5 Billion Claygent Runs Later, Clay's Own Growth Team Wrote Down What Actually Works

Five billion Claygent runs is the sample size behind Clay’s own growth team writing down what actually works for AI agents in GTM workflows — not a vendor pitch aimed outward, but an internal playbook the company is publishing as-is. It frames two core patterns: AI research (open-web data gathering) and data transformation (structuring messy internal data), then walks through specific applications built on top of them — account-signal discovery pulled from 10-Ks and job postings, personalized ABM campaigns with dynamic landing pages, competitive-intelligence tracking, deal postmortems assembled from Gong call transcripts, account health and churn scoring, persona classification from inconsistent job titles, and multi-angle outbound generation across insight, case-study, value, and what the post calls “desperation” angles. Claygent Builder lets teams build, test, and deploy these agents in plain language, with Account Agents in Workflows handling automated research inside existing account plays.

The operating guidance underneath the use cases is the more durable part: reserve agents for judgment-heavy tasks, use deterministic formulas for anything rule-based, keep prompts narrow, and keep reasoning visible to cut hallucination risk. On governance specifically, the post notes teams “can see which tables each agent runs in, set credit limits per agent, and cap how much teammates spend” — a lighter-weight, usage-metering approach to agent oversight than the audit-trail-and-evaluation-layer architectures larger platforms like Outreach have published, but one aimed at a different problem: cost and scope control inside a single tool rather than governing autonomous action across a full revenue stack.

The playbook also marks a real shift in what Clay is willing to run through an agent. Its July announcement of open-weight models (Kimi K2.6 and GLM 5.2) in Claygent was framed around cost — cheaper inference lets customers research an entire total addressable market instead of a sample — but this playbook shows what that cost reduction actually unlocked: the volume of use cases listed here, several requiring iterative multi-step reasoning, only pencil out at the lower price point the open-weight shift made possible.