How Variance built AI agents to catch fraud at scale

Executive overview

Most fraud systems rely on rigid rules, narrow classifiers, and slow human analysts — a patchwork that can't keep up with adversaries who adapt constantly. Variance replaces that patchwork with purpose-built AI agents that automate fraud detection, identity verification, and content review end-to-end.

Three building blocks drive the agents: compliance documents (standard operating procedures), tools, and data — internal and external. The result is a self-healing system that closes the feedback loop faster than any human team.

The key insight: AI agents replace the entire deterministic stack — rules engine, classifiers, and human reasoning — with a single adaptive layer.

What Variance does and who uses it

• Automates content review, fraud review, and identity verification for Fortune 500s and large marketplaces
• GoFundMe uses Variance agents to validate every fundraiser before it goes live — checking identity, sanctions exposure, and terms-of-service compliance
• Gig economy platforms use it to verify delivery driver identities against selfies and licence photos
• KYB (Know Your Business) reviews map complex ownership graphs across shell companies, multiple agents, and cross-border sanctions risks
• All of this work was previously done by human analysts; Variance makes it fully automated and consistent

The technical architecture

• Three building blocks: compliance documents (SOPs), tools, and data (internal + external)
• External data includes 100+ business registries worldwide and open-web access — web access was the missing piece that previously required a human to Google
• Customer data is often scattered across 5–10 systems; Variance pulls it via reverse ETL, API, or by spinning up a browser agent to scrape legacy internal tools built for humans
• Petabytes of unstructured data are ingested and unified into Variance's own data stores before agents reason over them
• No specialised classifiers needed — agents read SOPs and reason over images, text, and graph relationships directly

Why LLMs made this problem solvable now

• Legacy fraud stacks: rules engines (static), classifiers (narrow), humans (slow, inconsistent) — no true feedback loop
• AI agents materialise features on the fly, reason across entity graphs, and chain tool calls — the loop closes automatically
• During the 2024 elections, agents detected state-sponsored fraud rings pushing coordinated narratives — impossible with isolated classifiers
• Investigations have surfaced credible physical-harm threats, with findings handed to law enforcement
• GPT-4 launched during the YC batch; model improvements mid-pilot cut costs 10x and improved performance dramatically

Founding story and company culture

• Co-founders Karine Mellata and Michael met at Apple on the centralised fraud engineering team; she was a data engineer, he a machine learning engineer
• Built in stealth for three years — clients deal with sensitive fraud vectors and don't want their defences disclosed
• First customer IAC (Care.com, Angie, Ask Media Group) — took eight months to land; used Variance to automate compliance review of marketing content previously outsourced to a BPO
• In July 2024, at peak growth with revenue doubling month-on-month, Karine was hit by a truck and hospitalised for 10 days with a broken spine and leg
• Team of 12 (five engineers) operates with an AI-coding-maximalist culture; software output equivalent to ~25 people; even a non-technical CSM ships features via Cursor autonomously
• Series A: $21 million, announced March 2026