Hiring is where most managers leak the most time and make the most expensive mistakes. AI won't fix bad judgment, but it will pull the slow, error-prone parts of the loop — drafting, screening, debriefing — into the background so you can spend your attention where it actually matters: talking to humans.

Start With a Job Description That Filters

Most job descriptions are recycled from a half-decade ago and read like a wishlist crossed with a legal disclaimer. Candidates skim them, qualified people self-select out, and unqualified people apply because the criteria are vague. Fix the input and you fix the funnel.

Give the model the real shape of the role — not a title, but what someone does on Tuesday morning.

Draft a job description for a [role]. Here is the actual work:
- Top 3 outcomes in the first 6 months
- 3 things they'll do weekly
- 2 things this role does NOT do (so we filter out mismatches)
- Must-haves vs. nice-to-haves (be strict — 5 must-haves max)
- Salary band, location/remote policy, team they'll join

Write it in plain English. No "rockstar," no "ninja," no "passionate about
synergies." Optimize for clarity, not for SEO.

Then run it back through the model with a second pass: "Cut anything that sounds like every other job post. Flag any requirement that would screen out a strong candidate from a non-traditional background." You'll be surprised how often you were quietly gatekeeping on a credential you don't actually need.

Build the Rubric Before You See a Single Resume

This is the step almost everyone skips, and it is the single most important thing you can do for fairness and signal. If you don't know what you're scoring on before you meet candidates, you will score on vibes. Vibes are how bias gets in.

Write a rubric with 4–6 dimensions, each on a 1–5 scale, with concrete anchors at each level. Use AI to draft it, then sharpen it yourself.

For a [role], propose a 5-dimension scoring rubric. For each dimension:
- A one-sentence definition
- Concrete behavioral anchors for 1, 3, and 5 (what does each level look
  like in a real example?)
- One example interview question that surfaces signal on this dimension

Avoid proxies for personality. Score on demonstrated skill, judgment, and
relevant experience only.

Lock the rubric before screening starts. Share it with every interviewer. If a panelist wants to add a dimension after seeing candidates, that's a tell — they're rationalizing a gut call.

Screening Resumes Without Outsourcing the Decision

Here's the line: AI can summarize and surface, but it should not reject. Use it to compress 200 resumes into structured notes against your rubric. Make the cut yourself.

Score this resume against the rubric below. For each dimension, give a
score (1–5) with one sentence of evidence quoted from the resume. If a
dimension can't be scored from the resume alone, mark it "needs interview"
— do not guess.

Rubric: [paste rubric]
Resume: [paste resume]

Two warnings. First, never paste a candidate's resume into a consumer chatbot that trains on your inputs — use a workspace tool with data controls, or strip the name and contact info. Second, audit your AI screener. Run it against last year's hires (the ones who worked out and the ones who didn't) and see if it would have caught the good ones. If it wouldn't, your rubric is broken, not the model.

Resume-screening AI has been caught penalizing employment gaps, non-Western names, and women's colleges in published audits. You are responsible for what your tools do on your behalf. Read /courses/ai-ethics-responsible-ai before you deploy anything that touches a hiring decision — the section on disparate impact alone is worth the hour.

Interviews: Structured, Boring, Effective

The data on this is unambiguous: structured interviews — same questions, same order, scored against a rubric — predict performance roughly twice as well as unstructured ones. They also dramatically reduce bias, because every candidate is evaluated on the same evidence.

Use AI to generate a question bank tied to each rubric dimension, then run the same script for every candidate.

For the dimension "[dimension name, e.g., 'handles ambiguity']", generate:
- 3 behavioral questions ("tell me about a time...")
- 2 situational questions ("how would you approach...")
- For each, list 2 follow-up probes and what a strong vs. weak answer
  sounds like

Make the questions specific enough that they can't be answered with
generic platitudes.

Two more rules. Don't let candidates have AI as a silent third party — if it's a remote interview, ask them to share their screen for technical portions, and tell them upfront. And don't penalize candidates who openly use AI on a take-home; instead, ask them in the next round to walk you through how they prompted, what they kept, and what they threw out. That conversation is more diagnostic than the artifact itself.

Debriefs That Don't Devolve

Debriefs go sideways when the loudest person speaks first. Fix it structurally: every interviewer submits scores and written notes in writing, against the rubric, before the meeting starts. Then use AI to synthesize.

Here are the rubric scores and notes from 4 interviewers for one candidate.
Produce:
1. A one-paragraph summary of the consensus view
2. Dimensions where scores diverged by 2+ points (these need discussion)
3. Any concerns that appear in multiple notes
4. Any concerns that appear in only one note (single-source — probe these)
5. Open questions to resolve before a decision

Walk into the room with that document. Spend the meeting on the divergences and the single-source flags — those are where the real signal and the real bias both hide.

What Not to Automate

Don't automate rejection emails to the point of cruelty. Don't use AI to generate "personalized" outreach that the candidate will sniff out in three seconds. Don't run video-analysis tools that score facial expressions or vocal tone — they are pseudoscience and several jurisdictions have already banned them. Don't let the model write your offer call.

The hiring loop is one of the few places where being a human, audibly and visibly, is the entire job. Use AI to clear the runway. Land the plane yourself.