Variance.

Here is an unglamorous fact that everyone in banking knows and no one puts on a billboard: catching the fraud is the easy 10% of the job. A model flags a wire, an alert lights up, a name matches a watchlist. Fine. The other 90% - the part that eats payroll and time - is the investigation. Someone has to collect the evidence, read the documents, figure out whether the flagged customer is a money launderer or just a guy who sells vintage sneakers, and then write it all down in a way a regulator will accept. Variance is a company built on the theory that this 90% is where the money is.

The theory is more interesting than it sounds, because it inverts the usual pitch. Most fraud-detection startups sell you a better alarm. Variance's founders looked at the alarm industry and decided the alarm was mostly solved - what was broken was everything that happened after the alarm went off. "The real work begins afterward," the company says: "collecting evidence, interpreting context, deciding whether a case is part of a broader pattern." That is not detection. That is investigation, and investigation is expensive because, historically, it required a human with judgment to sit down for six hours per case.

Variance's product is a set of AI agents that do the six hours. They gather context, reason over a workflow, analyze documents, and produce an output a human can sign off on. The company is careful - pointedly, strategically careful - to say the agents produce auditable outputs. In a regulated industry, "auditable" is a more valuable word than "autonomous," because a bank does not want a black box that decides who is a criminal. It wants a fast, tireless junior analyst who shows its work.

What makes this fun, from a distance, is that Variance did not start out in banking at all. It started as a company called Intrinsic, and Intrinsic was in a completely different business.

Karine Mellata and Michael Lin met on Apple's fraud engineering and algorithmic risk team, where the job is to build systems that work reliably when the input is the entire population of iPhone users behaving adversarially. That is a specific kind of training. You learn that abuse happens at internet scale, that humans do not scale, and that the only thing that scales is a well-designed system. In 2023 they left to build one of their own, joined Y Combinator's Winter batch, and called the company Intrinsic.

Intrinsic sold trust-and-safety tooling: policy management, moderator workflows, content understanding - the machinery platforms use to keep bad posts, scams, and abuse off their services. It found real customers. The platforms you would expect to care about this - Reddit, Bluesky, the crowdfunding site GoFundMe, the marketplace Redbubble - are the ones publicly associated with the company's early work. In January 2024 it raised a $3.1 million seed round. This was a perfectly good content-moderation startup.

And then the founders noticed something. The exact problem they had built for - investigation at internet scale, sorting the genuinely-bad from the merely-flagged - was crushing a much larger, much better-funded industry: financial services. Banks and fintechs run enormous compliance operations. They do KYC (Know Your Customer) and KYB (Know Your Business) at onboarding, they monitor transactions for money laundering, they investigate fraud, and they staff all of it with analysts because the regulators require a human in the loop and a paper trail. It is the same shape of problem as content moderation, except the stakes are measured in fines and prison sentences rather than deleted posts.

So they pivoted. Intrinsic became Variance. The rename is itself a small tell - you change your name when your customer changes, and Variance's customer went from a platform's trust-and-safety lead to a bank's Chief Compliance Officer. The word "variance" is a statistics term for how far the data strays from what you'd expect, which is a reasonable thing to name a company whose product is finding the transactions that don't fit.