How to hire with AI: a stage-by-stage operating procedure

How to hire with AI comes down to one move: split the hiring process by stage, and give each stage to whichever party is actually good at it. AI drafts and distributes the job post, ranks applications against a scorecard a human wrote first, runs a structured first-round interview, and returns evidence for every score. A human owns every rejection and every offer.

That is the whole answer. What follows is the order to do it in, because the order is where almost every rollout goes wrong. If you want the background on what shifted in the market and why, we covered that separately in how AI is changing hiring. This page is the procedure.

The labels move around. AI powered hiring, use AI for hiring, or just the AI hiring process. They all describe the same sequencing problem, and the sequence is what nobody writes down.

Key Takeaways

Don't adopt AI across hiring. Give each stage to whichever party is actually good at it, and be explicit about which party that is.
A human writes the scorecard before any AI ranks against it. Ranking against a standard you never wrote down is how you automate your own bias.
AI may order a list, ask the next question, and quote the evidence behind a score. A human decides the criteria, every rejection, and every offer.
The four-fifths rule in 29 CFR 1607.4(D) is still on the books, but federal enforcement of disparate impact was deprioritized in April 2025, so the live exposure sits with private plaintiffs and states. Log selection rates from the first day you turn anything on either way.
Assume your candidate has AI too. That changes what a technical assessment can measure, and it kills the unsupervised take-home as a signal.
Most of the compliance work is record keeping, not software. If you can't reconstruct why someone was rejected, you don't have a process.

Start with the scorecard, not the software

Before any AI touches your pipeline, a human writes the scorecard: the criteria for the role, what each score level looks like, and which criteria are disqualifying. AI is a ranking engine, and a ranking engine needs a standard. Rank against a standard nobody wrote down and you have automated your existing preferences, now with a clean audit trail attached.

Most teams do this backwards. They buy a tool, let it infer "good" from the resumes of people they already hired, and call the output objective. What that encodes is last year's hiring pattern, including the parts nobody would defend out loud in a room.

A usable scorecard is short and boring. Four to six criteria, a description of what a 1, a 3, and a 5 look like on each, and the evidence that counts as proof. Write it with the hiring manager before you shop for software. If your team can't write it, tooling isn't your problem.

The five-stage split: what AI may decide, and what a human must decide

AI may decide the order of a candidate list, ranked against criteria a human wrote.
AI may decide which question comes next inside an approved question set.
AI may decide what evidence supports a score, quoted from the transcript or the candidate's code.
A human must decide the criteria themselves, and what each score level means.
A human must decide every rejection, by name, with the reason recorded.
A human must decide every offer and the compensation attached to it.

Everything below walks the funnel in order. For each stage: give this to AI, keep this human, and here is the prerequisite that has to exist first.

Stage 1: How to use AI on the job description and sourcing

This is the safest stage to start, because nothing here decides anything about a person. Give AI the drafting work: a first pass at the job description built from your scorecard, plain-language rewrites, and outreach variants for different seniority levels. Keep the requirements list human, because that list is your scorecard in public.

Prerequisite: the approved scorecard. If the AI writes the requirements, you have let a language model set your hiring bar by autocomplete, which is how job posts end up asking for eight years of a five-year-old framework.

Drafting an ad and targeting an ad are different activities under the law. The EU AI Act classifies recruitment and selection systems as high-risk under Annex III, point 4(a), including systems used "to place targeted job advertisements", subject to a narrow Article 6(3) derogation that never applies where the system profiles candidates. If you hire into the EU, ad targeting sits inside the regulated perimeter and ad copywriting does not.

Stage 2: How to screen resumes with AI without creating adverse impact

Give AI the ranking and the evidence extraction: order the pile against your written criteria, and show which line in the resume supports each criterion. Keep the rejection human. The gap between "this candidate ranked 240th" and "this candidate is rejected" is the entire compliance argument, and it costs one person about ten minutes a day to hold.

The rule itself hasn't changed. Under the Uniform Guidelines, a selection rate for any group that is less than four-fifths (80%) of the rate for the group with the highest selection rate "will generally be regarded by the Federal enforcement agencies as evidence of adverse impact" (29 CFR 1607.4(D)). Note the comparator is the group with the highest selection rate, not the highest scores.

What changed is enforcement. The April 2025 executive order Restoring Equality of Opportunity and Meritocracy directs federal agencies to deprioritize disparate-impact enforcement, and the EEOC's AI guidance was withdrawn. The regulation is still on the books, and private Title VII claims and state analogues still turn on it. Your exposure moved to private plaintiffs and state regulators, which is a reason to keep better records, not worse ones.

Prerequisite: a selection-rate log, running from the first day the screener is live. You cannot compute a four-fifths ratio retroactively for data you never stored, and "our vendor says it's fair" is not a record. More on what a screener can and can't see in our breakdown of resume screening software.

Stage 3: What an AI interview is allowed to decide

An AI interview may run a structured first round: the same criteria and the same rubric for every candidate, with follow-ups that adapt to the answer, and a report card that shows the criterion, the score, the transcript excerpt behind it, and the reasoning. What it may not decide is whether the bar itself is correct. That is a hiring-manager judgment, and no volume of interviews will produce it.

Structure is the part that carries the value, not the automation. Schmidt and Hunter's 1998 meta-analysis in Psychological Bulletin put structured interviews above unstructured ones for predicting job performance, and a 2022 reanalysis by Sackett and colleagues revised those coefficients downward while leaving the ordering intact. Structured interviews still came out highest of any predictor examined, at an operational validity of .42.

That is the real argument for handing this stage to a machine. Machines are relentlessly consistent, and humans are not at 4pm on a Thursday. We go deeper on the format in structured interview software, and on the mechanics in how AI interviews are scored.

Prerequisite: calibration before trust. Run a candidate you already have an opinion about through the interview, compare the rubric output to your own read, and tune until they agree. Expert Hire is an AI interview platform for technical hiring, and calibration mode exists for exactly this, because rank order you haven't checked is just a number with good posture.

Since we're asking you to trust scoring, here is where ours is documented. Our methodology page sets out the scheme: one score from 0 to 100, built from the skills and weightage you set against the job, with evidence attached to every number. It also states the limit plainly: scores and rankings go to a recruiter, and nothing is auto-rejected.

Two operational notes. Candidate consent flows are not optional in Illinois, whose AI Video Interview Act requires notice, explanation, and consent before an AI-analyzed video interview. And most teams we talk to lose roughly 50 hours per hire to first-round screens, which is the load this stage is actually removing.

Stage 4: How to assess technical skill when the candidate also has AI

Start from the assumption that your candidate has a model open in the next tab, because most of them do. That single assumption retires the unsupervised take-home as a hiring signal. What survives is live work: a real coding environment, follow-up questions about the code that just got written, and a conversation where somebody has to defend a design decision out loud.

Give AI the environment and the integrity layer: live coding with adaptive follow-ups, tab-switch awareness, screen monitoring, voice-consistency checks, and identity verification before the session starts. Keep the policy human. You decide whether AI assistance is allowed in your assessment, and what "good" looks like if it is.

Check what your plan actually includes before you design around proctoring. On Expert Hire the desktop proctoring app sits on the Growth plan and above, so the tier you buy decides which of those controls you actually get.

Prerequisite: write the AI policy into the invite, before the assessment, in plain words. "You may use an AI assistant and we will ask you to explain every line" is a defensible rule. "No AI" with no way to detect it is theater, and candidates can tell.

Watch the assessment design itself, not just the proctoring. California's automated-decision-system regulations warn that assessments, including "tests, questions, or puzzle games that elicit information about a disability", may constitute an unlawful medical inquiry. Timed logic puzzles and reaction-based games are squarely in that blast radius. Role-relevant coding work is not.

Stage 5: Where the humans go, and the decisions they must keep

Scheduling, reminders, calendar juggling, and ATS record syncing should all go to software and never come back. That is pure coordination cost, it is where days leak out of your process, and nobody was ever promoted for booking a room. We broke the arithmetic down in how to reduce time to hire.

Everything downstream of the shortlist stays human, and this is the part vendors like us are least motivated to say. The final round is a human interview, because by then you are assessing judgment, collaboration, and whether this person wants to build what you are building. AI is genuinely the wrong tool for that, along with closing a candidate, negotiating compensation, and delivering hard news.

Rejections deserve one specific rule: a named person owns each one. Drafting the message with AI is fine and honestly makes rejections kinder, because the alternative is silence. Deciding the rejection is not delegable, and if your process cannot name who made a given call, you do not have a process.

The audit layer: what to measure and what regulators now require

Measure two things: whether the split worked, and whether you can prove it was fair. Operationally that is time-to-shortlist, the selection rate for every group at every stage, and how often a human overturns the AI's rank order. That last one is the most useful number in the whole system, and almost nobody tracks it.

On the regulatory side, four rules matter most as of today:

NYC Local Law 144. An automated employment decision tool must "have been subject to a bias audit within one year of the use of the tool", the results must be publicly available, and candidates need notice 10 business days before use. DCWP enforcement began on July 5, 2023.
California. The Civil Rights Council's automated-decision-system regulations were approved on 27 June 2025 and took effect on 1 October 2025. Employers must keep employment records, including automated-decision data, for a minimum of four years.
EU AI Act. Annex III, point 4(a) makes recruitment and selection systems high-risk, including systems that "analyse and filter job applications" and "evaluate candidates". A narrow Article 6(3) derogation exists for systems doing only a narrow procedural or preparatory task, and it never applies where the system profiles candidates.
Colorado. SB26-189 became law on 14 May 2026. It repeals and reenacts the 2024 Colorado AI Act, pushes the effective date to 1 January 2027, and narrows the regime to four duties: notice, disclosure within 30 days of an adverse outcome, a right to correct inaccurate personal data, and meaningful human review. The duty-of-care and impact-assessment obligations most 2025 articles describe were stripped out, and nothing here is in force today.

The full jurisdiction map lives in our guide to AI hiring laws. None of this is legal advice, and the four-year retention rule alone is reason enough to get counsel to read your process once.

Frequently asked questions

How can AI be used in hiring?

Drafting job descriptions, ranking applications against written criteria, running structured first-round interviews with live coding, scheduling, and producing report cards with evidence. It should not make the rejection call, run the final round, or set compensation.

Are employers actually using AI for hiring, or is it mostly hype?

Both. Adoption is real, particularly in screening and scheduling, but a lot of what is sold as AI evaluation is keyword matching with a confidence score bolted on. The honest test is whether the tool shows you the evidence behind every score. If it can't, it isn't evaluating anything.

Which hiring decisions should AI never make on its own?

Rejections, offers, compensation, and the criteria themselves. AI can order a list and quote the evidence for each score. A named human decides who leaves the process and who gets an offer.

How do you stop AI resume screening from creating an adverse impact problem?

Write the criteria before the tool ranks anything, log selection rates by group from day one, and check the four-fifths ratio at every stage. Keep the rejection decision with a human. A vendor's fairness claim is not a substitute for your own numbers.

How do you run a technical interview when the candidate is also using AI?

Assume they are. Move the signal to live work: a real coding environment, follow-up questions about the code they just wrote, and a design decision they have to defend out loud. State your AI policy in the invite, and add proctoring and identity verification rather than pretending an unsupervised take-home still measures anything.

Do you have to pay to hire with AI, or can you start for free?

You can start free. Expert Hire's trial is free for 30 days and opens with 75 Hire Credits for AI interviews, resume screening, coding tests, and prompt assessments. A card is required, nothing is charged for 30 days, and paid plans run from $99 a month per workspace with unlimited team members.

How to hire with AI: the split is the product

The reason most guides to this topic read as menus is that a menu is easier to write than a sequence. But hiring is a sequence, and the decision that settles whether any of it works gets made before you buy anything.

A human writes the scorecard, and every AI in your process ranks against that document rather than against itself. Get that right and the tooling choice becomes small. Get it wrong and a better model just makes your existing bias faster.

If you want to see what a defensible first round actually produces, our AI interview platform page walks through the report card it returns: a score out of 100 with a verdict line, a skill breakdown scored from the transcript and the code, and the recording and transcript behind every number.

That page is a walkthrough, not a live artifact. Opening a real report card on your own role takes a trial round, which is the only honest way to find out whether it holds your bar.

By TK, Growth at Expert Hire. Last updated September 22, 2026. Reviewed by Anand Suresh, CPO at Expert Hire.