All posts

Blog

How AI candidate screening works (and where it fails)

How AI candidate screening works, from reading the resume to the first interview, and where it fails: undersold resumes, biased criteria, gaming and over-trust.

By the Haira team · Published · 5 min read

AI candidate screening is software doing the first pass on applicants: reading resumes, asking the first questions and, in some tools, holding the first interview, then telling a recruiter who looks worth meeting. Done well, it gives every applicant a fair read when a person would only have time to skim. Done badly, it rejects good people for reasons nobody can explain. This post walks through how it works, step by step, and then where it goes wrong.

Step 1: reading the resume

Screening starts by turning a resume into something software can compare with a job: roles and dates, skills, education, and what the person actually did in each job. This sounds mechanical, and it is where many tools differ most.

The weak version counts keywords. A resume that repeats "Python" ten times beats one that mentions it once inside three years of real Python work. That rewards people who know how to write for software, not people who can do the job.

The stronger version reads for evidence. A skill shown across real roles counts for more than one sitting in a skills list, and where the resume gives no evidence for a skill, the software should not assume it. Haira's resume ranking works this way, and it also reads resumes in the terms of the job's industry: the same three letters, "API", mean an interface in software hiring and an active pharmaceutical ingredient in pharma. The industry pages show what it looks for in each field.

Step 2: scoring against the role

Next, each applicant is scored against the specific job. Good screening lets the hiring team decide what matters most for each role (core skills over tenure for one job, the reverse for another), because a shortlist should reflect that role's priorities, not one fixed formula.

Two details separate a usable score from a number:

  • A reason for every score, taken from the resume, so a recruiter can see why someone ranked where they did and disagree.
  • Questions, not penalties. Overlapping jobs, a date that does not add up or a career gap are worth asking about. Haira flags them for a person to read rather than lowering the score, because a date inconsistency is a question, not a verdict.

The full detail is in how AI resume ranking works.

Step 3: asking the first questions

A resume does not say whether someone is still interested, when they could start, what they expect to be paid, or whether they can work where the job is. Those are the questions of a recruiter's first call, and AI screening increasingly asks them too, by chat, by text or by phone.

Haira's AI phone screen calls the candidate's own phone, checks it is speaking to the right person, and asks those questions one at a time. The answers are saved as fields you can filter, and a question the candidate did not answer stays empty instead of being guessed. Each call ends with a strong, maybe or weak fit, always with a reason. What is AI phone screening? covers a call from start to finish.

Step 4: the first interview

Some tools go a step further and hold the first interview. This can be a recorded video the candidate answers alone, a chat, or a live conversation with an AI interviewer. In Haira's AI video interview, the questions are written for each candidate from the job and their resume, and the interviewer follows up on what they actually say. Every candidate for a role is scored on the same four areas (job fit, technical depth, communication and confidence), and the score comes with the full transcript.

Step 5: a person decides

The output of screening should be a shortlist with reasons, not a decision. By default Haira scores and recommends and a person decides. If you switch on Auto-Process for a job, it rejects and advances candidates on their scores, so that is a choice to make deliberately, role by role.

Where AI candidate screening fails

Every one of these failures is real, including in the tools that handle them best. Knowing them is how you use screening well.

A resume that undersells good work

Screening reads what is written. Someone who did excellent work and described it in two vague lines will rank below someone who did ordinary work and described it precisely. Career changers, people returning from a break and people from less formal backgrounds are most exposed. The fix is human: read below the top of the list for roles where that matters, and use the interview to give people a second chance to show their work.

Criteria that encode bias

If the priorities for a role quietly favour one kind of background, software will apply them faithfully to every applicant. Any system that scores people can encode bias, and a vendor claiming otherwise is the one to distrust. What helps is criteria you can see and change, one format and rubric for every candidate, and a transcript behind every score. Is AI screening biased? sets out what Haira has in place, and what it does not, such as an outside bias audit.

Gaming the system

People write resumes for software: stuffed keywords, copied job descriptions, even hidden instructions meant for an AI. Reading for evidence blunts keyword stuffing, and in Haira a resume that tries to trick the AI with hidden instructions is caught and not ranked. No screening is immune to a well-written false claim, which is why the interview checks what the resume claims against what the candidate says.

Treating the score as the answer

A score is a reading order, not a judgment of a person. It cannot tell you whether the work described was done well, whether an achievement is real, or whether someone will do the job. Teams get into trouble when they stop reading the reasons and only sort by the number.

Roles a resume cannot judge

For some roles, a short work sample tells you more than any resume score: writing, design, some sales roles. Screening still saves time on the basics, but the decisive evidence comes from the work. AI screening for non-technical roles goes through where that line sits.

Candidates who never respond

A phone screen only works if the candidate picks up. Good screening retries at sensible hours, offers a callback, and hands anyone who does not answer to a recruiter rather than quietly dropping them.

Using it well

AI candidate screening is at its best when it does the reading nobody has time for and shows its working, and a person makes the call. Set the priorities for each role, read the reasons rather than just the scores, look below the top of the list when the role calls for it, and keep the transcript next to every decision.

To see it on one of your roles, start free, or book a 30-minute walkthrough first.

Hire faster with Haira. Start free today.

Free planNo card neededCancel anytime