How to screen candidates with AI without losing judgment or breaking the rules

TLDR: AI can remove screening’s mechanical side (sorting two hundred applications, summarizing each against the requirements) with one line never crossed: rejecting a person is decided and explainable by a human. There is no single federal AI-hiring law, but Title VII, New York City’s Local Law 144 and California’s automated-decision-system rules already reach this. This guide builds the flow that gains the speed without buying the problem: explicit criteria, AI that sorts with visible reasons, documented human decisions and bias vigilance.

This guide is for anyone receiving more applications than they can read attentively and tempted to let a machine decide. The temptation is understandable and the short answer fits in one sentence.

The machine can read for you, not decide for you.

Well built, that division is faster and also fairer than the classic eleven-at-night diagonal skim.

1. Understand the framework before automating

The starting point is not technical but regulatory, and it deserves clarity before configuring anything. The US has no single federal law written for AI hiring: what applies is a patchwork of an old civil-rights statute and a handful of state and city rules, and all of it reaches automated screening whether or not you notice.

Title VII of the Civil Rights Act still forbids employers to “fail or refuse to hire” or otherwise discriminate on race, color, religion, sex or national origin, and that applies to a selection tool exactly as it applies to a person. The twist worth knowing: the EEOC published technical guidance in 2023 on assessing AI-driven adverse impact under Title VII, and that guidance page no longer exists on eeoc.gov. The statute did not move. The agency just stopped explaining how it applies to a scoring algorithm, which leaves employers to work out disparate impact on their own instead of by the EEOC’s old four-fifths rule of thumb.

Two places add real, checkable obligations. If you hire for a role tied to New York City, Local Law 144 requires an independent bias audit within the year before you use the tool, a public summary of that audit, and candidate notice at least 10 business days before the tool is used. If you hire in California, regulations in force since October 1, 2025 bring any “automated-decision system” used in recruiting or screening under the state’s anti-discrimination law, whether you built the tool or a vendor did.

The practical translation in three obligations, wherever you hire: real human oversight (not a human mass-approving what the machine decided), a documented reason for every rejection, and the capacity to explain a decision if a candidate or a regulator questions it. Building the flow to that standard everywhere costs little more than building it for one city, and it survives the next state that passes a law you have not read yet.

None of this forbids using AI in screening: it forbids hiding behind it.

2. Define assessable criteria and write them down

Fair screening (and screening worth automating) starts with explicit criteria: what gets assessed, with what weight and why it is linked to the role. The document derives from the role definition in the job postings guide: the three essential requirements, the desirables with their weight and the fit signals.

That document has a double function: it is the instruction the tool uses to evaluate and it is your defense if a rejection gets questioned, because it connects every decision to role criteria rather than to the on-duty reviewer’s intuition.

The quality test for a criterion: two different reviewers applying it to the same application should reach the same conclusion. “Relevant experience” fails it. “Three years managing teams of five or more” passes.

3. Use AI to sort and summarize, not to reject

With criteria written, the division of labor is clean, and there are two ways to build it depending on what you already run.

In a complete system like Workable, the result is a reading order where the team starts with the best fits, summary of why included, with a trial available and with human oversight as part of the design.

Without a full platform, the same principle works artisanally with Claude, which has a permanent free tier. A hiring process means reading documents full of personal data and writing judgments you must be able to defend, and that double requirement outranks price when picking the assistant. Slower, same division: the machine reads and summarizes, you decide.

4. Document the process and watch for bias

The defensible flow leaves a trail at four points.

That log is the difference between explaining a decision and reconstructing it afterwards.

Bias vigilance is the other leg: systems learn from historical data and historical data carries bias. The practical control within your reach: periodically inspect the screening’s patterns (does the machine’s ranking systematically penalize a profile unrelated to the role?) and contrast the low-score sample from the previous step. Later interviews documented with Fireflies complete the comparable end-to-end process, as the documented interviews guide develops.

The complete system’s result: faster than the diagonal skim, fairer because everyone passes through the same criteria, and defensible because everything can be explained. The whole sector lives in AI for human resources.

Common mistakes

Letting the machine reject alone. It is the legal and ethical line never crossed, and bad business besides: the machine’s false negatives are good candidates going to your competitor.

Screening on implicit criteria. Without written criteria, the AI evaluates against guesses and the human against intuition: fast, unfair and indefensible at once.

Mass-approving the machine’s ranking. Pantomime oversight (click, click, click) is the modern version of no oversight, and it leaves you with the same liability as no oversight.

Never looking at the low-scored. The tail control sample is the cheapest bias detector there is, and skipping it is trusting the ranking blindly.

Assuming a headline law is already in force. Colorado passed the country’s first broad AI-in-employment law in 2024, then delayed it and replaced it before it ever took effect: the state’s current automated-decision-technology law does not apply until January 1, 2027. Build to the three obligations above, not to whichever state law you last read a headline about.

Frequently asked questions

Is AI screening legal in the US?

Yes, as support with real human oversight, a documented reason for every rejection and the capacity to explain a decision. There is no single federal AI-hiring statute, but Title VII already reaches discriminatory outcomes regardless of who built the tool, and NYC and California add specific, checkable requirements on top. This guide’s flow is built to clear all three at once.

Do I have to tell candidates?

Only New York City makes it a hard requirement, where Local Law 144 demands notice at least 10 business days before the tool is used. Outside NYC it is not yet a blanket legal requirement, but it is the clean practice anyway: mention it in the process information, in plain language, alongside the data notice you already provide.

How accurate is automatic screening?

It sorts well against clear criteria and fails where criteria are vague or the CV is atypical (non-linear careers, sector changes). That is why the low-score sample is mandatory: the false negatives live there.

Is it worth it without volume?

With twenty applications per process, manual screening against written criteria is plenty and AI adds little beyond summaries. The full system starts paying at the volume you can no longer read attentively.