Bringing AI into recruiting without handing it the decisions

Hiring quality shouldn’t depend on how much time one recruiter has that week.

Every search needs the same things. A sourcing string. A scorecard. An interview guide. A candidate assessment. Prep emails for candidates. Outreach that doesn’t read like outreach. A weekly update so the hiring manager isn’t chasing me.

When one person owns hiring for a company of about 190, across three regions, those things get done in proportion to how busy that person is that week. Senior commercial roles, product roles, and technical contractor searches all run at once. The quality of a search shouldn’t depend on my calendar.

A toolkit that starts every search at the same standard

The foundation is a set of Claude Projects that act as a persistent operating system for recruiting.

Recruiting Hub holds each live role and produces its toolkit: sourcing strings, assessments, interview guides, and candidate prep emails.
Pipeline Update turns the state of a search into the update a hiring manager actually reads.
Scorecard generators for recurring role families, so a developer search and a Head of Accounts search each start from a scorecard built for that job, not a generic one.
Outreach in my own voice. Sequences are drafted from how I actually write to people, not from a template that sounds like everyone else’s.

Everything sits on top of the stack we already used: Greenhouse, LinkedIn Recruiter, Google Workspace, Zapier, and Slack.

One intake conversation Senior Product Manager Team, must-haves, deal-breakers, timeline

Sourcing stringsBoolean and LinkedIn searches for the role
ScorecardBuilt for the role family, not generic
Interview guideSame criteria for every interviewer
Candidate prepWho they meet and what to expect
OutreachIn my voice, not a template
Pipeline updateFor the hiring manager, every week

Every candidate-facing piece is read and approved by a person before it goes anywhere.

What one intake produces. Run end to end by our recruiter, not just me. Fictional role.

AIDraftSourcing strings, scorecards, prep notes, and the right message for each stage.

A personReviewEvery candidate-facing draft lands in Slack: approve, edit, or send.

A personActThe message goes out, from a recruiter, in their own voice.

The boundary

Below this line, only a person acts: declines, offers, and anything that changes someone's outcome. A decline can't be approved in one click.

How the recruiting workflow runs today. The Slack approval step is live.

The rule that shapes everything: AI drafts, a person approves

On top of the toolkit sits a workflow with a trigger, an output, and a gate. Candidate communication is the clearest example. A stage change in the ATS triggers a draft of the right message for that stage, which lands in Slack to approve, edit, or send.

Most stages can be approved in one click. Declines can’t. Any message that tells someone a decision about them requires a person to read it, change it if needed, and send it. That’s a candidate experience decision as much as a risk one. A decline is the one message a candidate rereads.

See the approval step in Slack

#recruiting-approvals Illustrative example, fictional data

Draft ready9:41

Interview confirmation, Senior Developer candidate

Hi Jordan, thanks for making time. You're confirmed for Thursday at 10:00 with the engineering team. Here's what to expect...

Approve and sendEdit

Decline drafted9:44

Product Manager candidate, final round

Needs a person. Read it, make it yours, and send it yourself.

Open and editSend without reading

How the approval step works in Slack. Routine messages can be approved in one click. A decline can't be sent until a person rewrites or reads it in full.

The same principle runs through screening. The screening assistant writes a pre-read for the recruiter, never a verdict. A human always reviews before anyone advances or is declined.

The real test: it runs without me

A system that only works when its builder runs it isn’t a system. So I trained our recruiter to run the whole toolkit independently: pipeline management, scorecard drafting, candidate communications, and translating interview feedback from Spanish to English. Our technology and product hiring teams now use the structured scorecard and interview workflow too. That handover was recognized as an L2 AI Spotlight, and it’s the part I’m proudest of. The toolkit stopped being mine and became the team’s.

What changed

Candidate review, down from 15 to 20 minutes: Under 2 min
Batch review rounds, down from 1 to 2 hours: 10 min
Scorecard turnaround, down from 3 to 5 business days: 24 hrs

Before and after the toolkit. Zero scorecard follow-up reminders needed since rollout.

Every search now starts from the same standard, whoever is running it and however busy the week is. Hiring managers get updates they don’t have to chase. And the work that needs judgment, talking to candidates and making the call, gets more of the time.

What I’d do differently

When Orium’s leadership reviewed it for the AI Spotlight program, the point they called out wasn’t speed. It was that the time I’d saved for myself was now saving time for other people too. That’s the test I use now for anything I build.

I’d write the approval rules before writing a single prompt.

The prompts were easy to iterate. The rules about what a person must see, and what can go out on its own, were the decisions that mattered. Settling them first would have saved a round of rebuilding.