The best job candidates already have jobs. Finding them used to eat Curtis Key’s team alive, hours of manual searching every single day. Now an AI candidate sourcing setup built on Claude and Zapier MCP does the hunting, and it saves his recruiting agency more than 240 hours a week.
What Blue Castle actually built
Curtis Key has been building automation for about 20 years. First with scripts, then Salesforce, at a real estate brokerage that grew from 8 agents to 105. When hiring became that brokerage’s hardest problem, he turned it into a business: Blue Castle Agency, a recruiting firm just for real estate and property management.
His whole philosophy fits in one line. The best candidates aren’t looking. They’re already employed somewhere else, so a recruiter has to find them and earn a conversation. That search is repetitive, slow work, and repetitive slow work is exactly what AI handles best.
So Blue Castle gave it to Claude. Several times a day, Claude pulls active jobs from the agency’s applicant tracking system. It reads the company’s success profiles, the notes for each segment (multifamily, associations, affordable housing, luxury), then uses the agency’s own sourcing tools to build a candidate list. No human touches the process until the list is done.
Then comes the part most people skip. A second Claude run reviews that list later in the day, with the same context a human reviewer used to have. Two AI passes, one doing, one checking.
The numbers: human sourcers vs Claude
Here’s where it gets interesting. The agency compared its own people against the AI candidate sourcing workflow, and the results were not close.
| What gets measured | Human sourcers | Claude setup |
|---|---|---|
| Approval rate on candidate lists | 78-83% | About 94% |
| Time saved (sourcing + quality control) | Baseline | 240+ hours per week |
| How recruiters start the day | Searching from scratch | With a list already built and checked |
| Submission workload per recruiter | Full manual | 4+ hours a week returned |
Read that approval rate again. The AI’s lists get approved more often than the humans’ lists. Curtis expects human candidates to be approved 78 to 83 percent of the time. Claude’s come in around 94 percent.
That’s not because Claude is smarter than his recruiters. It’s because the AI follows the success profile every single time, never gets tired on a Friday afternoon, and never skims.
There’s a second automation too. When a recruiter is ready to submit a candidate, a separate process reads their notes and the Zoom transcript, checks everything against the job profile, scores the candidate, researches the companies on the resume, and writes the submission. Curtis estimates it hands each recruiter back 4+ hours a week.
Why Zapier MCP is the missing piece
Here’s the thing most “AI will change recruiting” articles leave out. Claude on its own can reason about a candidate all day long. It cannot reach into an applicant tracking system, a sourcing platform, and a Google Sheet by itself.
That’s the gap Zapier MCP fills. MCP stands for Model Context Protocol, an open standard created by Anthropic, and it’s the plumbing that lets Claude operate real software. Zapier’s version connects AI tools to around 9,000 apps, so Claude can read your ATS, push rows to a spreadsheet, and run your sourcing platform without anyone writing custom code.
Curtis put it simply: it all comes down to context. The technology already exists. The win comes from giving the AI understanding of your business, and the connectors make that possible.
His favorite part is scaling. When the agency needs more sourcing capacity, nobody starts an internal hiring process. They just turn the knob. For a small firm competing with big recruiting shops, that’s a structural advantage, not a gadget.
How to copy this AI candidate sourcing setup
You don’t need Curtis’s 20 years of automation experience to steal this pattern. You need one repetitive list-building task, and an afternoon.
1. Pick one repetitive workflow
Don’t start with your whole business. Pick the single most boring recurring task you have. Sourcing candidates is the recruiting example, but the same shape works for lead research, invoice chasing, inventory reordering, or vetting inbound requests. The pattern is always: pull data, apply judgment, produce a list.
2. Connect Claude to your tools
Set up Zapier MCP from zapier.com/mcp, then add it as a connector in Claude’s settings. MCP is an open standard, so this same pattern works with other MCP clients too. Start with read-only permissions: let Claude look at your ATS or your spreadsheet before you let it write anything anywhere.
3. Feed it your context first
This is the step people rush, and it’s the one that decides everything. Blue Castle stores success profiles for the company and every role, plus notes per segment. Claude isn’t guessing what a good candidate looks like. It’s reading your definition of good.
So before you connect anything, write down:
- What an obviously great candidate, client, or lead looks like for you
- Your instant disqualifiers, the things that make you pass on someone
- Segment notes, because luxury buyers behave nothing like volume buyers
- Two or three past wins you’d want the AI to repeat exactly
Write that down before you automate anything, or the AI will automate mediocrity for you at impressive speed.
4. Run a second pass for quality control
One AI pass generating output is a draft. A second pass checking that output against the same context is a system. This is what took Blue Castle’s approval rate past their human baseline, and it costs nothing but a second scheduled run. Schedule it later in the day so there’s a clear gap between doing and checking.
5. Keep a human on the final call
Notice what Blue Castle did not do. No candidate gets contacted without a recruiter deciding to contact them. Claude builds and checks the list; people still make the call. That’s the right division of labor for anything involving judgment about other humans.
What it costs
Less than you’d guess. Zapier MCP runs on your existing Zapier plan with no separate billing, and it’s available on every plan including the free tier. Each tool call through MCP costs two Zapier tasks, so the free plan’s 100 monthly tasks gets you roughly 50 AI tool calls to experiment with. Paid plans add task volume from there.
On Claude’s side, running AI candidate sourcing is subscription territory, not API territory. A Claude Pro seat at $20 a month handles personal experimentation. Claude Team, which is what an agency like this would actually run, costs $20 per person per month on annual billing ($25 monthly), with a two-seat minimum, and every seat includes connectors and admin controls. A two-person shop is into this for well under $100 a month between both tools.
Compare that to the 240 hours a week on the table and the math stops being interesting. It’s just yes.
The catch nobody mentions
Setup is the real price. Getting Claude connected to your ATS, your sourcing tools, and your sheets took Curtis real iteration time, and his two decades of automation experience made the build faster than it would be for most people. Your first month will be tuning, not harvesting.
The second catch: this only works because the context was written down. If your company’s knowledge lives in three people’s heads, the AI has nothing to work with. That’s not a tech problem, but it will block you like one.
If hiring is your bottleneck but sourcing lists aren’t, start somewhere else entirely. We covered which HR tasks to hand to AI first in stop letting AI hire people, automate the boring parts, and the connector setup overlaps heavily with what Claude Marketplace is and how to use it. Prefer a phone-based starter? These Gemini automations on Android follow the same logic at a smaller scale.
Your first move
Open your calendar and find the most repetitive list you build. That’s your candidate for an AI candidate sourcing workflow, whether it’s candidates, leads, or vendors. Connect one tool, feed the AI your standards, add the second checking pass, and keep the final decision human. Curtis’s numbers say the pattern pays. Start with the free plan and prove it on your own data before you spend a dollar.