Screen a stack of resumes without reading all of them

A structured summary of every applicant against your job-related criteria, so you spend your attention on the 20 candidates worth a real look instead of slogging through 280 obvious passes. The human ranks and decides; the AI only organizes.

What you'll have when you're done

A repeatable screening pass: you define what the role actually requires, drop the resumes into an AI workspace, and get back a clean table, each candidate's relevant experience, a yes / maybe / no against your criteria, and a one-line reason with a supporting quote. You read the yeses and maybes. You make every call. What used to be a weekend of reading becomes an hour of deciding, with a setup that keeps you on the right side of hiring law.

You posted one role and got 300 applicants

This is the moment AI earns its keep in hiring. A good role pulls hundreds of applicants, most of them clearly not a fit, and the job of finding the 20 worth interviewing buries you. So it gets done badly: skim the first 40, get tired, miss a great candidate in position 200. I have done exactly this, hired someone fine from the top of the stack, and wondered later who was sitting at position 180 that I never gave thirty seconds to. The honest problem is not that I screened harshly. It is that I never screened the back half at all.

An AI reads all 300 in minutes and hands you a structured comparison. But hiring is also the place where pointing AI in the wrong way creates real legal exposure, so this workflow is built around one rule: AI summarizes and organizes, you rank and decide. Stanford researchers studying four million applications found AI hiring tools systematically disadvantaged some groups (roughly a quarter of Black applicants and a sixth of Asian applicants faced algorithmic discrimination in that work), and the law (see adverse impact) holds you, the employer, responsible even when a vendor's tool caused the skew. Build it right and you get the speed without the liability.

What you need first

• Your job-related must-haves, written down first. Concrete, defensible requirements: "ran paid acquisition at $1M+/mo budget," "managed a P&L," not "rockstar" or "culture fit." Vague criteria are where bias hides.
• A Claude Project on a business plan, or your ATS's built-in AI review. Tools like Ashby redact personal details before the model sees them and cite their reasoning, which is the pattern you want.
• The resumes, ideally with names and photos removed (covered in Step 2).
• A clear decision: a human ranks and rejects. The AI never auto-rejects anyone.

Step-by-step

Step 1Write the criteria before you touch the resumes

List the role's objective, job-related requirements first, before any candidate is in front of you. This does two things: it makes your screen consistent, and it keeps you out of the "I'll know it when I see it" trap where bias lives. As one recruiter put it, culture fit is where great hiring goes to die, because vague criteria invite bias. Write the must-haves as things you could defend to a court: tied to the actual work.

Step 2Strip the bias signals

Names, photos, addresses, and graduation years all leak race, gender, and age, and the research is clear that models act on those signals whether you want them to or not. Either remove them before the AI sees the resumes, or use an ATS feature that redacts PII automatically. This single step removes the most direct path to a discriminatory screen.

Step 3Ask for a summary, not a score

In your Project, prompt for organized facts against your criteria, explicitly not a ranking:

For each resume, extract: years in relevant role, accomplishments with numbers,
and a yes / maybe / no against ONLY these criteria: [paste your job-related must-haves].
Give a one-line reason and a supporting quote for each. Do not infer culture fit,
personality, or anything not on the criteria list. Do not rank candidates against
each other; I will do that.

The "do not rank" instruction matters legally: a tool that scores or ranks candidates is more likely to count as a regulated automated employment decision tool. Keep the AI summarizing.

Here is the shape of what comes back, illustrative, three rows from a paid-acquisition role screen with names already stripped:

Candidate	Relevant experience	Criteria	Reason + quote
#047	6 yrs paid social, $1.4M/mo budget	YES	Ran budget at scale: "owned a $1.4M monthly Meta budget, cut CAC 22%"
#112	3 yrs, agency side, ~$200K/mo	MAYBE	Scale unclear: "managed paid for 8 SMB clients," no single large P&L
#178	8 yrs brand marketing, no paid	NO	Criteria is paid acquisition: "led brand campaigns and events"

Notice what the AI did and did not do. It pulled the facts, mapped each to your written criteria, and handed you the quote so you can confirm it in two seconds. It did not decide #178 is a worse person, or that #047 beats #112. It bucketed against your bar and left the ranking to you. You read the fifteen YES and MAYBE rows out of three hundred, and the NO pile is parked, not deleted, in case your bar was wrong.

Step 4You read, you rank, you decide

Sort by the yes/maybe/no buckets and read the yeses and maybes yourself. The AI compressed 300 resumes into a comparison you can hold in your head; the judgment is yours. Never let the AI's "no" auto-reject anyone, a borderline summary is your cue to look, not the AI's cue to discard.

Step 5Keep a record and check the funnel

Log who decided what (an audit trail is your protection). Periodically run a four-fifths check: if one protected group advances at less than 80% the rate of the top group, you have a skew to investigate before it becomes a pattern. And because hiring law varies by city and state, run your setup past employment counsel before you scale it.

The four-fifths check in plain numbers, because the rule is easy to state and easy to get wrong. Say 100 men and 100 women applied, and your screen advanced 40 men to interview, a 40% selection rate. Four-fifths of 40% is 32%. If fewer than 32 women advanced, say 25 (a 25% rate), then the women's rate is 25/40, about 62% of the men's rate, comfortably under the 80% threshold, and you have an adverse-impact flag. The flag is not proof of discrimination. It is your cue to look: is one criterion acting as a proxy, did the name-and-photo redaction actually fire, is the applicant pool itself skewed upstream? The point of running it on your own funnel is that you find it on your own schedule, with time to fix it, rather than reading about it for the first time in a demand letter.

How you'll know it's working

You interview a stronger shortlist in less time, and you stop the "I only read the first 50" problem, because all 300 got the same structured look. The deeper signal is that your screen is now defensible: written criteria, redacted inputs, a human decision, an audit trail, and a funnel you check for skew.

When it breaks

• The AI starts ranking or rejecting. Your prompt drifted. Restate: summarize against criteria, do not rank, do not reject.
• It infers things not on the resume ("seems like a culture fit"). Cut it hard, that is exactly the vague, bias-prone inference to ban.
• A great candidate got a "no." Your criteria are too narrow or the resume used different words. This is why you read the maybes and keep humans in the loop.
• You are tempted to let it auto-reject the "no" pile to save time. Do not. Auto-rejection by an AI tool is the highest-liability move in this entire workflow.
• Every summary reads the same, vaguely positive. Your criteria are too soft, so the model reaches for something nice to say about everyone. Tighten them to concrete, measurable must-haves and the buckets separate.
• Two equally qualified people land in different buckets. Usually the resumes used different words for the same thing. Give the model the synonyms ("paid acquisition, paid media, and performance marketing all count") so vocabulary, not capability, stops deciding.

Make it yours. The bucket pass is the big win on a high-volume role (the 300-applicant marketing opening). On a senior search with twelve hand-sourced candidates, screening is not your bottleneck. Point the same workflow at a different job: feed it each candidate's resume and your criteria and ask it to draft the three interview questions that would best probe where that specific person is unproven. Same setup, same redaction discipline, different output for a different stage.

Where this fits in your harness

This is the front of the hiring pipeline. The candidates who clear it move to interviews, where interview transcripts become a structured scorecard. The whole funnel works better when the role started with a sharp job description that filters for the right operator, fewer wrong applicants means a cleaner stack to screen.