A recruiting agency gave sourcing to Claude through Zapier MCP. Its lists get approved about 94% of the time, against 78 to 83% for humans

The most useful detail in this case study is not the 94%. It is the plan to take work away from the AI.

What happened

Zapier published a case study on Blue Castle Agency, a recruitment firm for real estate and property management run by Curtis Key. Several times a day, Claude picks active jobs from the firm's applicant tracking system, reads the success profiles and segment notes for each role, and uses the firm's sourcing tools to build a candidate list. A second Claude run checks that list later in the day, the quality control step a human used to do.

Curtis says human sourcers' candidates are approved 78 to 83% of the time, while Claude's are approved about 94% of the time. He estimates the sourcing and quality control automations save the team 240+ hours a week, and a separate submission process that reads recruiter notes and Zoom transcripts saves each recruiter another 4+ hours a week. Zapier MCP is how Claude reaches the ATS, sourcing platform and Google Sheets without a custom integration for each.

The firm has started building Next-Gen Zaps too. One runs at 2 a.m. on weekdays to build prep sheets for screening calls and post a morning note in Slack.

My take

Three things here map directly to client work.

First, context did the heavy lifting. Success profiles and segment notes are what made the output good. Most teams skip writing those down, then blame the model.

Second, the checker is a separate pass. Generating and reviewing in the same call tends to approve its own work. A second run with the same context is cheap insurance.

Third, and most important: Curtis is now asking where Claude repeats work a Zap would do the same way every time, and moving those steps into Zaps. That is the right instinct. Use the model for judgment calls, use deterministic steps for everything that never changes. It is cheaper, faster and much easier to debug.

These numbers are the agency's own estimates, so treat them as a direction, not a benchmark. The architecture is the part to copy.