FINN— private data layer for financial advisors

Your AI is only as honest as the data layer underneath it. Finn owns that layer.

Finn ingests brokerage exports and client documents, makes them faithfully queryable by the model of your choice, and redacts every word before it leaves your machine. The intelligence lives upstream in Anthropic or OpenAI. Finn's job is that the numbers are right — and that your clients' names never cross the wire.

IThe failure you can't see

A real export. Two annuity rows read “See contract.”

That's enough. The market-value column silently loads as text, the sum truncates “27,431.50” to 27, and the stack beneath most AI tools reports a fourteen-position portfolio as:

Actual portfolio value $207,727.45
What a plain SQL stack returns $203.00
Understated by $207,524.45

No error. No warning. A number an advisor repeats out loud to a client. Finn's coercion guards catch this class of failure — and when a figure cannot be verified, Finn does not guess. It withholds, and says why:

Total market value
Withheld. 2 of 14 rows hold non-numeric values (“See contract”, “Priced monthly”), so any total would be wrong. The unpriced holdings are listed below.

The industry's problem is the data layer, not the model: the best frontier score on the Vals AI Finance Agent benchmark is 64.37% — one in three entry-level analyst tasks failed, mostly on exactly this kind of file.

IIWhat Finn does
  1. Your export lands clean — no cleanup.

    Pershing, Schwab, Fidelity, Vanguard, and NetX360 profiles, plus dialect detection for the messy ones: European number formats, semicolon delimiters, opaque headers, merged rows. A Trust Report names the market-value column it found and shows the arithmetic that proved it.

  2. Nothing reaches the model unredacted.

    Names, account numbers, and identifiers are stripped locally before any AI call — the model only ever sees redacted text. A Boundary Report shows exactly what crossed, per conversation. Runs on your own machine or your own cloud; works with Anthropic, OpenAI, or fully local models.

  3. The client-meeting loop, closed.

    Prep a meeting in one call — IPS drift, concentration, wash-sale exposure across the household, tax-loss harvesting, aged action items — computed deterministically, with no LLM anywhere in the page, so nothing can be fabricated. Then record the meeting, and the notes and action items write themselves up.

IIIWho it's for

Fiduciaries who are personally accountable for every number.

Finn is built for solo RIAs and small advisory shops — the advisor doing their own meeting prep at 9pm, and the same advisor the next morning with the screen shared and a client watching. Every figure carries provenance; anything that could not be verified says so instead of guessing.

IVRequest access

Finn is in paid pilot with a small number of firms.

Access is granted personally. Write a line about your firm — custodian, rough household count, what meeting prep looks like today — and you'll get a reply from a person, usually with a working session on your own export files.

Already piloting Finn? Log in.