Says

The establishment survival tables published by the U.S. Bureau of Labor Statistics put first-year survival at about 78% for the most recent cohort with a full year of observation, and only about half of the establishments opened in 2020 were still operating five years later. The unit counted is an establishment with employees on unemployment insurance records, not every venture anyone starts. And the useful reading is not that most businesses fail: a plan rarely decides whether an idea is good, it decides whether the operation is alive in month nine, and that is cash arithmetic. The full BLS establishment age and survival data tables are public.

The Model's Default Number Is a Placeholder

Ask for a market size and you get a confident figure with a growth rate attached. Ask for gross margin and you get a percentage that sounds like an industry benchmark. Both arrive in the same tone whether or not anything sits behind them. A large-scale audit of the scientific literature found that fabricated, non-existent citations have grown sharply since late 2024 - a model filling a gap does not flag it. A number with no source, no date and no named owner is a placeholder, and in a plan a placeholder becomes a promise.

AI Business Plan Prompts: The Order of Work Matters

The common failure is inverting the sequence: ask for the plan first, then hunt for numbers that justify what the plan already says. That is writing backwards, and it produces the document that breaks on the first question from a banker or a co-founder.

Four Steps, In This Order

1. Read the figures that already exist.
2. Build an assumptions table that you own.
3. Write the plan on top of that table, one section at a time.
4. Answer by hand the questions no model can answer for you.

Step 1 - Read the Numbers You Already Have

Before writing anything, run an analyst pass over what you have. A structured prompt earns its price here: a general chat request summarises your statements instead of interrogating them.

Five Deliverables, and the Last One Is the Point

Claude Financial Statement Analyst takes financial statements or a year-over-year comparison and returns a sanity check on whether the data reconcile; key ratios (gross, operating and net margin, current ratio, debt to equity, days sales outstanding) with a one-line reading each; three-year trends in revenue, margin and cash; red flags such as falling margins, rising inventory, rising receivables, related-party transactions and one-off items, each carrying the number that proves it; plus a plain-language verdict with the three questions to answer before going further. That verdict is the deliverable that matters: it turns a stack of statements into the short list a human still has to resolve.

The Clause That Makes the Numbers Checkable

What makes that output usable is the instruction it carries: compute every number explicitly, showing formula and inputs; never invent data that is not in the statements; mark estimates as estimates. Any figure can then be re-derived by hand, and a figure you cannot re-derive is not a datum, it is a claim.

Step 2 - Keep an Assumptions Table the Model Does Not Own

Those ratios, trends and red flags become rows in a table you maintain outside the chat - the one artefact that has to survive every session.

Five Columns

Assumption, value, source, date, confidence (high, medium, low). Paste it at the top of each session. An assumption with no source is not deleted - it gets the label "unsourced - verify before showing this to anyone".

The Checks the Model Will Not Run For You

• Cash is not profit. A month can close green on paper and red in the bank: with 60 days of collection from Step 1, month one revenue arrives in month three, and a national survey of small firms run by the Federal Reserve banks keeps listing irregular cash flow among their top financial challenges.
• Delivery ceiling. Revenue divided by price equals units per month you must actually ship, with their cost, labour and time. A projection that grows revenue without asking who produces it is a wish.
• Compound versus flat. A 10% monthly rise applied in a straight line describes a different company from a 10% compound. Ask for the month-by-month table, never the percentage.

Break-Even in One Line

Fixed costs divided by gross margin percentage equals the revenue you must bill just to cover the fixed base. If that sentence cannot be said out loud, the financial section is decoration.

Step 3 - Write the Plan on Top of That Table

Only now does the plan get written, one section at a time, with the table alongside.

The Protocol

Business Plan in One Session works through seven sections and asks for what it needs one at a time: problem and solution; market sized honestly with a TAM/SAM/SOM frame, three named competitors and a differentiator; business model with revenue sources, pricing logic and unit economics (CAC, LTV, margin); go-to-market with channels, a first-hundred-customers plan and six-month milestones; financials as a twelve-month projection with revenue, costs, runway and explicit assumptions; the top five risks with mitigations; and a 250-word pitch. Answer section by section, checking each against your table, not as one long document at the end.

The Rule to Enforce Before You Accept a Section

The prompt carries its own honesty clause: be brutally honest about weak assumptions and never invent market numbers - use ranges and say what needs verification. Send back any section with a single precise market size and ask for a range and where the number would come from. If the model cannot say that it does not know, the document is not fit for a decision.

The Verdict Becomes the Risk Register

The three questions from the Step 1 verdict become open items in the risks block, each with a name and a date. That is where the plan stops being a text and becomes a worklist.

What the AI Cannot Decide For You

This is the honest boundary. Validating demand means talking to customers, not interrogating a model: the U.S. Small Business Administration frames market research and competitive analysis as the work that reduces risk before opening, and points to free or low-cost counselling through Small Business Development Centers and SCORE. Nor does it resolve your pricing power, which is tested by raising prices and losing customers, or your legal structure and tax treatment, which depend on jurisdiction. As one startup advisor put it in 2026, AI helps you write a clean plan but it does not create conviction - traction does. One more boundary: this is a prompting guide, not financial, accounting, tax or legal advice. Confirm figures and decisions with an accountant or advisor, and never paste identifiable financial records, bank details or third-party personal data into a prompt.

A 30-Minute Pass

1. Collect the numbers that already exist - statements, bank records, invoices, quotes - and anonymise what needs anonymising.
2. Run the Step 1 analyst pass and keep the ratios, the red flags and the three questions.
3. Build the assumptions table with value, source, date and confidence.
4. Run the Step 3 plan section by section with the table at the top.
5. Do the checks by hand and correct the table.
6. Rewrite the financial section from the corrected table. The plan is a living document; the table updates first.

Start With the Numbers You Already Have

If you are pre-revenue, the read is shorter but the order is identical. Claude Financial Statement Analyst reads what exists and hands back the ratios, the red flags and the questions; Business Plan in One Session turns those into a plan section by section, with the assumptions on the page instead of invented. The model writes the part that gets read. You decide the part that holds it up.