Clay's Monty Agent Hit 30-40% Weekly Use by Answering Only Governed Data

Thirty to 40 percent of Clay’s employees now use monty, an internal analytics agent, every week, and 84 percent of new users come back with more questions. Clay described how it built monty, a Slack-based agent that lets employees ask questions of the company’s Snowflake warehouse in plain English instead of writing SQL. The numbers come after two months of use.

The problem it targets is familiar to any RevOps team: data teams are buried in request queues, so only top-priority questions get answered while a long tail of useful ones never does. Clay set four principles. Monty should reason and respond like a trusted data team member, recognize questions it cannot answer and escalate them, query only validated and documented data models, and automatically inherit new models and skills as the data team ships them.

The architecture explains why that works. All analytical data flows through dbt, so the agent never navigates raw tables. The whole codebase, including dbt models, documentation, and skills, sits in monty’s filesystem and updates with each pull request. Hierarchical catalog files reveal context progressively to keep token costs and errors down. A three-layer fallback tries skills first, then pre-calculated metrics, then ad-hoc queries against approved models. The agent is built on Vercel’s eve framework and runs Opus for inference.

Typical users ask two to three questions a week, while power users ask ten or more. Clay argues this raises demand for governed data infrastructure rather than threatening analytics roles.

For sales and RevOps leaders, the lesson is the order of operations. The agent is only as trustworthy as the documented models beneath it, and the escalation rule keeps it from guessing. Before you point an agent at pipeline data, ask which metrics in your warehouse are validated and documented. Those are the ones it can answer safely today.