How to Write a Business Plan With AI, and What to Check Yourself

Use AI to structure the plan and draft sections from your own notes, not to supply the facts. Start with a facts sheet, let ChatGPT or Claude interview you to fill the gaps, draft one section at a time, and build the numbers in a spreadsheet yourself. Then check every market figure at its source, every competitor claim and every sum.

Done this way, you can have a full first draft in a day rather than a week. The checking takes about as long as the drafting, and it's the part lenders, landlords and partners actually test. Below: what AI should and shouldn't write, the prompts, a worked example of the numbers, and the checklist of what to verify yourself.

Follow me on Instagram@sagnikteaches

What AI writes well, and what only you can supply

The pattern is simple: AI handles structure, wording and questions; you provide facts, numbers and judgement. When a plan goes wrong, it's almost always because AI was allowed to provide facts.

Connect on LinkedInSagnik Bhattacharya

How long each stage takes

Most of the time saved is in the drafting. The spreadsheet and the checking take roughly what they always did, and they shouldn't be squeezed.

Subscribe on YouTube@codingliquids

Step 1: Write a facts sheet before any drafting

Spend an hour filling this in. It becomes the only source the AI is allowed to use, which is what keeps invented details out.

BUSINESS PLAN FACTS SHEET
Purpose of the plan:      [bank loan / partner / landlord / our own planning]
The ask:                  [amount, what it pays for, how and when it's repaid]
What we do (2 sentences):
Who pays us and why they choose us:
Premises and area served:
Trading history:          [revenue each year, customers per month]
Prices:                   [list]
Team:                     [roles, hours, cost]
What changes with this plan:
Known risks:
Numbers I'm unsure of:    [list them honestly]
Sources I already have:   [reports, supplier quotes, lease terms]

Filled in, the key lines for the osteopath in the worked example below might read like this. The figures are hypothetical, and the honesty in the last two lines is the point:

Purpose of the plan:      bank loan, and our own planning
The ask:                  $15,000: $9,000 fit-out and equipment,
                          $6,000 working capital for the first months;
                          repaid over 3 years, starting month 4
Who pays us and why:      adults with back, neck and sports injuries;
                          most find us through doctor and physio
                          referrals or a friend
Trading history:          first clinic about 1,650 appointments last
                          year, about 140 a month now; waiting list
                          of roughly two weeks since spring
Team:                     owner full time; associate for the new site,
                          paid 50% of each fee, starts month 1
Numbers I'm unsure of:    how fast the new site fills (my guess: 15
                          more appointments each month); whether one
                          associate can cover five days
Sources I already have:   landlord's heads of terms, fit-out quote,
                          last two years' appointment reports

Step 2: Let the AI interview you

Before any writing, have the AI find the gaps. This keeps the content yours and surfaces the questions a lender would ask anyway.

You are helping me prepare a business plan for [purpose]. Here is my
facts sheet. Do not write any of the plan yet.

Ask me questions one at a time, up to 25 in total, to fill the gaps a
[bank / investor / landlord] would care about. Prioritise questions
about money, demand and risk. If one of my answers seems inconsistent
with the facts sheet or an earlier answer, say so before moving on.
At the end, give me a list of the facts I still haven't provided.

[paste facts sheet]

Answer honestly, including "I don't know". A plan that admits two uncertainties and explains how you'll test them is more convincing than one that has an answer for everything.

The inconsistency rule earns its place quickly. In an illustrative run on a facts sheet like the one above, the AI's seventh question was:

"Your facts sheet says the associate is paid 50% of each fee, but your answer to question 4 says they'll work 30 hours a week 'on a set wage'. Which is it? A fixed wage raises your monthly costs whether or not the diary fills, so it changes when the new site breaks even."

That is exactly the question a bank manager would ask across the desk, and far better to meet it in a chat window. The owner's answer (fee share for the first year, then a review) went back into the facts sheet, so every later draft used the same terms.

Step 3: Draft one section at a time

Drafting the whole plan in one go produces smooth, generic text. Section by section gives you control. Use this prompt for each:

Using ONLY the facts sheet and my interview answers above, draft the
[section name] section of the plan, about [300] words, in plain
language for [reader]. Short paragraphs. No marketing adjectives.

Where you need a fact I haven't given you, do not invent it.
Write [NEED: what's missing] in square brackets instead.

The [NEED] placeholder is the most useful line in the whole process. When the drafts are done, search the document for "NEED" and you have your to-do list. Without it, the AI fills gaps with plausible numbers that look exactly like real ones, the problem covered in catching made-up figures in AI drafts.

Here is what a draft with placeholders looks like in practice. A two-person bicycle repair workshop, planning to add a van for call-out repairs, asked for its market section and got this back (illustrative):

MARKET AND CUSTOMERS
Most of our work comes from commuters and families within a short
ride of the workshop. Last year we completed [NEED: number of repair
jobs] jobs, of which [NEED: share from repeat customers]% were repeat
customers. Demand for cycle servicing continues to grow as more
people cycle to work. The call-out van targets customers who cannot
easily bring a bike in, such as [NEED: which groups, and how you know
they want this].

Two of the three placeholders are easy to fill from the workshop's job book. The third is the real test of the idea, and the owner answered it with a count: 23 phone enquiries in the last six months asking whether someone could come out, logged on a notepad by the till. The sentence the AI didn't flag is the one to worry about. "Demand for cycle servicing continues to grow" has no [NEED] and no source; the model wrote it because such sentences usually appear in market sections. Either find a dated source for it or delete it. That pattern, where a claim slips through without a placeholder because it sounds general, is the one the checking step below is built to catch.

Then edit each draft for specifics. A typical first-draft line reads: "Our clinic delivers high-quality, patient-centred care to a growing community." It says nothing a reader can check. The edited version reads: "The first clinic sees about 140 appointments a month, 60% of them returning patients, and has had a two-week waiting list since spring." Every section should have sentences like the second one. If you can't write them, the [NEED] list isn't finished.

A good order: services, operations, market, competitors, marketing, risks, then the executive summary last. If you've already done a SWOT, paste it in for the market and risks sections; running a SWOT analysis with AI shows how to build one on evidence.

Step 4: Build the numbers outside the chat

Don't let AI produce your financial totals inside a conversation. Language models are unreliable at multi-step arithmetic, and why AI is bad at maths explains the mechanism. More importantly, a lender wants a spreadsheet they can follow and question. Use AI to help you design it, write formulas and challenge assumptions. Build it yourself.

Start with an assumptions table at the top of the spreadsheet, so every figure below it is a formula that points back to one of these cells.

Worked example: an osteopath opening a second clinic

Imagine an osteopath with one established clinic planning a second site, staffed by an associate paid half of each fee. All figures here are hypothetical.

Monthly fixed costs are $2,450 while marketing runs, then $1,850. At $35 kept per appointment, the site covers its monthly costs at 70 appointments a month during the marketing period, and about 53 after it. With appointments rising by 15 a month, the site breaks even on monthly costs in month 3 and, once the early losses are added up, recovers the $9,000 fit-out in month 9.

Now the reason for the spreadsheet: change one cell. If appointments rise by 10 a month instead of 15, the fit-out is recovered in month 11, not month 9. That single sensitivity line, "if growth is a third slower, payback moves by two months", is worth more to a lender than a page of optimistic prose. Ask AI to suggest which assumptions to test this way, then run the tests in the spreadsheet yourself.

A useful prompt at this stage:

Here are the assumptions and formulas from my business plan
spreadsheet [paste]. Don't calculate totals. Instead:
1. Check each formula does what its label says.
2. Tell me which three assumptions a lender is most likely to challenge,
   and what evidence would satisfy them.
3. Suggest which assumptions I should stress-test, and by how much.

Run on the osteopath's spreadsheet, the reply included this (illustrative):

1. Row 12 "Marketing" is =$B$6 in every month from 1 to 24, but
   your assumptions say marketing runs for six months only. Months
   7-24 overstate costs by $600 each.
2. Likely challenges: (a) growth of 15 appointments a month - what
   is it based on? (b) the associate's 50% share - is it in a signed
   agreement? (c) no allowance for the associate's holidays.
3. Stress-test: growth at 10 a month; fee at $65; one month in
   twelve with no associate cover.

Point 1 was a real error. The owner had copied the marketing formula along the whole row, so the spreadsheet disagreed with the back-of-envelope sums, and only a formula check would have caught it. Point 2(c) was the useful surprise: an associate on a fee share who takes five weeks off a year earns nothing for the practice in those weeks, while the rent still falls due, so the model now assumes 47 working weeks, not 52. The $65 fee test was less useful. The fee already sat on the first clinic's price list and had no reason to fall, so that test was dropped and the other two kept.

For a loan, you'll usually need a cash-flow forecast too; preparing a business loan application with AI help covers what goes alongside the plan.

Step 5: The checklist of what to verify yourself

Numbers

• Every figure in the text matches the spreadsheet. Search the document for each number and compare.
• Totals and percentages recalculated by you, not taken from the AI's prose.
• Monthly and annual figures aren't mixed up anywhere.

Market claims

• Every market statistic traces to a named, dated source that you opened yourself. Delete any statistic you can't find.
• No "studies show" or "research suggests" without the study named.

How this tends to show up: an early draft of the osteopath's plan said demand for osteopathy "has grown by 12% a year, according to industry reports". No report was named, and a search turned up nothing that said it. The model had produced a figure that looked like a statistic, and it would have been the first thing a lender asked about. It came out. In its place went something the practice could prove: the first clinic's two-week waiting list and the number of patients turned away last quarter.

Competitors

• Claims about competitors match their current websites. Keep them factual and fair; you don't want to put an inaccurate statement about a named business in writing.

Rules and obligations

• Licences, professional registration, insurance and lease terms confirmed with the regulator, your professional body, your insurer or a solicitor. AI summaries of rules can be out of date or simply wrong for your situation.

Consistency and voice

• Staff numbers, prices and dates are the same in every section.
• No "[NEED" placeholders left.
• The ask is the same everywhere. A typical slip: the fit-out quote rises from $9,000 to $9,800 after the first draft, the spreadsheet is updated, and the executive summary still asks for $15,000. Search for the ask amount last of all.
• It reads like you. Cut words like "passionate", "innovative" and "world-class". Readers see them as filler, and increasingly as a sign nobody edited the draft.

Write the risks section as if the plan had already failed

AI will happily produce a risks section of the standard kind: "increased competition", "economic downturn", "staff turnover", each followed by a reassuring sentence. Readers skip it. A better approach is to ask: "It's two years from now and this second clinic has closed. Give me the ten most likely reasons, specific to an osteopathy practice with one associate." Then keep the three that genuinely worry you and write a real response to each: what you'd notice early, and what you'd do. For the associate-leaves risk, that might be a notice period in the contract and a list of locums. The same thinking, in workshop form, is in running a pre-mortem before you launch.

For the osteopath, three of the ten reasons made the cut, and each got an early warning a lender could see the practice would act on:

The seven that didn't make it included "economic downturn" and "changing patient preferences". Both are true of every clinic, and neither gives the owner anything to watch.

Keeping confidential figures safe while you draft

A business plan contains your most sensitive numbers: revenue, margins, borrowing, sometimes personal guarantees. Draft it on a business plan of your AI tool, which doesn't train on your content by default, or switch off the model-training setting in a personal account's privacy settings first. Leave out personal financial details of partners and guarantors entirely; the plan can refer to them without the AI ever seeing them. Customer details don't belong in a business plan at all, so anonymise any examples you paste in.

What readers of your plan will notice

People who read many business plans skim the prose and interrogate the assumptions. Expect questions like "where does 15 new appointments a month come from?" and "what happens if the associate leaves?". If your answer is "that's what the plan says", you've lost them. Every number should come with a sentence of evidence: last year's growth at the first clinic, enquiries already on a waiting list, a comparable site.

You can rehearse that interrogation before the meeting. Open a fresh chat with no history, paste the finished plan, and ask: "You are a small-business lender reading this for the first time. List the eight questions you would ask the owner, and for each, say whether the plan answers it and where." In an illustrative run on the osteopath's plan, six were answered and two weren't: what happens if the landlord won't renew the room after the first year, and whether the first clinic stays as busy while the owner spends time setting up the second. Each deserved a paragraph. Use a fresh chat for this, because the conversation that helped write the plan already "knows" the answers and will mark them as covered.

They'll also notice generic text. A plan where every section is smooth and nothing is specific reads as unedited AI output, whether or not it is. The fix isn't to disguise anything; it's to make sure each section contains details only your business could have written. If the plan is for a grant rather than a loan, check the funder's guidance on AI use first; using AI for grant writing without funders rejecting it covers that.

Business plans and AI: follow-up questions

Which AI tool is best for writing a business plan?

Any of the main assistants, ChatGPT, Claude or Gemini, can draft the text well from a good facts sheet. The difference between a strong and weak plan is your inputs and checking, not the tool. Use one that can read your documents and, if the plan contains sensitive figures, a business plan that doesn't train on your content by default.

Can AI make my financial projections?

It can help you design them: which lines to include, how to lay out assumptions, and the spreadsheet formulas. Don't let it produce the totals inside a chat, because sums and percentages slip and there's no working to follow. Build the model in a spreadsheet you can change, then ask AI to review the formulas and question your assumptions.

Do I need to tell a lender I used AI?

Lenders generally judge the plan and the person, not how the text was typed. What matters is that every number is yours and you can explain it without the document in front of you. Grant funders are different: some have rules about AI-written applications, so read the funder's guidance before you submit.

Further reads

• How to Build an Investor Pitch Deck With AI, and What to Check — Turn the finished plan into a pitch deck.
• How to Use AI for Scenario Planning: Best, Worst and Likely Cases — Build best, worst and likely versions of the numbers.
• How to Use ChatGPT as a Business Adviser Without Being Misled — Use AI to challenge the plan without being flattered.
• How to Test a New Service Idea With AI Before You Launch It — Test demand before you put it in the plan.
• How to Build a 13-Week Cash Flow Forecast With AI Help — The short-term cash view lenders often ask for next.
• How to Do Competitor Research With AI in One Afternoon — A fuller method for the competitor section.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.