Ask a compliance team about AI and the conversation goes to one question almost immediately: is a model good enough to be trusted with the decision. Good enough to say that this case should be escalated, that this account should be closed.
That question already has a reasonable answer. A model returns a likelihood; a supervisor asks for a reason. When a file is reopened two years later, the institution has to say what it looked at and why it concluded what it did, and a score from a model since retrained is not something anyone can stand behind. The decision stays with people, working to rules they can point at and repeat.
The tempting shortcut: hand AI everything and ask what it thinks
So, letting a model decide is out of the question, and not compliant. The idea is still worth playing with, so suppose for a moment that it was allowed. Take everything the customer sent, hand the lot to a capable model, and ask it for the finished case.
Tempting, but most crypto-related files that get stuck never reach a decision at all.
Five data processing layers, and one language problem on top
A Source of Wealth review runs through five heavy data processing layers before anyone can assess anything.
- Acquisition. Establishing what the customer’s accounts and addresses actually are, and pulling the transaction history of each.
- Normalisation. Translating every record into one vocabulary, so that totals mean the same thing across the file.
- Augmentation. Filling the gaps: what a contract call was, which address a deposit arrived from, what an asset was worth on the day.
- Computation. Turning that into balances, flows and gains that reconcile, each figure traceable to what produced it.
- Attribution. Putting names and risk signals against counterparties.
Interpretation sits on top, and interpretation is where language models are strong. The five are not writing tasks. Ask whether AI can do a Source of Wealth review and the honest answer is “maybe” for the layer on top, and “no” for the five underneath it.
Acquisition: before anything can be analysed, it has to be collected
Fifteen, sometimes thirty accounts and wallets across exchanges and chains is an ordinary customer. Each exchange hands its history over as a CSV in its own layout, or through an API where one exists; each chain has to be read through a provider, and no provider is best for all of them. ChainComply connects to those sources directly and picks the best provider for each, so a history arrives as data rather than as an inbox of files.
What is collected limits everything after it, and the first collection is rarely the last. In ChainComply’s own case work, roughly three quarters of files connect on exchange inflow and outflow data, about a fifth only once a wallet the customer had not mentioned comes to light, and in around one in twenty the transaction data does not connect to the rest of the story at all. A wallet arriving late is the ordinary course of a case rather than a mishap, and everything after it runs again. No model moves those proportions. The customer does, so the work is interactive and iterative - and needs a workspace, not a black box.
Normalisation: fifty thousand decisions that have to agree with each other
Every exchange, every chain and every tax-tool export records the same event in its own terms. A staking reward, a fee, an internal transfer, a swap: each has a different name in different places, and sometimes the same name means two different things.
Every record has to be assigned to one taxonomy, identically at record fifty thousand and at record three hundred. Applied inconsistently, a category produces no visible error, only a total that is quietly wrong. A rule is checked once and holds everywhere it touches; a reading that varies between runs does not. ChainComply resolves every source into one vocabulary of twenty-two kinds of movement, treating a gas-only leg, a spam token, and two assets with the same ticker name and an entirely different valuation the same way every time.
Augmentation: three things the records do not contain
What a transaction was: a contract call arrives as an address and a method, not as a staking withdrawal. Pattern recognition on contract behaviour is a genuine classification job, supplied by specialist providers, and measurable against a known answer. That kind of model classifies a record. It does not narrate a life.
Where the money came from: an exchange-supplied deposit line carries the amount, the asset and the timestamp, and not the address it arrived from. It usually carries a transaction hash, and the counterparty sits on chain behind it, one lookup away through a provider with access to that chain.
What any of it was worth: converting an amount on a date into a figure takes market data from outside the customer’s file, and which source answered has to travel with the number.
None of the three is in what the customer sent, and none of them can be reasoned out. Each one is fetched, from a source somebody pays for.
Computation: aggregation is where data becomes information
A few hundred thousand records are not evidence of anything. What a reviewer reads is a handful of figures: what entered the estate in each year, what it was worth when it entered, how much of it came from salary converted into crypto, from trading, from a position bought early and held for a decade, and what all of it is worth now. Aggregation is the step that produces those figures, and it is where data turns into information.
Each of those figures is the sum of thousands of movements, every one of them priced at the moment it happened and carried forward through every transfer in between. That is millions of operations before a single line of the file is written. The totals then have to be reconciled, per asset and per account, against what the venue says the customer holds, and any difference declared rather than quietly absorbed. And each figure has to keep the movements underneath it attached, because a total that cannot be opened and checked is an assertion rather than evidence.
Attribution: labels are bought, not inferred
Labels are what turn a list of addresses into a picture of who the customer actually dealt with, and on a complex file that is usually the most valuable thing on the page. Once an address is in hand, whether it belongs to an exchange, a mixer or a sanctioned entity is a matter of fact held by commercial analytics providers, on licences that run from EUR 10,000 to 140,000 a year. A model will still answer the question, but poorly, because it has no access to that data.
The same data carries the risk signals, and those need reading carefully. A flag is an opening position rather than a verdict, and what matters is the residual risk once the exposure has been sized and explained. Getting from the first to the second means tracing the flagged addresses back, asking the client what a particular counterparty was, and setting the exposure against the rest of the file, which is again a workspace rather than a black box.
Where the judgement actually sits
Put the five layers end to end and the shape of the work is clear. Collecting the history, resolving it into one vocabulary, filling in what the records do not contain, aggregating it into figures that reconcile, and putting names against counterparties is most of a Source of Wealth review. All of it is an automation task built on gated data sources. The assessment at the end is rarely the expensive part. The argument about whether a model can be trusted with the conclusion is being held over the smallest piece of the file.
Yet, models do have real work here.
- Classification, below the interpretation layer. What a contract call was, which pattern a transaction matches: trained on labelled data, measurable against a known answer, and bought from people who specialise in it.
- Language, at the top. Drafting narrative around facts already established, summarising an analysis already done, and triaging a queue so that the hard files reach a person sooner.
Both are a meta automation, sitting on top of the five layers rather than doing their work. They take repetitive effort away so that the hours go on judgement, not an analyst being replaced.
What is left is the assessment. It turns on one question: do the flows and the investment returns plausibly explain the balance? The answer does not have to be perfect. Ninety percent explained is often a reasonable place to stop. Deciding this, on this customer, at this size, is the part a person is answerable for, and it is the one part of the review that was never a data problem.
That is why the opening question is the wrong one to spend the argument on. When the file is reopened two years later, the institution shows two things: the data underneath every figure, and the reason a person gave. Both outlast whichever model happened to be in use at the time.
One last practical note. If you are dealing with holders who prepare their own evidence, they meet the same five layers from the other side. That is covered in why AI cannot build your source of wealth file, published by CashoutReady, the consumer service ChainComply runs.