The agentic fraud and AML investigation platform. Its AI agents read the documents, cite the evidence, and hand risk teams an audit-ready verdict.
Not by an analyst with sixty browser tabs and a lukewarm coffee. By an agent that has already read the dispute PDF, cross-checked the transaction log, glanced at the merchant's website, and written down why it decided what it decided.
Picture the compliance floor of a large bank at 9 a.m. Thousands of alerts sit in a queue. Each one is a small trial: gather the evidence, weigh it, decide, and - this is the part that matters - document it well enough that an examiner nodding over your shoulder in eighteen months will agree. The unglamorous truth of financial crime is that someone has to read all of it, so that nobody launders money through the gaps.
Roe AI's wager is that most of that reading can be done by software, and done with receipts. Its agents reason across the messy 80% of enterprise data - the PDFs, screenshots, recordings, and web pages that a traditional data warehouse politely ignores - and produce what the company calls an audit-ready disposition. Cited. Logged. Reproducible.
That is a stranger sentence than it sounds. Most AI at work today writes emails and summarizes meetings, tasks where being confidently wrong costs nothing. Fraud and anti-money-laundering is the opposite corner of the room: a wrong answer moves real money, and "the model said so" is not a defense that survives a regulator.
So Roe built the boring part first. Every case carries an immutable, per-case audit trail. The agents follow the customer's own standard operating procedures rather than improvising. Deployment is single-tenant, inside the customer's own cloud. The cleverness is deliberately hidden behind a paper trail - an agent that investigates like a good analyst and documents like a nervous one.
Figures published by Roe AI; treat as company-reported, not independently audited.
By the following week, Richard Meng had founded Roe AI. He had been the Gen AI tech lead at Snowflake, and before that had worked on Knowledge Graph and skill assessment at LinkedIn - a resume built around teaching machines to understand messy human data. His co-founder, Jason Wang, came from the infrastructure trenches of Meta, Robinhood, and Retool, the kind of engineer who has architected inference systems at a scale most people never see.
Their first product was not a fraud platform at all. It was a next-generation data warehouse: a way to point SQL at your documents, images, websites, and video and query them like an ordinary database table. Ask a question in plain language, get an answer from a pile of PDFs. It was clever, and it was general, and being general was the problem.
Because when they watched where customers actually felt pain, one industry kept raising its hand: financial services. Risk and compliance teams were buried under exactly the kind of unstructured evidence Roe could read - and they had the budget, the regulatory pressure, and the offshore review costs that made a better answer worth paying for. So Roe narrowed. The general tool became a specific one. The pivot is a small master class in listening to your users instead of your pitch deck.
Each agent takes a slice of the investigation floor. Together they aim to automate the high-volume work and accelerate every human escalation - not replace the judgment behind it.
Works cases end to end across structured data plus PDFs, images, sites and recordings, returning cited evidence and an audit-ready disposition.
Runs first-line anti-money-laundering reviews aligned to your SOPs, with immutable per-case logging built in.
Drafts customer responses and manages the appeals and dispute correspondence that pile up behind every decision.
Assesses merchant compliance and onboarding risk for marketplaces and platforms.
Reads across case volume to surface emerging fraud patterns and new attack types before they scale.
Normalizes data across disparate systems so every other agent can actually reason over it.
Underneath it all sits the original engine: SQL + natural language over multimodal data, connected to S3, Snowflake and Databricks.
"The next-generation data warehouse that uses AI to process unstructured data - natural language, documents, images, and structured tabular data."
- ROE AI, on the idea that started it all
Richard Meng leaves his Gen AI tech-lead role at Snowflake; Roe AI joins Y Combinator's W24 batch.
Announces a $3.5M seed round led by Gradient Ventures to build an AI-powered warehouse for unstructured data.
Sharpens onto financial services - repositions as the agentic fraud and AML investigation platform with per-case audit trails.
Reports deployments with a median time-to-decision of about one week across fraud, dispute and AML reviews.
Roe plays in the crowded, high-stakes lane of financial-crime tooling - up against Sardine, Nasdaq Verafin, NICE Actimize, Hummingbird, and ComplyAdvantage. But its loudest competitor isn't software at all: it's the offshore review floor, the room full of humans reading documents by hand that Roe is trying to make faster, cheaper, and easier to audit.
Back on that compliance floor, the queue is shorter.
The 9 a.m. pile of alerts is thinner now, and the ones that remain are the hard ones - the escalations that deserve a human. The reading has been done. The evidence is cited. The paper trail is already written. Roe AI didn't replace the analyst. It handed them back the part of the job that needed a person.