Build your Innovator Founder Visa business plan inside Claude

The instinct, when you sit down to write an Innovator Founder Visa business plan, is to open an AI assistant and type "write me a business plan for my startup." What comes back reads beautifully and fails endorsement. The problem is not that you used AI. The problem is that you asked AI to do the wrong job.

A general-purpose model is trained to produce persuasive prose. Endorsing bodies are not assessing persuasion — they are running a rubric and verifying evidence. The right way to use AI for this document is to keep the reasoning and writing yours, and to use the model for what it is genuinely good at: structure, scoring, and finding the holes before an assessor does.

Why "write me a business plan" is the wrong prompt

Ask a model to generate a full plan and it will reach for the conventions it has seen most often: investor-pitch headings, a top-down market figure ("capture 1% of a £5bn market"), team bios that sound like LinkedIn summaries, and a hockey-stick revenue chart. Every one of those is a red flag to an endorsement assessor.

That is the failure mode in one sentence. An assessor who decides your plan was machine-generated stops reading it as your business and starts reading it as a submission to be caught out. And they will catch it — in the formal presentation interview, where you have to reason about your own business without the document in front of you.

The deeper issue is that a generic model does not know the rules. It does not know the three endorsing bodies score against innovation, viability, and scalability. It does not know the 24-month runway rule or the £100k-per-employee scalability benchmark. It cannot run the displacement test. So it produces a document that is fluent and structurally wrong.

The model that works: structure first, writing yours

There is a better division of labour. Use AI for the parts that are mechanical and rule-bound; keep the parts that require your judgement and your evidence.

What AI should do

• Impose the right structure. A visa plan should be organised so an assessor can trace each section back to innovation, viability, or scalability. AI is good at scaffolding this once it knows the rubric.
• Score the idea before you write. Running your concept through a scoring framework first tells you whether the idea is endorsement-grade at all — and where it is weak — so you do not spend three weeks drafting a plan for an idea that fails on innovation.
• Find the gaps. "Your plan does not name a break-even month. Your plan does not evidence customer engagement. Your plan does not state the barrier to entry." A model running a checklist against your draft is genuinely useful.
• Build the financial scaffold. A bottom-up model with UK-specific assumptions — employer NI at 13.8%, realistic London rent bands, a 20% Prince2 contingency — that you populate with your own numbers.

What stays yours

• The substance and the voice. The assessor cross-checks the written plan against how you speak. If the plan is more polished than you can naturally produce, the gap shows inside ten minutes of Q&A.
• The evidence. Letters of intent, signed pilots, bank statements. No model should synthesise these. A model that flags their absence is useful; one that fabricates them is fraud and gets the application rejected for fabricated documentation.
• The track record. Your CV is what you actually did. Let a model rephrase it into executive-speak and you will be fact-checked on the inflation in the interview.

Doing this inside Claude

This is the model the TorlyAI connector for Claude is built around. Rather than a separate app that generates a plan for you, it adds six UK Innovator Founder Visa specialists to the Claude you already use — McGonagall for journey orchestration, Dumbledore for strategy and scoring, Harry for drafting structure, Hermione for financial modelling, Kingsley for compliance, and Dobby for evidence and interview prep.

The important distinction: your own Claude does the writing and reasoning. The connector supplies the visa-specific structure, the deterministic 4F score, and the financial scaffolding — never the inference. So the plan that comes out is written in your voice, in your Claude session, with the visa rules wrapped around it. You keep the assistant you already trust; you add the domain expertise it lacks.

A typical session looks like this. You describe your idea. Dumbledore scores it on the 4F Innovation Matrix — Product–Market Fit, Founder–Market Fit, Business–Model Fit, plus a Fortune factor — and returns a composite verdict of pursue, refine, or reconsider, with a gap analysis. If the verdict is refine, you fix the weak dimension before writing a word of the plan. Then Harry scaffolds the sections against the rubric, you write them, and Kingsley reviews the draft for caseworker readiness.

The financial model is where most plans collapse

If there is one section where generic AI does the most damage, it is the financials. A model asked for projections will produce a top-down narrative — pick a market size, claim a percentage, draw a line up and to the right. An endorsing body reads that as fiction.

What survives assessment is a bottom-up model: a cohort-based revenue build, a line-by-line cost base, employer NI and pension on every hire, realistic UK overheads, and a break-even month that falls out of the maths rather than being asserted. The 24-month runway has to be covered by verifiable, liquid funds, and the projections have to clear the £100k revenue-per-employee scalability test.

This is deterministic work — arithmetic, not prose — which is exactly why the connector computes it rather than asking the model to imagine it. The financial modelling tool returns real 3–5 year projections, a sensitivity analysis, the funding gap, and the break-even month, so the numbers are defensible because they were calculated, not generated. You still supply the assumptions; the engine does the maths the same way every time.

A realistic eight-week sequence

If you are starting today, order the work to avoid rework:

1. Week 1 — Score, don't write. Run the idea and your founder profile through the 4F matrix. Fix the weakest dimension first.
2. Weeks 2–4 — Draft in your voice. Scaffold the sections against the rubric; write the substance yourself; have the structure tightened, not the content invented.
3. Weeks 3–5 — Build the financials. Bottom-up model, UK assumptions, contingency line, break-even month.
4. Weeks 5–7 — Collect and check evidence. Letters of intent, bank statements, agreements. Run the draft against the gap checklist.
5. Week 7 — Match and review. Confirm your endorsement body fit and run a caseworker-readiness review.
6. Week 8 — Interview prep, then submit. Rehearse the endorsement interview against the plan you actually wrote.

The order matters more than the tooling. A founder who scores the idea in week one and finds it needs refining has saved a month. A founder who writes the whole plan first and discovers the innovation pillar is empty has wasted one.

External context

The Home Office Innovator Founder Visa guidance is the authoritative rules source. The endorsing bodies operationalise those rules — check Envestors and Innovator International for current criteria. For how the connector fits into a session, see the Claude connector overview.

Key takeaways

• "Write me a business plan" is the wrong prompt — it produces fluent, structurally wrong documents that endorsing bodies reject as not genuine.
• Use AI for structure, scoring, and gap-finding; keep the substance, voice, evidence, and track record yours.
• The TorlyAI connector for Claude follows exactly this split: your own Claude writes; the connector supplies visa-specific structure, the deterministic 4F score, and computed financials. £24/month.
• The financial model is where generic AI does the most damage — insist on bottom-up, UK-specific, contingency-backed numbers that reconcile.
• Sequence the work: score first, draft second, model the financials third, evidence fourth, interview prep last.

• ai-tools
• business-plan
• claude
• endorsement
• financial-model