Skip to content
G·01
← All guides
AIvaluationLLMtooling

Valuing companies with AI: when ChatGPT is enough — and when it isn't

MisanoUpdated

Frontier LLMs are genuinely good at parts of investment analysis, and pretending otherwise would be silly — Misano itself runs on Claude and OpenAI models. The honest question is not 'AI or no AI' but which jobs a raw chat window does well, and which jobs quietly fail in ways you only discover later.

What a raw LLM chat is genuinely good at

  • A qualitative first pass: summarize this teaser, list the risks, what would a skeptic ask.
  • Market context: how this sector works and what typically drives value in it.
  • Drafting: turning your conclusions into a readable memo.

Where the chat window quietly breaks

  • Scanned filings: Czech (and most European) registry statements are often photographed paper. Chat uploads process images at premium token prices, and the numbers come back with silent errors — a decimal comma read as thousands separator, a 'v tisících Kč' scale missed. Nothing warns you.
  • No memory across companies: every valuation starts from zero. The same public filing is re-read — and re-paid for — every time anyone asks about it.
  • Inconsistent scoring: ask the same model about two companies on two days and the implicit criteria drift. Ranking a pipeline requires one versioned model applied to everything.
  • Math in the wrong place: an LLM doing a DCF in its head is a liability. The projection arithmetic, WACC build-up and terminal-value cap belong in deterministic code; the model's job is judgement, not multiplication.
  • Invented multiples: ask for 'the sector multiple' and you get a plausible number with no provenance. A benchmark you can inspect — including its evidence quality — is the difference between a valuation and an opinion.

What a purpose-built layer adds

The economics come from routing and reuse, not magic. Misano routes each job to the cheapest model that does it well — text extraction is not a frontier-model problem, investment judgement is. Public filings are parsed once and cached for every user; each analysis job is metered instead of open-ended chat. And the deterministic parts — DCF math, ratio computation, hard gates, the terminal-growth cap — run as code the AI cannot hallucinate around. The result, costed line by line in the founder's blog post, is roughly 90% cheaper per company than doing the same work in your own chat subscription — with guards a chat window doesn't have.

The honest boundary: if you evaluate one or two companies a year, a chat window plus a methodology guide is fine, and cheaper. The platform pays for itself when there is a pipeline — when consistency across companies, registry ingestion, and not re-paying for the same documents start to dominate the cost of thinking.

Frequently asked

01

Can ChatGPT or Claude value a company?

They can produce a structured first opinion and are excellent at qualitative analysis. They are unreliable at extracting numbers from scanned filings, they invent sector multiples without provenance, and they have no consistency across companies or sessions. Use them for judgement; keep extraction, benchmarks and DCF math in deterministic tooling.

02

Is AI-assisted valuation reliable enough for real investment decisions?

As reliable as its guardrails. The failure modes are known and mechanical — unit-scale errors, decimal commas, hallucinated multiples, arithmetic drift — and each has a code-level guard. A pipeline with those guards plus AI judgement outperforms both a raw chat and a purely manual process on speed at equal rigor.

03

Which AI model should do which valuation job?

Route by job, not loyalty: cheap fast models for document extraction and classification, frontier models (Claude-class) for investment judgement and thesis writing, and no model at all for arithmetic — DCF and ratios belong in code. Single-model-for-everything is the most expensive possible configuration.

04

Is it safe to upload confidential financials to an AI tool?

Ask two questions: where does the data rest, and does any model train on it. Misano's answers: EU data residency (Frankfurt), and zero-retention AI — LLM calls go to Anthropic and OpenAI under terms where neither trains on the data. Any tool that can't answer both questions plainly hasn't earned confidential documents.

See also: the full valuation-multiples benchmark at misano.ai/guides/multiples.