AI in Hiring: How It Works
What each type of hiring AI does with candidate data, where errors and bias can enter, and what recruiters can check

Recruiters now leverage AI in hiring at several points in a search, from the job posting a candidate reads to the ranked shortlist a recruiter opens. The label covers quite different technologies, though, and each one turns candidate data into an output in its own way.
In plain terms, AI in hiring is software that reads candidate and job information, then parses, matches, ranks, writes, or summarizes it to support a recruiter's decision. Knowing how a tool reaches its output shows you where errors might show up and what to check before a shortlist goes out.
Key takeaways
- Ask which functions in a tool use AI, because parsing, ranking, writing, and scheduling each fail in different ways.
- Compare parsed fields and match reasons with original resumes for a sample of candidates before you trust a ranked list.
- Review a few candidates below the cut line, since people in one study followed the AI's picks closely.
- Confirm the rules where the role is based, because New York City, Illinois, and California each regulate AI in hiring decisions.
What AI in hiring covers
Most vendors' AI in hiring falls into seven types:
- Resume parsers
- Matching and ranking engines
- Candidate search
- Writing assistants
- Chatbots
- Interview analysis
- Agents
All of them sit under the AI recruiting umbrella, but they differ in what they do with a candidate.
A practical way to sort them is by whether they organize information or judge the candidate. Parsers, search, writing tools, and schedulers prepare information that a person then acts on. Matching engines and interview scoring make assessments that can decide who moves forward, so they deserve a closer review.
In the Society for Human Resource Management's AI in HR 2026 research, 39% of organizations used AI in human resources (HR) and 27% applied it in talent acquisition. No other HR area ranked higher. The most common uses were writing job postings, screening and matching resumes, and scheduling interviews. That survey covered in-house HR teams, but agency recruiters meet the same functions in their own systems and in client processes.
What each type of hiring AI does with candidate data
Each type takes in different data and produces a different output, so errors enter at different points.
Resume parsing
A resume parser reads a resume file and copies what it finds into structured fields such as titles, employers, dates, and skills. From then on, your applicant tracking system (ATS) searches and matches on the parser's version of the candidate, not the original. A misread field looks just as clean as a correct one, so it's worth knowing how parsed resume fields come out for the formats your candidates send.
Matching and ranking
A matching engine compares parsed fields with the job's requirements and produces a match score, a number or label estimating how closely a profile fits. Ranking sorts candidates by that score, so the order depends on how the tool weights each criterion. That weighting may not match how a hiring manager or client would trade experience against a specific skill.
Tools that show the reasons behind each score make the weighting visible. That's why reasons matter so much in AI resume screening, since an unexplained score gives you nothing to compare with your intake notes.
Candidate search
AI candidate search turns a plain-language request into a structured query across a database of profiles. Some tools use semantic search, which looks for related meaning rather than exact keywords. That surfaces people who describe a skill in different words, but it also means the tool interprets your request, so check how it translated your must-haves.
Generative writing
Generative writing tools run on a large language model, a system trained on large amounts of text to predict likely next words. Recruiters use them to draft postings, outreach, and interview questions. The model produces plausible text rather than checking facts, so it works best when you supply the pay, location, and requirements yourself.
Chatbots and scheduling
Recruiting chatbots answer candidate questions, collect screening answers, and book interviews against interviewers' calendars. A scripted bot chooses from answers someone wrote in advance, so it can only be as wrong as its script. A language-model bot writes its own replies and can field unexpected questions, so that kind of recruiting chatbot needs a clear handoff to a person.
Interview analysis
Interview analysis tools transcribe recorded interviews and summarize them, and some also score candidates. Those functions carry different levels of risk. A transcript or summary is a memory aid you can check against the recording, while the interviewers make the decision. A score is an assessment in its own right, so if it moves candidates forward or out, it needs the same scrutiny as a match score.
AI agents
An AI agent chains several of these steps, such as finding candidates, drafting outreach, and booking replies, without a person approving each one. Each step takes the previous step's output as its input, so the practical decision is which steps to hand off and where to keep approval. Agency recruiters have an added reason to approve outreach, since it goes out under the firm's name.
The table summarizes each type of AI in hiring, with its most likely failure point and what to check.
| Function | What it takes in | What it produces | Where it can go wrong | What the recruiter checks |
|---|---|---|---|---|
| Resume parsing | Resume files | Structured fields in the ATS | Misread dates or titles from unusual layouts | Parsed fields against a few original resumes |
| Matching and ranking | Parsed fields and job criteria | Match scores and a ranked list | Weighting the wrong criteria or a proxy | Match reasons and who fell below the cut |
| Candidate search | A plain-language request | A list of profiles | Treating a must-have as optional | The first page of results against the intake notes |
| Generative writing | Your notes and prompts | Draft postings and messages | Adding details nobody supplied | Each draft against the intake notes |
| Chatbots and scheduling | Candidate questions and calendars | Answers and booked interviews | Wrong answers to unscripted questions | A sample of transcripts and bookings |
| Interview analysis | Interview recordings | Transcripts and summaries plus scores | Scores that measure the wrong thing | What the score measures and the recording |
| AI agents | A goal and access to tools | A chain of completed actions | Early errors passing down the chain | Approval points before candidate contact |
Where bias enters AI hiring tools
Bias can enter AI hiring tools through training data, job criteria, and proxies. Training data is the set of past examples a model learns from, so if past hiring favored certain groups, a model can learn to prefer similar profiles. A matcher also applies every criterion it's given to every candidate, including ones the role doesn't need. A proxy is a data point that tracks a protected characteristic without naming it, such as a zip code.
A University of Washington study had three open-source models rank over 550 real resumes with swapped names. The models favored white-associated names 85% of the time versus Black-associated names 9% of the time. These were research models, not commercial hiring tools, so the results are a reason to test your own tool, not proof about any product.
Ordinary errors can also push a candidate down the list. Consider a candidate whose recent work is a series of contract roles through one staffing firm, and a parser that captures only the latest contract's dates. The matcher scores them low on years of experience, and the ranking drops them below the cut line. A recruiter reading the match reasons would notice the short experience figure, open the resume, and see the full history.
How to review AI recommendations before you use them
A good review starts with your own view, formed before the tool's ranking shapes it. In a later UW study, 528 online U.S. participants picked candidates with simulated AI recommendations. Without suggestions, their choices showed little bias, and with severely biased recommendations, their choices followed the AI around 90% of the time.
That was a lab study, not recruiters at work, but it's a fair reason to build independent judgment into your process. These are recommended practices rather than legal requirements:
- Look at who was dropped: Open several candidates just below the cut line and ask whether any would have passed your own screen.
- Check the reasons: Confirm each recommended candidate's match reasons reflect the job's requirements rather than a proxy like a school name.
- Compare a sample by hand: Screen a small batch yourself before viewing the ranking, then compare the lists to see which criteria the tool weighs differently.
Agency recruiters can use these notes when a client asks how candidates were screened, and in-house recruiters can use them when a hiring manager asks why someone wasn't advanced.
See why each candidate ranked where they did
Book a demoRules on AI in the hiring process
Several jurisdictions regulate AI in the hiring process, and these three are examples, not an exhaustive list. New York City's Local Law 144 bars employers and employment agencies from using an automated employment decision tool unless it has had a bias audit within one year. The audit information must be public, and notices must go to candidates or employees, so agency recruiters filling city roles may be covered.
Illinois amended its Human Rights Act effective January 1, 2026, to bar employers from using AI that has a discriminatory effect. The amendment also bars employers from using zip codes as a proxy for protected classes and requires them to notify employees when they use AI.
In California, Civil Rights Council rules took effect October 1, 2025. They say an automated-decision system may violate state law if it harms applicants or employees based on protected characteristics. Employers must also keep automated-decision data with employment records. Requirements depend on where the role is based, so confirm current rules with counsel.
How to check whether a hiring tool does what the vendor says
The best evidence about a hiring AI tool comes from the vendor's answers to specific questions and from a test on candidates you already know.
- Which steps use AI: Ask which functions parse, score, rank, write, or decide, so you know which outputs need the closest review.
- What a score is based on: Ask which criteria feed the match score and whether you can see reasons for each candidate.
- How it was tested: Ask how the vendor tested for bias, including any bias audit if you hire for New York City roles.
- What you can export: Ask whether you can export scores, reasons, and decisions for your records.
For a practical test, run a closed search with candidates whose outcomes you know and see whether the people you advanced rank near the top. You can also submit one resume as a plain document and as a designed template, then compare the parsed fields. A good result doesn't prove the tool is free of bias, but a poor one shows early where its judgment differs from yours.
If you're still comparing products, sort them by AI software type to see which questions matter most for each.
Check the outputs before they shape a shortlist
A practical first step is to pick one AI function you already rely on, such as the ranking in your ATS. On your next batch of candidates, screen a handful yourself first, then compare your picks with the tool's. Where the two disagree, the match reasons often point to a specific field, criterion, or proxy you can correct or raise with the vendor. The same compare-first approach works as you switch on other AI features in your ATS, one at a time.
Start with a search you can check
Book a demoFrequently asked questions
Do candidates have to consent to AI in the hiring process?
The rules covered here focus on notice: New York City requires notices to candidates or employees, and Illinois requires employers to notify employees when they use AI. Whether consent is also required depends on where the role is based, so check with counsel.
Will AI replace recruiters?
The research cited here doesn't measure effects on recruiting jobs, so any forecast would be a guess. Each tool described automates a narrow task, and its outputs still need someone who knows the role to decide who moves forward.
About the author

Founder & CEO of Rotto, building tools that help tech recruiters source better candidates, faster.





