A three-person startup used Claude and Zapier to send 19,569 expo emails with a 0.37% bounce rate. The checks did the real work

A three-person hiring startup ran expo-scale outreach with Claude and Zapier, and every step that mattered had a check in front of it.

What happened

Zapier published a customer story on Vetano, a three-person hiring platform founded by Chris Fairley. Its welcome emails already ran through Zapier: each sign-up landed in Slack, a Zap read the message and sent the right email. When the product changed, Fairley estimated two to three hours of careful edits, and he remembered an earlier filter mistake that emailed companies who should not have received it.

He connected Claude to Next-Gen Zaps, which Zapier unveiled at ZapConnect 2026. You describe a workflow in an MCP-enabled tool like Claude and deploy it to Zapier. Now Claude builds a new Zap for each change, the old one is switched off and kept, and a test email goes to Fairley's inbox for approval before anything goes live. The rebuilt flow skips test accounts, validates addresses with ZeroBounce, never welcomes anyone twice, and still sends if ZeroBounce is down.

For the Small Business Expo in New York, Vetano contacted nearly 9,000 registered businesses. Claude scored each weekly export from T0 to T4 against a master Google Sheet. Sending ran on 37 mailboxes across 13 domains in Instantly, and a second agent flagged any mailbox above a 3% bounce rate. ZeroBounce kept 2,874 dead or risky addresses off the list. Of 19,569 emails sent between September 29 and October 14, only 73 bounced. Replies post to Slack, Claude drafts an answer, and Fairley sends it. Interested leads go into HubSpot right away. On expo day the welcome Zaps ran 71 times with zero errors, about 12 seconds per sign-up.

My take

The headline numbers come from the boring parts. Validation before sending, a duplicate check, a fallback when a vendor is down, a test email before go-live, and the old version kept instead of deleted. Those are the same controls I put into client builds, and they matter more once an AI is the one making edits.

Notice where the human sits: on replies and on going live. The AI drafts and builds, a person approves anything a customer sees. If you want AI to edit your automations, start by deciding which steps need a test run and a sign-off, then let it work inside those lines.