AI for HR: hiring, onboarding, performance reviews

Most HR people I know lose about four hours a week to mechanical work. Not the work that actually shapes a team. The work around it. Reformatting resumes into a screening doc. Writing the eighth rejection email this week. Stitching a manager's three sticky notes into a real performance review. Drafting an offer letter from a template you've used forty times.

That's the four hours. AI handles it cleanly, if you put it in the right places and keep it out of the wrong ones.

Sourcing without becoming a Boolean wizard

You don't need a sourcing course. You need a job description and twenty minutes.

Paste the JD into Claude or Perplexity and ask for the shape of the candidate, then the search strings to find them. Perplexity is better when you want to validate the profile against actual people in the wild. Claude is better when you want a clean list you can paste straight into LinkedIn Recruiter.

Below is a job description for a Senior Product Designer at a 60-person 
B2B SaaS company.

Give me:
1. Three candidate archetypes who'd fit this role (current title, 
   company size, years of experience, the kind of product they ship).
2. For each archetype, three LinkedIn Boolean search strings I can 
   paste directly into Recruiter or Sales Nav.
3. Five companies whose designers are likely to be open to this move 
   (similar stage, recent layoffs, or known design-team friction).
4. Two non-obvious titles I should also search (people who do this 
   work but aren't called "Product Designer").

[paste JD]

You'll get a list that beats whatever you would have brainstormed at 9am with one coffee.

Screening, and the bias trap

Two hundred applicants land on Monday. You have three hours before standup.

The temptation is to ask Claude to "rank the top ten." Don't. Resume ranking by an LLM amplifies whatever bias is already in the training data and your JD. Stanford and MIT have published on this. It's not theoretical.

The honest use is different. Ask the model to summarize each resume in two lines: one on relevant experience, one on a gap or question worth asking. You read 200 two-line summaries in twenty minutes and decide who advances. The model compresses, the human filters. That's the line.

Summarize this resume in exactly two lines:
Line 1: Most relevant experience for a [role title] at a [company stage].
Line 2: One gap, ambiguity, or question I should ask in a screen.

No ranking. No "strong candidate" language. Neutral tone.
[paste resume]

Run it across the batch. You're faster, and you haven't outsourced the judgment call.

Onboarding that doesn't feel like a Notion graveyard

Most 30/60/90s read like they were written by someone who'd never met the role. Because they were.

Feed the model the JD, the team's last quarterly goals, and a paragraph from the hiring manager about what "good in 90 days" actually looks like. You get a plan that ties to real work instead of generic milestones.

Build a 30/60/90 day plan for a new [role] joining [team].

Inputs:
- The JD: [paste]
- The team's current quarterly goals: [paste]
- Hiring manager's note on what success looks like: [paste]

Structure:
- Day 30: learning goals, people to meet, one shippable small win.
- Day 60: first owned project, who they're collaborating with, 
  how we'll know it's on track.
- Day 90: a clear deliverable tied to one of the team's quarterly 
  goals, plus the questions their manager should be asking by then.

Be specific. No "ramp up on the codebase" filler.

Send the draft to the manager, let them edit for ten minutes, ship.

Performance reviews from a pile of notes

Managers don't write bad reviews because they're lazy. They write bad reviews because they have eleven scattered observations and no structure to put them in.

That's the part to automate.

Below are a manager's raw notes on a direct report from the last 
six months. They are messy, out of order, and include both 
positive and constructive observations.

Organize them into a draft review with this structure:
1. Strengths (3, each with one specific example from the notes)
2. Areas for development (2, each with one example and one 
   suggested next step)
3. Impact on the team (a paragraph, grounded in what's in the notes)
4. Open questions the manager should reflect on before finalizing

Do not invent examples. If something isn't supported by the notes, 
flag it as "needs more input."

[paste notes]

The manager still owns the judgment, the rating, the conversation. The model just stops them from staring at a blank Google Doc on Sunday night.

What NOT to outsource

The hiring decision. The actual conversation with a candidate or an employee. The PIP call. The offer negotiation. The "we're letting you go" meeting. Anything that decides someone's career, their paycheck, or how they feel about your company a year from now.

If it's the moment that matters to the person on the other end, you show up. The model drafts the email afterward.

Up next

The next pillar 2 guide moves from people-team work to revenue retention: "AI for customer success: spotting churn before it happens." Same playbook, different stakes. The signals are noisier, and the cost of missing them shows up in next quarter's net revenue.