Recruiting has more AI vendors than any other HR function — and it’s not close. Three reasons: volume (a single job posting can generate hundreds of applicants), clear metrics (time-to-fill, cost-per-hire, quality-of-hire are measurable), and massive cost savings (reducing time-to-fill by even a few days saves thousands per requisition). AI thrives where there’s lots of data, clear success metrics, and a strong financial incentive. Recruiting checks all three boxes.

Traditional sourcing meant writing Boolean strings: "software engineer" AND "Python" AND "San Francisco". The problem? It only finds people who use exactly those words on their profile. AI sourcing tools like HireEZ, Entelo, and SeekOut go further: they understand that “ML engineer” and “machine learning developer” are the same thing. They find passive candidates who aren’t actively looking. They predict likelihood to respond based on signals like recent activity, job tenure, and company changes.

Semantic search: Understanding intent behind words — genuinely useful, works well.
Response prediction: Reasonably accurate based on behavioral signals.
Personalized outreach: AI-drafted InMails get higher response rates when done well.

AI resume screening happens in two stages. First, parsing: NLP (natural language processing) extracts structured data from an unstructured document — name, contact info, skills, job titles, dates, education. Second, scoring: a matching algorithm compares the parsed data against the job requirements and produces a fit score. The difference between keyword matching and semantic understanding is critical:

Resume screening is where the most candidates are eliminated in the shortest time with the least human oversight. A screening AI might reject 80% of applicants before a human ever sees them. That means:

Scale of impact: Thousands of people affected per posting
Invisibility: Rejected candidates rarely know AI made the call
Historical bias: Models trained on past hires inherit all the biases in your hiring history
Proxy variables: A candidate’s name, university, address, or even formatting choices can become proxies for protected characteristics

Scheduling: AI coordinates across multiple interviewers’ calendars, handles rescheduling, sends reminders. This is high-value, low-risk automation — exactly the kind of AI you should embrace.

Structured interview scoring: AI can help interviewers stick to structured rubrics, flag when scoring drifts between interviewers, and identify calibration gaps. This actually reduces bias by enforcing consistency.

Interviewer analytics: Which interviewers are the best predictors of eventual performance? AI can surface patterns across hundreds of interviews that no human could track manually.

HireVue pioneered video interview analysis — using AI to score candidates on facial expressions, word choice, and vocal tone. The idea: predict job performance from how someone acts on camera. The backlash was severe. Researchers flagged disability bias (people with facial differences or speech patterns scored differently), racial bias (facial analysis algorithms perform worse on darker skin tones), and fundamental validity questions (does a micro-expression actually predict job performance?). In 2021, HireVue dropped facial analysis entirely. The lesson: just because AI can analyze something doesn’t mean it should.

In 2018, Reuters revealed that Amazon had built an internal AI recruiting tool that systematically penalized women. Trained on 10 years of resumes (in a male-dominated tech workforce), the model learned that male candidates were more likely to be hired. It downgraded resumes containing the word “women’s” (as in “women’s chess club captain”) and penalized graduates of all-women’s colleges. Amazon scrapped the tool.

You can remove “race” from the data and still have a racist model. Here’s how:

Zip code → race: Residential segregation means zip code is a strong proxy for race in many cities.
University → socioeconomic status: Favoring certain schools penalizes first-generation college students.
Name → ethnicity: Research shows that “distinctively ethnic” names trigger different callback rates.
Gaps in employment → gender: Career breaks disproportionately affect women (caregiving).

Status updates: “Where is my application?” is the most common candidate complaint. AI chatbots can provide real-time updates 24/7 without recruiter involvement. This is a clear win.

FAQ handling: Benefits, relocation, visa sponsorship, interview logistics — AI handles 80% of candidate questions without human touch.

Scheduling: Self-service interview scheduling eliminates the endless email chains. Candidates love it.

Personalized career sites: AI that surfaces relevant job openings based on a candidate’s profile, similar to how Netflix recommends content.

Impersonal rejection: Being rejected by a chatbot with a generic message after investing hours in an application is a terrible experience. Candidates who feel “processed by a machine” tell others — and they have platforms to do it (Glassdoor, Blind, LinkedIn).

The black hole: AI that screens candidates out before a human sees their application creates the feeling that “nobody even looked at my resume.” Because nobody did.

Accessibility: AI chatbots and video platforms may not accommodate candidates with disabilities, non-native speakers, or those with limited internet access.

Bias audit results: Has the vendor conducted (or had a third party conduct) a bias audit on their AI? Can they share the results? If not, assume the worst.

Model card access: A document describing what the model was trained on, what it optimizes for, known limitations, and performance across demographic groups.

Opt-out mechanisms: Can candidates opt out of AI-based screening and be reviewed by a human? Some jurisdictions now require this.

Explainability: Can the system explain why a candidate was ranked high or low? Not just the score — the reasoning.

Human review requirement: Define which decisions require human review before action. At minimum: all rejections of candidates who meet minimum qualifications, any decision affecting protected classes disproportionately, and all final hiring decisions.

Regular bias audits: Conduct adverse impact analyses quarterly (at minimum annually). Compare selection rates across race, gender, age, disability status, and veteran status. Document everything.

Candidate notification: Inform candidates when AI is used in the screening process. Some jurisdictions require this by law (Illinois, NYC) — but doing it everywhere is best practice.

Appeal process: Give candidates a way to request human review of an AI-driven rejection. This isn’t just ethical — it’s increasingly a legal requirement.