Methodology

How Finautor approaches AI in finance

Our methodology — how the system works, and the discipline behind it.

Finautor is an AI research assistant for investors. It reads live market data, the public record — company filings, economic data — and, when you connect it read-only, your own portfolio, and turns all of it into analysis you can act on. In finance, how an answer is produced matters as much as the answer itself: a confident number from the wrong source is worse than no number at all. This page explains how we build Finautor, and the discipline we hold ourselves to.

Augmenting human judgment, not replacing it

Finautor is built to make a person a sharper decision-maker — not to make the decision for them. It does the legwork a careful analyst would: gathering the data, cross-checking it, laying out the case on both sides, surfacing what you might have missed. Then it hands the judgment back to you. It analyses and explains; it does not tell you what to buy or sell. And it knows where its lane ends: when a question crosses from research into a personal recommendation to buy or sell a specific instrument — the kind of advice that is legally reserved to accredited advisers — Finautor recognises the line and stays on the research side of it, rather than dressing up a recommendation as analysis. The person stays in charge.

Grounded in real data, not the model's memory

A general-purpose model will happily answer a market question from its training data — which may be months stale and was never meant to be a price feed. Finautor doesn't work that way. The figures that matter are pulled from a live source at the moment you ask, each carrying the date and the field it came from, traceable back to where it originated. The model's job is to reason over real, current data — not to recall it from memory.

Structure over scale

More data is not the same as better answers. The harder problem — and the bigger lever on quality — is organising what the system knows, so it can find the right fact and use it correctly rather than re-deriving everything from scratch each time. It's a principle well established in the field, and it shapes how Finautor is built: around a structured, durable knowledge base, not just a bigger pile of text.

Data honesty as an engineering discipline

We treat the reliability of AI output as a first-class engineering problem — written as rules, enforced in code, and tested in our pipeline, not left to the model's good intentions:

• no figure without a source;
• every number carries the time window and the field it came from;
• a statistic stated in one breath stays consistent in the next.

These aren't suggestions in a prompt the model may or may not honour. They're part of the system we test and ship — checks that run automatically before any change reaches production.

We measure our own failure modes

We know the specific ways an AI can mislead in finance — for instance, stating a figure as a fresh fact when it never actually looked it up. So beyond writing rules against them, we instrument the system to measure how often they still happen. Automatic checks run quietly in the background and quantify each failure mode, so we can drive its rate down and decide — on evidence, not instinct — where to add hard enforcement. A research tool earns trust around money by holding itself to a measured error rate and working to shrink it.

Verification and traceability

What Finautor keeps, it keeps with its receipts. Durable analysis is retained alongside a record of the sources and the session it came from, and generated reports carry their sources — so a conclusion can be traced back to what it rests on. Traceability isn't bolted on at the end; it's how the workflow is built, because a research tool you can't audit isn't one you should trust around money.

It won't act on its own

Finautor connects to your portfolio read-only — it can see your positions, never trade them or move your money. Anything that changes your workspace — a saved note, a dashboard, an update to its memory — happens only when you ask for it, never on the system's own initiative, and changes to its memory can be undone. The assistant proposes; you decide and act.

Measured, not asserted

We hold the work to outside yardsticks rather than our own marking. Finautor scores 88% on an independent third-party accuracy benchmark, and an independent security assessment rated the platform 85 / 100. We re-measure as the system evolves — the standard only counts if it's checked by someone other than us.

We learn from how it's actually used

Some of our sharpest improvements come from a small group of power users who put Finautor through real, demanding work and tell us precisely where it falls short. We treat those reports as the engineering signal, not noise — most turn into a shipped fix within days. The discipline on this page exists because real use showed us where it was needed.

Why we publish this

A tool that helps people make financial decisions should be willing to show how it thinks. We publish our methodology so users and partners can hold us to it — and so the standard we set for ourselves is on the record.