The conversation about AI in Recruiting usually starts with the spectacular: automated screening, matching, interview summaries, generated scorecards, "perfect" outreach.

The problem is that if your process is messy, AI won't fix it. It amplifies it. And that can get expensive.

  • Expensive in time (because you automate errors).
  • Expensive in quality (because you reject the wrong people).
  • Expensive in reputation (because candidate experience degrades).
  • Expensive in internal credibility (because the business stops trusting Recruiting).
  • Expensive in tokens!
Step zero isn't choosing a tool. It's getting the system in order.

1) Map the real process (one page)

Before adding AI, build an end-to-end map of how you hire today:

  • Opening: who requests the req, who approves, what inputs exist.
  • Attraction: where talent enters, from which recruiting sources.
  • Evaluation: stages, interviews, tests.
  • Decision: who decides and with what evidence.
  • Close: offer, negotiation, signature.

Add two layers:

  • Tools (ATS, LinkedIn Recruiter, scheduling, tests, parallel spreadsheets, etc.)
  • People (recruiter, hiring manager, interviewers, leadership, legal/finance if applicable)

If the map shows spreadsheets "because the ATS isn't enough," duplicated steps, or fuzzy approvals, you've already found friction.

2) Source-of-hire traceability (without this, you're flying blind)

Measure performance across at least 4 sources:

  • Applicants (inbound)
  • Referrals
  • Direct sourcing (cold candidates)
  • External / agencies

Look at conversions, not just volume:

  • % who reach interview
  • % who reach final
  • % who receive an offer
  • % who accept

And if you can: some quality signal (HM feedback, early tenure, initial performance). Without this, it's easy to "optimize" toward the wrong source.

3) Bottlenecks: where value gets lost

Look for three things:

  • Stages where delay accumulates (aging by stage)
  • Sharp drops in the funnel (conversions)
  • Variability: similar roles with totally different timelines

Aging helps, but only when interpreted with context: not every req should take the same amount of time. What matters is spotting avoidable delays (decision waits, redundant stages, confusing handoffs).

4) Interview and decision audit

AI builds on what you give it. If evaluation is inconsistent, automation amplifies inconsistency.

Quick check:

  • Do you have scorecards, or just "I liked them / I didn't like them"?
  • Are you evaluating what was agreed in kickoff, or does each interviewer invent their own criteria?
  • Is evidence documented, or does everything live in a conversation?

5) Stack inventory: what we pay for and why

Make a list with 4 columns:

  • Tool
  • Estimated cost
  • What it was bought for
  • What it's used for today

This exercise usually reveals paid features nobody uses, duplicate tools, and parallel processes. If you're already paying for an expensive stack, first make it work with intention.

Guiding principle: don't automate steps that shouldn't exist

Elon Musk said Tesla went too far with factory automation and that it was a mistake: automating for automation's sake can make the system worse. The same happens in Recruiting: if a stage doesn't add value, automating it just makes it faster—and more harmful.

"Step Zero" checklist (7 days)

  1. End-to-end map on 1 page
  2. Defined stages + advance criteria per stage
  3. Performance by source (the 4 minimums)
  4. Conversions by stage (even if imperfect)
  5. Tool inventory + real purpose
  6. Scorecards and evidence-based decisions
  7. Eliminate 1 step that doesn't add value
After this, AI stops being a toy and becomes leverage.