Why generic AI writes terrible business plans
The standard critique of AI-generated business plans is that the writing is bad. It isn't. The writing is fine. In some cases it's better than what a tired founder would produce at 11pm the night before a bank meeting. If you grade a ChatGPT plan as a piece of prose, it scores reasonably.
That's exactly the problem. Business plans are not a writing problem. They're an evidence-and-judgement problem dressed up in prose, and AI is being asked to perform the wrong task.
What a business plan is actually for
A business plan is a document a third party uses to make a decision: lend us money, give us a grant, take a board seat, sign the lease. The third party is not reading the plan to enjoy the writing. They're reading it to answer two questions:
- Is the picture this document paints of the world true?
- Is the person behind it credible enough that I'd back them?
Neither question is about prose. The first is about evidence — has this founder actually checked their market size, talked to real customers, modelled their cash position correctly. The second is about judgement — does this founder understand which risks matter, which assumptions are load-bearing, where their own plan is weakest.
A model trained on text can imitate the surface of both. It can produce a paragraph that sounds like someone who checked the market size. It can write a risk section that reads like someone with judgement. But the underlying acts — checking, judging — never happened.
The TAM problem, in one paragraph
Here is the canonical example. A founder asks an LLM to estimate the UK total addressable market for, say, AI-assisted tax software for sole traders. The model produces a number. The number sounds confident. It is wrong, but not in any way the founder can detect, because the model has helpfully rounded it, sourced it to "industry reports", and stated it with the cadence of a McKinsey deck.
A competent adviser, reading the same paragraph, will ask: how many sole traders are there actually, what percentage have any willingness to pay for software, what's the realistic price point, and does the resulting number bear any relationship to the figure you've quoted. Usually it doesn't. Usually the LLM has multiplied two plausible-sounding fractions and produced a number that's an order of magnitude off.
The model cannot do this challenge. The model generated the implausible number; asking the model to critique it is asking the cat to guard the milk.
AI can make your TAM sound plausible. It cannot tell you whether it is. The two activities feel similar from the inside and are entirely different from the outside.
Why "human in the loop" isn't enough
The standard response is "we'll just have a human review the AI output". This is necessary but not sufficient, and the reason is reputational.
A reviewer who is paid an hourly rate to skim AI-generated plans will skim AI-generated plans. They have no stake in the document being right beyond catching the obvious errors. The plan goes to the bank with a polite tick-box from someone who has not done the underlying check, and the founder takes the hit when the lending officer pulls apart the numbers in the meeting.
The model that actually works is adviser in the loop. The adviser owns the relationship with the client. The adviser's name goes on the plan, in some form or another. The adviser's reputation with the local Growth Hub, the local bank, the local grant body, depends on the plans they put their name to being defensible. They have a real stake in the numbers being checked, because the bad outcome lands on them, not on the AI vendor.
This is structural, not motivational. You don't get to "adviser in the loop" by exhorting reviewers to care more. You get there by making sure the person responsible for the output is the same person whose career suffers if it's wrong.
What AI is actually good for here
Within the adviser-in-the-loop model, AI earns its place by doing the parts of the work that genuinely are writing problems:
- Producing a clean first draft of the executive summary from notes the adviser took during the intake call.
- Restructuring a messy founder-supplied risk list into something coherent.
- Generating the three plausible scenarios for the sensitivity table once the adviser has agreed the input ranges.
- Translating an accountant's mid-paragraph aside into the kind of prose a non-finance reader will actually read.
None of these tasks require judgement about the world. All of them require fluency with language. This is the work AI is well-suited to, and the work most founders are weakest at. It's a reasonable division of labour.
The sane architecture
Generic AI writes terrible business plans not because it can't write, but because writing was never the bottleneck. The bottleneck has always been the gap between a founder who is too close to their own idea and a reviewer who knows what to check. AI bridges nothing in that gap. It just produces more polished documents on the wrong side of it.
The fix is structural: put a named adviser between the AI and the client, and make sure that adviser owns the relationship and the reputational risk. That's not a workflow tweak. It's the only configuration that produces plans worth reading.
BusiPlanly is built on that premise — a small UK product for advisers who already do this work and want the intake half automated without losing the review half. Read more, or get in touch about the early adviser programme.