Deterministic, not hallucinated: the rule for AI in visa financials

Last week we launched TorlyAI Visa Master on Product Hunt, and the question I kept getting in comments and DMs wasn't about the connector. It was about the numbers: "If Claude is generating my financials, won't an assessor see straight through them?" Good instinct. Let's talk about why that fear is correct — and how to design around it.

The number a chatbot will happily invent

Ask a general AI chatbot to "write my three-year revenue forecast" and it will give you one. Confidently. It will pick a plausible-looking growth curve, sprinkle in a customer-acquisition assumption, and hand you a tidy table that reads like it came from a finance team.

It didn't. It came from a language model predicting what a forecast usually looks like. There is no spreadsheet behind it, no headcount it ties to, no unit economics it can defend. It's narrative wearing the costume of arithmetic.

For most uses, that's a harmless first draft. For a UK Innovator Founder Visa endorsement, it's a landmine.

Why hallucinated figures are an endorsement-killer

Endorsing bodies assess your business against three pillars — innovation, viability and scalability — and viability is where the maths lives. An assessor (and later, at contact-point meetings, a reviewer) is trained to do one thing relentlessly: ask where did this number come from?

• "Whatever revenue you project for year two — what's the customer count and average contract value behind it?"
• "Whatever gross margin you've stated — what does your cost of delivery actually consist of?"
• "However many people you're hiring in year one — what does the salary line do to your runway?"

If your answer is some version of "the model produced it," you have already failed the question. The figure isn't wrong because it's optimistic; it's wrong because it's unfounded. You cannot trace it back to an assumption you chose. And an applicant who can't derive their own numbers does not look like a founder who can run the business.

AI-written narrative: fine. AI-invented figures: dangerous.

Here is the distinction that should govern every tool you point at your application.

AI-written narrative is fine. Phrasing your innovation claim crisply, structuring your market section, tightening a clumsy paragraph into something an assessor can read quickly — this is exactly what a large language model is good at. The words are yours; the model is a faster, better-read editor.

AI-invented figures are dangerous. The moment a model is generating the quantities — your 4F score, your revenue line, your burn rate, your break-even month — you have outsourced the part of the application you most need to own and least can afford to fabricate.

So the design rule we built TorlyAI Visa Master around is blunt: Claude does the writing; TorlyAI does the maths. The six specialists give Claude the visa structure and the prompts, but the 4F scoring and the financial projections are computed by fixed models, not generated as text.

What "deterministic" actually means

Deterministic is a plain idea dressed in an intimidating word. It means: the same inputs always produce the same outputs, via a formula you can inspect.

Run the model twice with identical assumptions and you get identical numbers — every time. Change one input and you can see exactly which outputs move and by how much. There is no creativity in the calculation, which is precisely the point. Creativity belongs in the strategy and the prose; it has no business in the arithmetic.

Concretely, in our design:

• The 4F Innovation Matrix score is computed by deterministic evidence-matching, not by asking a model "how innovative is this, out of 100?" Thin input scores low by design — because thin input genuinely is weak, and an assessor will say so.
• The financial model is a real model: revenue derives from a customer count and a price; costs derive from a headcount and a salary assumption; cash flow follows from the two. Change the price, watch the break-even month move.

That reproducibility is what makes a number defensible. When an assessor asks where a revenue figure came from, you can answer: "this many customers at this price, and here's the acquisition ramp that gets us there." That's a founder talking. That's the conversation you want.

A checklist for spotting hallucinated financials

Before you submit anything, run every figure in your application through these four questions:

Any number that fails even one of these is a number you should not be putting in front of an endorsing body.

The free web tier gives you five AI assessments with no card, and the connector (add it from /install) puts the six specialists inside the Claude you already pay for, at £24/month. But the principle stands whatever tools you use: let AI write, make the maths deterministic. For more on how we think about the 4F matrix and viability, browse the rest of /insights.

Key takeaways

• A general chatbot will invent a forecast on request; an endorsement assessor will catch an unfounded figure in seconds, because viability questions are where-did-this-come-from questions.
• Separate the two jobs: AI-written narrative is fine and useful; AI-invented figures are a liability you cannot defend.
• "Deterministic" means same inputs → same outputs via an inspectable formula — that reproducibility is exactly what makes a number defensible to a reviewer.
• Run every figure through four checks: can I derive it, does it reproduce, does it tie to a real assumption, can I defend it out loud?
• Any 4F or financial values you see in tooling (e.g. 0.72) are illustrative — the point is never the specific number, it's that it computes and traces to evidence.

• financial-projections
• ai-tools
• business-plan
• viability
• common-mistakes