Fifty applications, five interview slots, one afternoon. This is where hiring quietly goes wrong: not in the interview, but in the rushed first cut. AI resume screening promises to fix it, and a bare chatbot half-does. Paste a CV, ask "is this any good?", and you get a confident answer with no criteria behind it. Ask about ten candidates the same way and you get ten answers judged by ten subtly different standards.
There is a second problem that most articles on this topic skip entirely, and it is the one that can cost you a claim rather than a candidate: screening software is regulated in a growing number of places, and the obligations land on the employer, not the vendor. Before the how-to, that part is worth getting straight.
Where AI screening is regulated, as of September 2026
This is not legal advice and the rules move; check your own jurisdiction and take proper advice before you deploy anything. But here is the shape of it, because most guides on this keyword are still quoting dates that have changed.
| Where | What applies | Since |
|---|---|---|
| New York City | Local Law 144. An automated employment decision tool needs an independent bias audit within a year of use, a public summary of the results, and notice to candidates at least 10 business days before use. | Enforced from 5 July 2023 |
| Illinois | HB 3773 amends the Human Rights Act: notice required when AI is used in recruitment, hiring, promotion and related decisions, and using zip code as a proxy for a protected class is prohibited. | 1 January 2026 |
| California | FEHA regulations on automated decision systems. Evidence of anti-bias testing, or the absence of it, is relevant to a discrimination claim, and records relating to automated decision systems must be kept for four years. | 1 October 2025 |
| Colorado | SB 26-189 replaced the earlier Colorado AI Act. Advance notice, a plain-language explanation within 30 days of an adverse outcome, meaningful human review, and three-year records. | Takes effect 1 January 2027 |
| European Union | The AI Act treats recruitment and selection as high-risk. The Digital Omnibus, in force since 27 July 2026, moved those obligations from 2 August 2026 to 2 December 2027, so the deadline many articles still quote has passed without being the deadline. | High-risk duties from 2 December 2027 |
| United States, federal | Title VII and the ADA apply as they always did. Disparate impact is the employer's liability whether the screening was done by a person, a vendor, or a model. | In force |
A criteria-first screening rubric, advance / hold / no with two or three concrete reasons per candidate, red flags framed as screen-call questions, plus Boolean search strings and sourcing hit lists for when the applications are not arriving in the first place.
View the Alex skill →Why generic AI screening is risky
- The standard drifts. Nothing anchors a bare chatbot, so the tenth CV is not judged like the first. That is not a shortlist, it is ten separate opinions.
- It reacts to the wrong signals. An unguided model can pick up on a name, a university, or an employment gap that has nothing to do with whether someone can do the job. In Illinois, using zip code as a proxy is now explicitly prohibited, and the same logic applies to every other stand-in for a protected class.
- It is over-confident. It hands down a clean reject where a careful human would have flagged something worth a five-minute call.
- It leaves no trail. If a rejected candidate asks why, or a regulator does, "the model said no" is not an answer. A rubric and a reason per candidate is.
What changes with a criteria-first skill
The order is the whole point. Alex refuses to judge a CV until it knows what it is judging against: the role brief, the genuine must-haves, and the nice-to-haves kept separate from them. Only then does it read applications, and every candidate after that is measured by the same ruler.
Verdicts come back as advance, hold or no, with two or three specific reasons rather than an essay, formatted to be scanned across a stack of 200. Red flags are framed as questions to explore rather than automatic disqualifiers: "three short tenures in a row, worth asking about in a screen call" is useful, "unstable employment history" is not. And evaluation runs against job-relevant evidence only.
What it actually produces
| Output | What you get |
|---|---|
| Screening criteria | Must-have eliminators separated from nice-to-have differentiators, and a scoring rubric written down before a single application is read, so the standard is defensible after the fact |
| Shortlist | Advance / hold / no per candidate with the top two or three reasons, built to be scannable at volume |
| Flag analysis | Career progression, impact metrics, promotion patterns, title inflation, vague accomplishments and unexplained gaps, each framed as something to explore on a call |
| Boolean strings | Ready-to-paste LinkedIn, GitHub, Dribbble and Google X-Ray searches, with each component explained and variants to widen or narrow |
| Sourcing | Competitor talent mapping, channel recommendations by role type, and hit list templates for when the inbound pipeline is the problem |
| Phone screens | Role-specific question guides, screen note templates, and conversion analysis by sourcing channel |
The prompt to start with
Test the behaviour before buying anything. The trick is that the first message contains no CVs at all:
You are screening candidates for the role below. Do NOT look at any CV yet. Step 1. From this job brief, propose a scoring rubric. Separate: - MUST-HAVES: absence of this is a genuine eliminator - NICE-TO-HAVES: differentiators between otherwise equal candidates For each one, say what evidence in a CV would satisfy it. Challenge any must-have that is really a preference, and say why. Step 2. Wait for me to approve or edit the rubric. Step 3. Only then, for each CV I paste, return: - advance / hold / no - the two or three specific reasons, each tied to a rubric line - anything to ask about on a screen call, phrased as a question Rules: - Judge on job-relevant evidence only. Ignore name, age, address, photo, nationality, school prestige and career gaps unless the brief makes them genuinely job-relevant. - Never turn a gap or a short tenure into a rejection on its own. Flag it as a question. - If the CV does not contain the evidence you need, say "not evidenced" rather than inferring it. JOB BRIEF: [PASTE THE ACTUAL BRIEF, NOT THE JOB TITLE]
Approving the rubric before any CV is read is what makes the screen consistent, and it is also the artefact you want if anyone later asks how the decision was made.
How to keep a person in the decision
- Keep the rubric, dated. It was written before you saw the candidates, which is the single most useful fact about it.
- Read every application yourself at the hold and no boundary. The summary is a prompt for your attention, not a substitute for it.
- Do not rank by the model's score. That is exactly the pattern the NYC rules describe, and it is the difference between a tool that assists and one that decides.
- Write down why you overrode it, in both directions. The overrides are the evidence that a human was actually in the loop.
- Check what your ATS is already doing. Plenty of teams worry about the chat window while their applicant tracking system has been auto-ranking for years, which is far more likely to meet the definition.
- Take advice on notice and audit duties before this touches a real vacancy in NYC, Illinois, California, Colorado or the EU. The obligations sit with you as the employer.
Where it fails
It cannot verify anything. Every claim on a CV is taken at face value, so dates, titles and degrees are yours to check. It has no idea which of two similar employers is the stronger signal in your market. It cannot tell you whether your must-haves are the right must-haves, only whether a candidate meets the ones you wrote. And it will screen a badly written brief just as confidently as a good one, which is why the first two minutes matter more than the next fifty CVs.
Who this is for
In-house recruiters working through volume, founders making early hires without a talent team, and hiring managers who want a faster and more consistent first screen. It runs in Claude, ChatGPT, Gemini or Copilot. The legal, HR and people ops collection holds 52 skills across the rest of the employee lifecycle, and the recruiter skills cover sourcing, interviewing and offers.
Related Skill Guides
Alex - Resume Screener AI Skill
One file. The rubric comes before the CVs, every verdict carries its reasons, red flags arrive as questions rather than rejections, and the decision stays yours. No subscription, yours permanently.
KissMySkills is a marketplace of over 1,900 AI skills, 60+ prompt packs, 90+ agents & free tools for Claude, ChatGPT & any AI chat.
