Key Moments

This AI Company Catches Fraud Across the Internet

Y CombinatorY Combinator
Science & Technology5 min read32 min video
Mar 31, 2026|2,693 views|109|8
Save to Pod
TL;DR

Variance uses AI agents instead of human analysts to detect fraud and ensure compliance for major companies, but this automation comes at the cost of keeping their advanced techniques secret from the public.

Key Insights

1

Variance automates fraud detection, content review, and identity verification for Fortune 500 companies and platforms like GoFundMe using purpose-built AI agents.

2

The company has been operating in stealth for three years, processing petabytes of data and identifying sophisticated fraud rings, including state-sponsored actors.

3

Variance's AI agents can process unstructured data from various sources, including scraping information directly from customer UIs built for human interaction.

4

The company attributes its success to AI agents that can replace rule-based systems, classifiers, and human analysts, enabling a fully self-healing and dynamic fraud detection system.

5

A key technical challenge for Variance was consolidating and processing unstructured data, often scattered across 5-10 different systems and sometimes hidden behind human-facing dashboards.

6

Variance is highly focused on leveraging AI coding agents, with a small team of five engineers producing software output comparable to a 25-person team.

AI agents tackle complex fraud and compliance at scale

Variance is emerging from stealth after three years of development, announcing a $21 million Series A funding round. The company builds purpose-built AI agents designed to automate risk and compliance processes, including fraud detection, content review, and identity verification, for large enterprises and platforms like GoFundMe and Fortune 500 companies. These AI agents analyze fundraisers for legitimacy, ensuring money goes to the intended recipients and not to sanctioned countries or fraudulent causes, a task previously handled by human analysts but now automated for greater consistency and scale. Variance operates discreetly, often dealing with sensitive data and aiming to equip 'good guys' with tools previously used by adversaries, thus maintaining secrecy to avoid creating more opportunities for fraud.

Detecting fraudulent fundraisers on GoFundMe

A specific example highlighted is GoFundMe, a platform where fundraisers for crises often spike. Variance's AI agents are tasked with distinguishing genuine fundraisers from fraudulent ones. For instance, following a public tragedy, numerous fraudulent fundraisers might appear, claiming to be from the victim's family. Variance agents use behavioral signals, identity information, past account activity, and fundraiser details (like images and bios) to determine legitimacy against GoFundMe's terms of service. This automated review process validates every fundraiser request before it goes live, a significant shift from manual analysis.

Automating identity and business verification

Variance also powers identity verification for marketplaces and gig economy platforms. When individuals sign up as, for example, delivery drivers, their identities are verified using selfies and driver's licenses, with AI agents reasoning over this data against company standards. Furthermore, Variance handles complex Know Your Business (KYB) verifications, which are crucial for businesses operating online or with financial institutions. This involves confirming if an entity is truly linked to the business it claims to own, identifying potential shell companies, connections to sanctioned countries, or individuals with adverse media records like those involved in money laundering. These complex investigations are typically manual and time-consuming.

Leveraging diverse data sources with AI

To build these AI agents, Variance utilizes three core building blocks: compliance documents (standard operating procedures), tools they've developed, and data. This data comes from both internal customer sources and external databases, including hundreds of global business registries and the open web. Access to the open web has been particularly crucial, mirroring how human analysts would use search engines to gather context on individuals or entities. Variance excels at connecting to disparate data sources, pooling unstructured data into their own stores, which is essential for building comprehensive risk graphs.

The challenge of unstructured and scattered data

A primary technical hurdle for Variance has been managing and processing vast amounts of unstructured data (petabytes) from diverse sources. This data is often scattered across 5-10 different internal systems with no defined schema. In some cases, data is only accessible through human-facing dashboards or UIs, requiring Variance's AI agents to scrape this information directly. The ability to consolidate this scattered, unstructured data and then reason over it has been a significant engineering achievement for the company.

Evolution from deterministic systems to agentic AI

Previously, fraud detection relied on a patchwork of deterministic systems, including rule-based engines (e.g., 'if transaction over $1,000, then do X') and specialized classifiers for specific abuse types, supplemented by human analysts who provided contextual reasoning. However, these systems had slow feedback loops and lacked adaptability in dynamic environments like fraud. Variance's AI agents, by contrast, can materialize any feature a rules engine could, reason over unstructured data and images to identify fraud types without specialized classifiers, and replace human reasoning entirely. This creates a fully self-healing system that can evolve rapidly, enabling companies to ship faster and launch new product lines without bottlenecks.

Detecting sophisticated fraud rings and misinformation

Variance agents have identified complex fraud rings, including those orchestrated by state-sponsored actors during elections by analyzing entities' relationships within broader networks. This sophisticated detection goes beyond isolated content analysis. Moreover, their ability to detect misinformation and harmful content can have significant real-world implications, potentially preventing physical violence by identifying threats and plans at scale, leading such findings to be reported to law enforcement. This capability demonstrates the profound impact of their technology on societal safety.

Lean team leveraging AI for massive output

Despite its impressive capabilities, Variance maintains a lean team of 12, including five software engineers. The company operates with a strong ownership culture, treating engineers almost as managers of small AI agent teams. This approach, coupled with AI coding agents, allows them to achieve a software output comparable to a 25-person team. Even non-technical staff, like customer success managers, can now use AI agents to handle feature requests autonomously, shipping new functionalities within hours without direct engineering involvement.

Common Questions

Variance builds purpose-built AI agents that automate content review, fraud detection, and identity verification at scale for risk and compliance. They power major companies to handle sensitive data and complex reviews.

Topics

Mentioned in this video

More from Y Combinator

View all 578 summaries

Found this useful? Build your knowledge library

Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.

Get Started Free