Building Stripe's LLM-Assisted Response Workflow
Identified a repeated communication bottleneck and independently built an LLM workflow to eliminate it.
Pooja Naik
Reserach and Investment Analyst at TresVista
From their time as
Risk Operations Associate
Stripe • 2024 - 2025
Overview
Pooja noticed that a particular type of investigation at Stripe involved repeated, templated communication at every step of the workflow. Drafting and sending these emails to merchants and financial partners was consuming several minutes per case, and the pattern was consistent across the team.
The Story
Pooja noticed that a particular type of investigation at Stripe involved repeated, templated communication at every step of the workflow. Drafting and sending these emails to merchants and financial partners was consuming several minutes per case, and the pattern was consistent across the team.
Rather than flagging it as a problem, she decided to investigate whether she could solve it herself first. She spent time mapping the existing process, identifying which parts were repetitive, and understanding the range of scenarios the team handled regularly. Having been in the role for seven to eight months, she had a strong grasp of the different query types and what a good response looked like for each.
She then built an LLM-assisted response workflow by training the model on real scenarios and expected outputs. The approach was straightforward: given a merchant ID and a scenario type, the LLM would generate a ready-to-send email. The team could proofread it once and send it directly, with no drafting required.
The build took two to three days. She tested the approach across multiple scenario types, refined the outputs, and made sure the workflow was practical enough for the full team to adopt. The result was a reduction in average handling time of five to seven minutes per case across everyone using it.