Candidate Matching

How AI candidate matching works

Most matching systems follow the same steps.

1. Extract structured data. Resume parsing pulls out job titles, employers, dates, skills, certifications, education and location. The job description is parsed the same way.
2. Normalize. "CSR," "customer care rep" and "call center agent" are mapped to the same role; "Excel" and "spreadsheets" to the same skill. Good systems use a skills taxonomy for this.
3. Compare. The system scores how closely the candidate fits on each dimension: skills, years of relevant experience, required credentials, location or commute, availability and pay expectations.
4. Rank or recommend. Candidates are ordered by overall match, sometimes with an explanation ("meets 6 of 7 required skills; 4 years in similar role").

The biggest technical difference between systems is how they compare. Keyword matching looks for the exact words in the job description; it misses good candidates who describe the same experience differently. Semantic matching uses language models to compare meaning, so a candidate who "resolved billing disputes for a telecom provider" can match a job asking for "customer issue resolution." Some systems also learn from past hiring decisions, which is where most of the bias risk comes from.

Where it is used

• Inbound applicants: ranking hundreds of applicants for a requisition so recruiters review the strongest first.
• Rediscovery: searching an ATS or CRM with 200,000 past applicants for people who fit a new role, often the fastest source of hires for repeat requisitions.
• Sourcing: finding external profiles that fit, a core part of AI sourcing.
• Job recommendations: showing candidates other roles they fit when they apply or visit a career site.
• Staffing: matching available workers to open shifts or assignments by skills, location and availability.

A worked example

A staffing agency gets a job order for 12 medical billing specialists. Its database holds 80,000 candidates. A keyword search for "medical billing" returns 1,900 people, many of them only mentioning it in a course list.

The matching system instead scores candidates on CPT and ICD-10 coding experience, years in billing roles, experience with the client's billing software, location within 25 miles, availability and pay expectation within the bill rate. It returns 140 strong matches ranked by score, with the reason for each. Recruiters call the top 60 and fill the order in four days instead of two weeks.

Notice what made it work: the criteria were set by a recruiter who knew the job, the system explained each match, and humans made the contact and the decision.

Bias and legal risk

Matching tools can screen people out at scale, which makes them high-impact. Risks include:

• Learning from biased history. A model trained on who was hired in the past can learn to prefer the same backgrounds, colleges or employment patterns.
• Proxy variables. Zip code, graduation year, gaps in employment or certain activities can stand in for race, age, disability or sex.
• Rigid requirements. Matching on degree or years of experience the job does not really need excludes capable candidates.

Under Title VII, employers are responsible for disparate impact from selection tools, including vendor tools. A tool that ranks or filters candidates can be an automated decision system under laws such as NYC Local Law 144, which requires an annual independent bias audit and candidate notice, the Colorado AI Act and the EU AI Act. Ask vendors what data the model uses, whether it learns from your decisions, what adverse impact testing they run and whether results are explainable. Rules vary by state and country.

Best practice

• Use match scores to prioritize review, not to auto-reject. If you do auto-reject, limit it to clear, job-related knockouts.
• Define criteria with the hiring manager from the actual job, and remove nice-to-haves that act as hidden filters.
• Check pass rates by demographic group at each stage where matching is used, using the four-fifths rule as a first screen.
• Prefer systems that show why a candidate matched, so recruiters can spot errors.
• Review a sample of low-ranked candidates regularly. If good people are sitting at the bottom, the model or criteria need fixing.