How Cleo puts AI agents to work on governed company data

About Cleo

Cleo builds an AI financial assistant that helps people in the US and UK budget, save and borrow. Its data sits in a Redshift warehouse of about 4,000 tables, used by fraud, payments, marketing, growth, support and operations.

A 4,000-table warehouse and a big data queue

Using the warehouse took knowledge of its tables. Fraud, marketing, engineering and the business teams all waited on analysts for answers.

Fraud and disputes analysts spent 30 to 60 minutes assembling context before they could decide a case. Engineers wanted to query the warehouse without leaving Cursor or Claude Code.

A fixed semantic layer would not have been enough. The questions were often ambiguous, and answering them took context beyond metric definitions, from across the whole warehouse.

Cleo put its definitions under the tools people already use

Cleo’s data team chose to make the warehouse safe to use from anywhere. It certifies metrics and table descriptions on its own Redshift in Rig, and staff reach them from Claude and Cursor.

Every query is tested against those definitions and Cleo’s usage rules before it runs. The data team reviews what gets certified, and the rest of the company answers its own questions.

Analysts hand whole investigations to agents

Since August, a Growth Specialist has run investigations through Claude agents. The agents work through Rig’s governed layer. They use certified metrics and table descriptions, and test each query before running it.

Several agents often work on one question at once, each taking part of the research. In 30 days they tested queries 2,632 times and ran 1,309, and on the busiest day they made 794 requests.

Recent work includes swings in day-zero premium conversion, cash advance take-up, income-classifier accuracy and rebuilding a past experiment.

Every team builds on the same definitions

The same governed layer sits under the apps Cleo’s teams build for themselves. Finance and disputes, marketing, operations, go-to-market and customer support each run their own alerts, reviews and digests on it.

Cleo’s payments analytics team built a daily check that flags unusual payment failures by route and posts them to Slack each morning.

A quarter of Cleo’s requests now come from agents

Agent requests rose from 6,973 in August to 11,204 in September. Over the same two months, total use grew from 33,828 to 41,136 requests, from 137 people in September.

Most of that use runs through Claude and the other tools people already work in, and a shrinking share through Rig’s own chat and app screens. Governed context is what made that safe. Each question reaches the right definitions and permissions on the right tables, wherever it is asked.