Consilient.

Here is a fact about anti-money-laundering that ought to be more scandalous than it is: the overwhelming majority of the alerts a bank's transaction-monitoring system produces are wrong. Not slightly wrong. In some trials, more than 90% of flagged transactions turn out to be nothing - a retiree wiring money to a grandchild, a small business paying a supplier. Compliance analysts spend their careers clearing these ghosts, one at a time, while the actual launderers, who have read the same rulebooks the banks use, route around the tripwires. The industry's response, for two decades, has been to spend more money doing the same thing. Consilient's founders looked at that and concluded the whole model was built backwards.

Consilient is a financial technology company, headquartered in New York with an office in Washington, D.C., that launched in October 2020 as a joint venture between two parents you would want in the room for this problem: K2 Integrity, a regulatory and compliance advisory firm, and Giant Oak, a behavioral data-science shop. The premise is deceptively simple. Criminals collaborate across institutions. Banks, hemmed in by privacy law and competitive instinct, mostly cannot. That asymmetry is why financial crime keeps winning, and closing it is the entire point of the company.

The mechanism is federated machine learning, and the one-line version - which Consilient repeats often enough that it functions as a company motto - is share the algorithm, not the data. In a conventional model you would pool everyone's data in one place and train on the pile. You cannot do that with bank transactions; the lawyers would have a collective heart attack, and in cross-border cases it is simply illegal. Federated learning inverts the arrangement. The model goes to the data. It learns from behavioral patterns inside Bank A, carries what it learned - the math, not the names - to Bank B, learns more, and comes back smarter. No customer record ever leaves the building. What travels between institutions is intelligence, stripped of identity.

If that sounds like having your cake and eating it, that is roughly the pitch, and the interesting part is that it appears to work. In bank trials, Consilient reported reducing false positives from north of 90% down to about 12% while increasing the rate at which the system caught genuinely suspicious activity. Those two numbers usually move in opposite directions - tighten the net and you catch more of everything, loosen it and you miss the real thing. Getting both to improve at once is the kind of result that makes a compliance officer look up from their desk.

Most RegTech companies promise to make compliance faster. Consilient is making a subtler bet: that the problem is not speed but correctness. A monitoring system that fires ten thousand alerts and buries the three real ones is not slow - it is wrong, expensively. By training on behavior rather than static rules, and by letting the model keep learning as launderers change tactics, Consilient is trying to build a defense that moves when the target moves. Rules can be reverse-engineered; a continuously-adapting model is a harder thing to game.

The elegance is that the company's central technical choice was forced on it. Consilient could not move bank data across institutional or national borders - so it moved the model instead. The regulatory constraint that would have killed a data-pooling startup is exactly the thing that makes federated learning valuable. Privacy and collaboration, which usually pull against each other, both come out ahead. That is a rare shape for a business to have, and it is worth pausing on: the hardest limitation in the industry is the reason the product exists.

It is early. Eleven people, a $3 million seed, a handful of trials and two patents pending is a modest footprint against a compliance-spend problem measured in the tens of billions of dollars a year. Consilient competes in a crowded field of AI-driven financial-crime vendors, and the question that matters - whether banks will trust models trained partly on their rivals' behavior - is a commercial and cultural one as much as a technical one. But the underlying idea has the quality good ideas tend to have: once someone explains it, the old way of doing things looks slightly absurd.