Every recruiter you know is already using AI. Not because the company bought a platform, but because someone got tired of writing the fourteenth job description that month and pasted the brief into a chat window. The tools arrived from the bottom up, and most hiring policies have not caught up.
That is worth sorting out, because AI genuinely helps at some stages of hiring and quietly causes damage at others. The difference is not the technology. It comes down to one thing: is the task admin, or is it judgement?
The line that matters
Split every recruiting task into two piles. In the first pile the correct answer is knowable and checkable: find a slot both calendars share, send the reminder, summarise what was said, format the offer letter. In the second pile the answer is a judgement about a person: is this candidate better than that one, should this application be dropped.
AI is very good at the first pile and unreliable in the second. Most disappointing rollouts happen because someone bought a tool that promised the second and was only ever good at the first.
A model trained on your last five years of hiring will reproduce your last five years of hiring. If you were not happy with those results, automation makes them arrive faster.
Stage by stage
| Stage | What AI does well | Where it goes wrong |
|---|---|---|
| Sourcing | Widening a search, surfacing candidates outside your usual pipelines, drafting outreach | Volume without targeting, which fills the funnel and burns recruiter time downstream |
| Screening | Extracting structured fields from unstructured resumes | Ranking and auto-rejecting, where it encodes proxies for background rather than ability |
| Scheduling | Finding slots, rescheduling, sending reminders, chasing feedback | Very little. This is the safest and highest-return use in hiring |
| Assessment | Scoring work samples against a defined rubric | Inferring personality or fit from video, tone or facial expression |
| Interviews | Transcription, notes, filling the scorecard so interviewers can pay attention | Replacing the interviewer, or scoring the candidate without one |
| Offer and close | Drafting letters, answering routine candidate questions | Negotiating, or anything a candidate would resent hearing from a bot |
Sourcing: more candidates is not the goal
AI sourcing tools are good at finding people your keyword search missed. Someone whose title says Analytics Lead when you searched for Data Manager, someone who spells the skill differently, someone two degrees outside your network. That is a real gain, because the best candidate is often the one your search string excluded.
The trap is treating volume as success. Doubling applications while keeping the same screening capacity means your recruiters spend the saved time on a longer list. Measure the quality of the shortlist and the number of hires, not the size of the funnel.
Screening: the stage that needs the most care
Resume screening is the most automated part of hiring and the least examined. The pitch is compelling, because a recruiter reading three hundred applications is slow and inconsistent and knows it. The problem is what the model learns from.
A ranking model trained on who you hired before learns the patterns in that data, including the ones you would not defend. Colleges you happen to recruit from. Companies whose names appear often. Career gaps, which correlate with caregiving. Address, which correlates with a great deal. None of these needs to be an explicit input for the model to find a proxy for it.
Extraction is different and much safer. Pulling years of experience, certifications, notice period and location out of a resume into structured fields removes the tedious part without making the decision. Use AI to read, not to reject.
Scheduling: the easiest win in hiring
Interview coordination is pure administration and it consumes an enormous share of a recruiter's week. Panels of four across two time zones, a reschedule, a no-show, a chase for feedback that never arrives. Automating this is low risk, immediately measurable and improves the candidate experience at the same time.
If you adopt one thing, adopt this. It also builds internal confidence in the tooling before you go anywhere near screening.
Assessment and interviews
Scoring a coding exercise or a written work sample against a rubric is a reasonable use of a model, because the rubric is explicit and a human can check the output. Inferring conscientiousness from a video interview is not. The research supporting facial and vocal analysis for hiring is weak, and several vendors have quietly withdrawn those features after regulatory attention.
Interview transcription and automatic scorecard filling deserve more attention than they get. Interviewers write notes badly because they are trying to listen at the same time. A tool that captures what was said, so the interviewer can actually pay attention, improves the quality of the decision without making any part of it.
What the law now expects
Hiring is a regulated use of automated decision-making in a growing number of places, and the direction is consistent even where the details differ.
- New York City requires an annual independent bias audit of automated employment decision tools, published results, and notice to candidates
- The EU AI Act classifies recruitment and candidate selection as high risk, with obligations covering risk management, data governance, logging and human oversight
- India's Digital Personal Data Protection Act requires a lawful basis, notice and purpose limitation for processing candidate data, which includes feeding it to a third-party model
- Illinois requires consent and disclosure before AI analysis of video interviews
Two practical consequences apply almost everywhere. Keep a record of what the tool did and who decided, because you will be asked to show it. And be careful what candidate data leaves your systems, since pasting a resume into a public model is a disclosure.
How to introduce AI without breaking hiring
- Fix the process first. Write the scorecard and agree the must-haves before automating anything, because automating an undefined process produces faster confusion
- Start with admin. Scheduling, drafting, reminders and notes save hours immediately and carry almost no risk
- Run any ranking tool in parallel for a full cycle. Let it score while humans decide, then compare. Look hard at who it dropped. Everyone checks the top of the list and nobody checks the bottom
- Keep a named person on every rejection. A tool may rank, flag or summarise. A person makes the call and that is recorded
- Audit outcomes quarterly by gender, age band, college tier and location, against the same numbers before the tool arrived
- Write down what may be pasted into which system. Two pages beats a committee that meets next quarter
- Tell candidates. Disclosure is becoming a legal requirement and is already a reasonable expectation
Questions worth asking a vendor
- What data was the model trained on, and was any of it ours?
- Can you show me an adverse impact analysis, and when was it last run?
- What happens to candidate data, where is it stored, and who else can see it?
- Can a recruiter see why a candidate was ranked where they were?
- Can we turn the ranking off and keep the extraction?
- Who is accountable if a rejected candidate challenges the decision?
A vendor who cannot answer the second and the fourth is selling you a liability with a dashboard on it.
Frequently asked questions
What is AI used for in recruitment?
Most commonly for sourcing candidates, extracting information from resumes, scheduling interviews, drafting job descriptions and outreach, transcribing interviews and answering routine candidate questions. Ranking and shortlisting are also common and carry the most risk.
Can AI replace recruiters?
No, and the framing misses what recruiters actually do. AI removes coordination work and speeds up drafting. It cannot assess judgement, sell a role to a hesitant candidate, manage a hiring manager who keeps changing the brief, or take responsibility for a decision.
Is AI resume screening biased?
It can be, and the bias comes from the training data rather than the algorithm. A model trained on past hiring decisions learns whatever patterns those decisions contained, including proxies for college, gender, age and location that were never explicit inputs. Regular adverse impact testing is the only way to know.
Do we have to tell candidates we use AI in hiring?
Increasingly yes. New York City requires notice for automated employment decision tools, Illinois requires consent for AI analysis of video interviews, and the EU AI Act imposes transparency obligations on high-risk uses including recruitment. Even where it is not mandated, disclosure is becoming an expectation.
What is the safest place to start with AI in recruitment?
Interview scheduling and coordination. The task is pure administration, the output is easy to verify, it saves recruiters several hours a week, and it improves candidate experience without touching any hiring decision.
The useful version of this technology is unglamorous. It books the interviews, writes the first draft, captures the notes and leaves the recruiter with more time to talk to people. The version that promises to tell you who to hire is the one to test hardest and adopt last.
