Why AI cannot build your source of wealth file

The letter from the bank asks for proof of funds. So you paste it into ChatGPT, follow the structure it gives you, paste in your exchange export, and a few seconds later you are holding a well-organised document with headings, a summary and a timeline of your crypto life.

It is a good document, and that is the problem.

AI can write two pages of a thirty-page file very well. The question is who does the other twenty-eight.

Almost none of those twenty-eight pages are a writing job. They are a collecting job, a matching job and an arithmetic job, and most of what they need sits behind data that nobody hands out for free.

It reads what you paste, and it cannot go and get the rest

Everything AI does happens after the evidence is already in front of it. It cannot sign in to an exchange account, and it cannot reach blockchain transactions. It cannot, to take one ordinary example, expand an xpub, the extended public key behind your hardware wallet, into the individual addresses that key actually used.

It can ask you good questions, and it should: did you ever use a hardware wallet, did anything go into DeFi, what happened to the coins that left in 2021. What it cannot do is check the answer. It cannot follow that withdrawal to the address it landed on, or see that the address moved the money onward in 2019 to an account you have forgotten. It can prompt your memory. It cannot audit it.

People forget wallets and exchange accounts they opened years ago, including ones with money still sitting on them, and that is normal rather than careless. In about nine cases out of ten, the CashoutReady platform surfaces an account or a wallet that was never declared. It cannot remove one from the file. It can only tell you that it found one.

A gap you flag yourself is a conversation. A gap your bank finds is a decline.

A wallet turning up late is the normal course of a file rather than a mishap, and everything after it is worked out again, which is why this is built somewhere you can keep working rather than delivered in one answer to one prompt.

It reads every row in the dialect of whoever exported it

Every data source speaks its own dialect. A CoinTracking export, a Koinly export, a Kraken statement and the blockchain itself each have their own name for a staking reward, a fee, a swap, a transfer you made to yourself. Some of them use the same word for different things.

Before any of it can be added up, every row has to be translated into one vocabulary, and row thirty thousand exactly as row three hundred was. That translation decides what counts as a deposit, what counts as gain, what is a swap and what is simply your own coins changing pockets. Get one category wrong in a way that repeats, and every total underneath it is wrong.

A language model does not make fifty thousand identical decisions. It matches patterns, and it drifts. Nobody is going to re-check fifty thousand of its judgements afterwards, whereas a deterministic rule, like a formula in a spreadsheet, can be checked once and then relied on for every row it touches. Done properly, every input is read the same way.

It cannot tell your own money from other people’s

Moving coins from your own wallet to your own exchange account is neither income nor a deposit from a third party that a bank will ask you to explain. In a raw export it looks identical to money arriving from a stranger.

AI receiving that export counts it as an inflow, because on the face of the data it is one. Your document then claims wealth you never had, supported by a total nobody can check.

The CashoutReady platform uses battle-tested algorithms to separate your own transfers from money arriving from outside, across wallets, exchange accounts and bridges, before anything is summed.

Your exchange statement has gaps

An exchange deposit line records the amount, the asset and the date. It very often does not record the wallet the money arrived from, and a withdrawal line does not record where it went. The two things a bank most wants to know about a movement are the two the statement leaves out.

What the line usually does carry is a transaction hash. The origin and the destination sit on the chain behind it, and putting a name to either takes a licence somebody pays for. Paste a bare hash into AI and it will explain what a hash is; the CashoutReady platform puts it to the chains it reads and looks up what sits on the other side. Neither answer is in your export, and neither can be worked out by thinking harder about the rows you already have.

Adding it up is where the numbers come from

A few hundred thousand rows are not evidence of anything. What the bank reads is a handful of figures: what came in each year, what it was worth on the day it arrived, how much started as salary, how much as trading, how much as a coin bought early and held, and what the whole of it is worth now. Producing those figures is the work.

Each one is the sum of thousands of movements, every one priced at the moment it happened. AI does not execute arithmetic at all: it predicts what the answer to a calculation looks like, which is a different thing and looks the same on the page. Nor does it know what a token was worth on 14 March 2021, and it will not tell you which of its figures it guessed.

The totals then have to be checked against what the exchange or wallet says you actually hold, coin by coin and account by account, with any difference declared rather than absorbed. A line that does not match means transactions are missing, doubled or mistyped, and that is worth learning while the file is still yours to change.

Every figure keeps its workings, so a reviewer who asks where a number came from gets an answer for that row. A chat transcript has none to give. Imagine explaining that to your bank.

It cannot say who you dealt with

Labels are what turn a column of addresses into a picture of who you actually transacted with, and on a complicated file that is the most useful thing on the page. Whether the address that sent you two ETH in 2021 belongs to an exchange, a mixer or a sanctioned entity is a matter of record. It is held by the same providers that put a name to the wallet behind a deposit, on licences that run from EUR 10,000 to 140,000 a year. Asked the question, AI will still answer, and answer badly, because it has no access to any of it.

A flag is not a verdict either. What matters is what is left once the exposure has been measured and explained, and getting there means tracing the address back and saying what that counterparty was to you. Screening findings stay in the claim whether or not they help.

Where AI genuinely helps: your own two pages

Two pages of a source of wealth file belong to you and to nobody else: where the original money came from, and the story of how it grew. Nobody can rebuild those from the chain, because the facts are in your life rather than in the ledger. They are also the pages most often left out altogether, which is why a file can be full of data and still not say anything.

Writing plainly about fifteen years of your own decisions is harder than it sounds, and that part is a language problem. Give AI the contract from the flat you sold, the payslips from the year you started buying and the loan agreement you repaid, and let it help you turn them into prose.

CashoutReady reached the same conclusion, which is why that last step is built into the service rather than left to you. The model that drafts those pages never sees a raw export. It receives the figures the rest of the work has already produced, so the story and the numbers underneath it cannot drift apart, and it is trained on files of this kind, so what comes out reads the way a reviewer expects.

Even with that in place, a person checks it against the file before anything is sent to you. AI is amazing - if controlled.

Nobody signs a transcript, and nobody is introduced by one

A bank can ask who prepared the file. A CashoutReady claim has an expert author a compliance officer can telephone, and an analyst reviews it before it goes anywhere.

There is a second thing a person does that no tool can. An introduction to a partner bank willing to read the file is made by someone who knows both sides (an option on files above EUR 1 million), and no model will answer this partner’s follow-up question three weeks later.

Which returns to the two pages. AI can write them very well, and it should. The question was always who does the other twenty-eight.