AI resume screening starts with a scoring rubric that finds real fits

A must have is a requirement without which a person cannot do the job in the first six months, even with reasonable support. Most job descriptions list eight or ten of them, and most of those are really preferences. Every false must have shrinks the pool and quietly filters out people who would have been excellent. A useful test is to ask the hiring manager whether they would reject an otherwise brilliant candidate who lacked this one item. If the honest answer is "probably not", it is a nice to have.

Aim for three to five must haves per role and phrase each as a capability rather than a credential. "Has shipped and maintained a production web service in Python" says more than "five years of Python", because it describes what the person did rather than how long they were near the language. Capabilities are also easier to find evidence for, and they are fairer to people with nonlinear careers.

Nice to haves separate good candidates from each other. They should carry real but modest weight, so that a candidate with every must have and no extras still scores well above someone with every extra and a missing fundamental. Industry experience, a specific secondary tool or experience mentoring are classic nice to haves. When a nice to have keeps dominating the ranking during calibration, it is either overweighted or it is secretly a must have, and the hiring manager should decide which.

Deal breakers are not scored at all. They remove a candidate regardless of the total, so they must be rare, objective, lawful and directly tied to the job. Work authorization when the company cannot sponsor, a required professional license, or availability for a schedule the role genuinely demands are reasonable examples. "Not a culture fit" is not a deal breaker, it is an unexamined preference, and it has no place in a rubric.

Weights express how much each criterion matters relative to the others. A simple approach is to distribute 100 points across the scored criteria, giving the must haves most of them. In practice, must haves together usually hold 65 to 80 points and nice to haves the remaining 20 to 35. Resist the urge to give every criterion a similar weight, because a flat rubric behaves like a keyword counter.

Weights should follow the work, not the job title. Two senior engineering roles at the same company can need very different rubrics if one is about building new products and the other is about keeping a payments system stable. Ask the hiring manager to describe the hardest problem the new hire will face in their first quarter, then make sure the criterion closest to that problem carries the most weight.

Every criterion needs a description of what evidence looks like at a high, medium and low level. Without it, scores drift, two reviewers disagree and nobody can explain a ranking. Evidence should be observable in a CV, a profile, a screening answer or a skills test. "Strong communicator" cannot be checked from a profile, while "has written public technical documentation or led customer facing presentations" can.

When evidence is missing, the score should say so plainly instead of guessing. Absence of evidence in a two line profile is not evidence of absence, which is exactly why screening questions and short skills tests exist. A good pipeline uses the rubric twice: once to score profiles and choose whom to contact, and again to score screening answers and test results for those who reply. That second pass is where AI candidate screening fills gaps that a profile alone cannot.

The most damaging rubric errors are criteria that look neutral but track protected characteristics. Graduation year and total years of experience caps track age. Employment gaps often reflect caregiving, illness or disability. Home address or commute distance can track race and income. Prestige of university mostly tracks family background. "Native English speaker" tracks national origin, when the real requirement is usually clear written and spoken English at a professional level.

The fix is to replace each proxy with the capability it was standing in for. Instead of "graduated from a top school", score "can design and explain a data model for a multi tenant product". Instead of "no gaps", score recent, relevant evidence of the skill. In the US, federal laws such as Title VII, the Age Discrimination in Employment Act and the Americans with Disabilities Act apply to screening criteria just as they do to interviews, and some jurisdictions, such as New York City, require bias audits and notices for automated employment decision tools. Review your rubrics and tools with counsel.

This example suits a product company hiring a senior engineer to own backend services. The weights lean toward production ownership and data modeling, because those are the problems a senior hire faces first. Adjust them to your own context before you use them.

This version fits a mid market B2B software company selling to operations leaders. Sales profiles are full of quota claims, so the evidence column asks for specifics such as deal size and sales cycle rather than adjectives. Screening questions do much of the work here, because many strong sellers keep their profiles brief.

Design hiring tends to overweight polished visuals. This example puts more weight on problem framing and on shipped outcomes, which are harder to fake than a beautiful mockup. A portfolio link is the main evidence source, and a short exercise in the skills test can confirm how a candidate reasons through a flow.

No rubric is right on the first draft. Calibration means scoring a small sample, around ten to fifteen profiles, and asking the hiring manager to rank the same profiles blind. Where the two rankings disagree, you have found a rubric problem. Usually it is a nice to have weighted too heavily, a must have described too narrowly or an unspoken preference that the manager holds but never wrote down.

Unspoken preferences are the most valuable thing calibration surfaces. If the manager keeps rejecting people from large companies, ask why. Sometimes there is a real, job related reason, such as needing someone comfortable without a platform team, and it should become an explicit criterion. Sometimes it is a bias, and writing it down is how you notice it. Either way, the rubric gets better and the next search runs faster. An AI recruiter that lets you edit weights and rescore the pool in seconds makes this loop cheap enough to do on every role.

A rubric organizes judgment, it does not replace it. Scores should rank and explain, while people decide. In a well designed pipeline, scoring determines whom to contact and in what order, screening and testing add evidence, and a hiring manager approves or rejects every shortlisted candidate with the reasons visible. That keeps accountability with a person and keeps you clear of rules that restrict decisions made solely by automated means.

Keep each rubric version with the role so you can show what was measured and why if a candidate or regulator ever asks. Over time, this collection becomes a practical hiring playbook for your company. If you want to see how a rubric feeds every later stage of a search, our walkthrough of how AI recruiting works follows a role from intake to shortlist.