It Takes a Village: Building Gusto’s First AI Risk Agent

Xao Yang26 Jan 2026

Problem

Gusto, a payroll and HR platform, scores every new customer with ML models to estimate bad-actor risk. Companies above the approval threshold or with too few signals go to manual review. Onboarding volume was growing faster than the team could review without hiring non-linearly, and slow approvals lowered conversion and delayed first paychecks. The Risk team also could not admit bad actors in exchange for speed.

Example from the article: a prompt told the AI to treat a "low" risk score as a positive signal, but it sometimes wrote that "the low score introduces some risk", so the same word was read two ways.

Solution in brief

A small team of engineers and operations specialists built an LLM agent (GROW) that reviews companies in the manual queue. The prompts were written like code from the operations SOP, with every data field and score range defined. Operations specialists acted as the quality bar, and shadow prompts tested changes without touching production.

Questions this case answers
  • How to automate an expert manual review queue with an LLM agent while keeping the error rate near human level and the downside bounded
  • How to encode tacit domain knowledge and ambiguous data fields into prompts so an LLM interprets them consistently
  • How to cut manual risk reviews of new customers with an LLM agent without letting bad actors through?
  • How to turn operations experts' SOPs and tacit fraud patterns into prompts an LLM follows reliably?

Tools and methods

  • Prompt as specification — explicit field meanings, null handling and score thresholds
  • SOP-to-prompt conversion — an LLM restructures the narrative SOP, then operations specialists rewrite and verify it
  • Shadow prompts — candidate prompts run in parallel with the active one and are promoted when they win
  • Human-in-the-loop review — operations specialists compare AI decisions with their own judgment
Continue reading — it's free

Google or a one-time email code. No password.