The trust problem.
Most early-stage CRMs are 30% useful and 70% theatre. Stages are wrong, contacts are duplicated, "next steps" hasn't been updated in weeks. The founder spends Monday morning re-asking each rep "so where is this actually." The fix isn't more discipline — it's removing the discipline tax.
The 4 jobs the agent does.
- Dedupe. Same company, three deal records — merged based on domain match.
- Stage validation. A deal in "Proposal" with no proposal sent in 14 days gets flagged and proposed for downgrade.
- Auto-summarize. Claude reads the last 10 emails on the thread and writes a 3-bullet "where this stands."
- Score. Each deal gets a freshness × engagement × ICP-fit score, sorted weekly.
The build.
Clay polls HubSpot nightly via the deals API, runs the dedupe + validation rules, and pushes back the cleaned-up record. Claude runs as a separate column for the email summarization (read-only — never auto-writes back). The weekly digest lands in a Notion DB on Sunday night, ready for Monday's pipeline review.
What it gives the founder back.
Pipeline review goes from 90+ minutes of "wait, what's the status of X" to 20 minutes of "approve or downgrade these flagged stages." The forecast becomes defensible because every deal has a fresh, AI-written status note attached.
The Clay table, validation rules, and Claude summary prompt ship in the weekly. Browse all workflows in the loop.