AI Startup Business Plan: Sections, Costs and a Worked Example

An AI startup business plan follows the sections that lenders and investors commonly request, as the SBA lists them, then adds four AI parts: data rights, model costs, compute and a moat. This page works one plan through for a fictional company, with every number shown.

By the 15minuteplan.ai team · Facts checked 2026-11-05

Model cost and gross margin calculator, per customer

The calculator starts on the worked example's month 12. Use the presets for the cases in this article, or enter your own numbers.

Section order investors read first

The U.S. Small Business Administration (SBA) lists these sections for a traditional business plan: executive summary, company description, market analysis, organization and management, service or product line, marketing and sales, funding request, and financial projections. The SBA page says lenders and investors commonly request this plan, and that you do not have to keep to its exact outline. Our SBA template page has the outline as a copyable list.

For a fast read we suggest a different order from the SBA's. It is our recommendation for a reader with five minutes, not a rule the SBA sets:

1. Executive summary. One page: the problem, the product, the model in one sentence, the cost per customer and the ask.
2. Funding request. The amount, debt or equity, the terms you want, and the period the request covers.
3. Financial projections. The SBA page asks for a five-year outlook, with the first year quarterly or monthly. Build cost of goods from the AI costs below. A percentage of revenue is a shortcut; if you use one, show where it comes from.
4. Technology and data. The model, the data you use and its rights, and the usage per customer.
5. Market and customers. Who pays, how many, and how fast they arrive.
6. Competition and moat. Who you compete with, and the moat as a test (below).
7. Team and organization. Who builds the product and who sells it.
8. Risks. Including a rise in the model price and the fixed costs before you scale.

AI parts to add: data, model costs, compute and moat

Data: what you own and what you may use

Write three things: what your terms grant customers about the text the product generates, whether customer data trains a model, and which privacy law applies. The rights depend on your contract and on the law that applies, so state the terms you actually offer and link them rather than assume an answer. If personal data appears in the text, such as a named staff member in a product biography, the GDPR can apply if your company is established in the EU or offers goods or services to people in the EU. Ask counsel to confirm.

The example budgets US$3,000 for a one-time legal review of the data terms. That figure is a placeholder for a quote.

Model costs: count tokens per customer

Model cost grows with use, so count it per customer. Multiply requests per customer per month by input tokens per request, then by the input price per million tokens. Do the same for output tokens at the output price. Add the two. Then add 20% for retries and checks.

The example uses 400 requests a month per shop, 1,500 input tokens and 300 output tokens per request. The token counts are assumptions to measure on the prototype before the plan goes out. The prices are a Sonnet-class list price: US$2 per million input tokens and US$10 per million output tokens, from Anthropic's pricing page as checked on 2026-11-05. Check the provider's price page on the day you use it.

Input: 400 × 1,500 = 600,000 tokens, or 0.60 million, at US$2 is US$1.20. Output: 400 × 300 = 120,000 tokens, or 0.12 million, at US$10 is US$1.20. Together that is US$2.40. With the 20% overhead it is US$2.88 per shop per month, about US$0.007 per description.

Compute: hosting is fixed, training is variable

In this example, hosting, storage and the job queue cost a fixed US$400 a month. That is an assumption. Real hosting can rise with use, so model usage tiers or get a quote where they apply. Replace the US$400 with two quotes. If you train your own model, add GPU hours times the hourly rate. The example uses a third-party model and trains nothing, so it has no GPU line.

Moat: a test an investor can check

A moat claim needs a number the founder can show. Four tests can be measured:

• saved brand-voice profiles reused each month;
• the share of generated text a customer accepts without edits;
• months of a customer's saved catalogue data kept inside the product;
• live integrations with the shop platforms your customers use.

Give each test its data source and its period. If the number is not known yet, write 'not measured yet' and the date you will measure it.

Standard plan and AI plan, side by side

Table 1. The SBA page's listed sections and the AI additions. Standard column: the SBA page, checked 2026-11-05. Numbers: the worked example below, checked 2026-11-05.

Worked example: Lanternfish AI, a fictional company

Lanternfish AI is invented for this example. It sells an AI assistant that writes product descriptions for small online shops, at US$79 a shop per month. It has 20 paying shops in month 1 and 300 in month 12. Table 2 lists every input, and each step of the arithmetic follows.

Table 2. Inputs for Lanternfish AI. A placeholder is a value the founder replaces before the plan goes to an investor. An assumption is a value to test.

Month 1 and month 12

Table 3. Lanternfish AI, months 1 and 12, in US$ a month. Model cost is 20 × 2.88 and 300 × 2.88. Payment fees are 3% of revenue. Gross margin is gross profit divided by revenue.

Model cost is small at this price, 3.6% of revenue in both months. Under these assumptions, fixed hosting is the larger cost in month 1: 25.3% of revenue (400 ÷ 1,580), against 3.6% for the model. By month 12, hosting falls to 1.7% of revenue (400 ÷ 23,700).

A customer who uses the product twice as much

If the average shop sends 800 requests a month instead of 400, model cost per shop doubles to US$5.76. At 300 shops it is US$1,728 a month, 7.3% of revenue. Cost of goods becomes 1,728 + 400 + 711 = US$2,839, or 12.0% of revenue, and gross margin falls to 88.0%.

A model price that doubles

If the provider doubles its price and shops keep sending 400 requests a month, model cost at 300 shops doubles to US$1,728 a month. That matches the usage case above, so gross margin is again 88.0%.

If usage doubles to 800 requests and the price doubles too, model cost at 300 shops is US$3,456 a month, 14.6% of revenue. Cost of goods becomes 3,456 + 400 + 711 = US$4,567, and gross margin falls to 80.7%: (23,700 − 4,567) ÷ 23,700. A plan should show each risk on its own line, and the two together.

The generator's money step for an AI company

The generator's money step has ten cost lines. Nine are a percentage of revenue. Income tax is a percentage of earnings before tax. Cost of Goods Sold starts at 40% of revenue. The AI business plan generator guide describes the steps before this one.

This example puts model cost, hosting and payment fees in cost of goods, so its table has one gross-margin line. That grouping is this article's choice, not an accounting rule. So the 68.0% and 91.7% margins include hosting. With hosting below gross profit, both months read 93.4%. The app's help text describes cost of goods as costs that go up as you sell more. The help text's test is that a cost rises with sales. Model use and payment fees do. Its examples, raw materials and line wages, are production costs, so this grouping is ours. Fixed hosting does not rise with sales, so in your own plan, show it on its own line.

At month 12 the built-up cost of goods is 8.3% of revenue (1,975 ÷ 23,700). If cost of goods stays at the 40% default, gross margin reads 60% instead of 91.7%. Replace the default with your own build-up, and show the arithmetic in the plan.

Common questions

Which AI cost should a plan show per customer?

The model cost per customer per month: input tokens at the input price plus output tokens at the output price, plus 20% for retries and checks. In this example it is US$2.88 per shop per month.

What should a plan say about a model provider’s price?

Name the price you used, its date and its source; the price can change.

Is Lanternfish AI a real company?

No. It is a fictional company; Table 2 lists its inputs.

Sources and date checked

1. U.S. Small Business Administration, "Plan your business", the traditional and lean plan sections, the funding request and the financial projections: sba.gov, plan your business. Date checked: 2026-11-05.
2. Anthropic, pricing page, the Claude Sonnet 5.5 row (US$2 input and US$10 output per million tokens): platform.claude.com, pricing. Date checked: 2026-11-05.
3. 15minuteplan.ai, the generator's money step: ten cost lines, the 40% default for Cost of Goods Sold, and the income tax rule, read from the product repository. Date checked: 2026-11-05.