AI Recruitment: How to Screen 200 CVs a Day, Objectively

The admin vacancy goes up on Monday. By Wednesday afternoon, applications have poured in from JobStreet, Glints and LinkedIn, plus PDFs staff forwarded on WhatsApp and a few emails from acquaintances. The person sifting through them is the operations manager, after hours, in between their actual job. A few dozen CVs in, their eyes start drifting to the photo, the university name, the year of birth. They don't mean to play favourites. They're just tired.

AI recruitment tools can read hundreds of CVs in a day and hold every applicant to the same standard. The objectivity, though, doesn't come from the AI. It comes from a weighted rubric you write before the first CV is uploaded, from requiring the system to show the evidence behind every score, and from a person who signs off every rejection. Without those, AI just speeds up the same gut feel.

Why isn't AI screening objective by default?

AI models have leanings of their own. In a controlled experiment, large language models preferred CVs they had written themselves over human-written ones 67% to 82% of the time, depending on the model. Applicants using the same model as the employer's screener were 23% to 60% more likely to be shortlisted than equally qualified applicants who wrote their own (Source: Xu, Li & Jiang, University of Maryland, 2026). The same researchers showed that simple interventions cut that bias by more than half.

Another study had four models write around a million CV summaries, changing nothing but the combination of applicant names. The facts in the summaries stayed fairly stable, yet the evaluative tone shifted with the name, in patterns that can slip past an ordinary fairness audit (Source: Nghiem et al., 2026). The lesson is simple: never judge a candidate by a summary paragraph the AI wrote.

Meanwhile, the pile of applications isn't going to shrink. BPS, Indonesia's statistics agency, counted 7.24 million unemployed people out of a labour force of 154.91 million in February 2026, an open unemployment rate of 4.68% (Source: BPS, 2026). Even an ordinary vacancy can draw a flood of applicants.

Recruitment is, unsurprisingly, where AI most often enters HR. A SHRM survey of 1,908 HR professionals found 39% of organisations already use AI in HR, with recruiting the most common use at 27% (Source: SHRM, 2026). Governance is lagging. Only 49% have an AI policy, and 56% of HR functions don't formally measure whether their AI investment is working (Source: SHRM, 2026).

Six steps to AI recruitment for a single opening

These steps are sized for one person screening alongside other work. Start with one role and one rubric; an afternoon is enough to set it up.

Step 1: write the rubric before you open a single CV. Derive the criteria from the job's needs (must-haves, nice-to-haves and soft skills) and weight them to total 100%. Phrase each criterion as evidence, not as a job title. A line like "used to recapping daily orders and stock in a spreadsheet" is far more useful than matching the words "Admin Officer", because the same job can be written as Admin Staff, Admin Officer or Sales Admin, often in two languages at once.

Step 2: bring every application through one door, and mask what must not be scored. Every application, WhatsApp-forwarded CVs included, goes to one place and is scored against the same rubric; a referral CV that skips the rubric is where unfairness creeps back in. Before the AI reads anything, hide the name, photo, date of birth, religion, marital status and address, fields still common on Indonesian CVs. Ministry of Manpower Circular No. M/6/HK.04/V/2025 prohibits discrimination in recruitment and allows an age requirement only where the nature of the job genuinely demands it (Source: Kemnaker, 2025).

Step 3: ask for scores with evidence. Every number should point to the part of the CV that supports it, and every candidate needs a list of gaps alongside their strengths. If the most polished CVs keep rising to the top, be suspicious first. The model may be favouring AI-written text while the best candidate sits further down.

Step 4: review the shortlist and a sample of the rejections. Most teams read only the top candidates. Each round, also pull a few random CVs from the rejected pile; that is where flaws in the rubric show up. No applicant should be rejected by the system alone, not least because Article 10 of Indonesia's Personal Data Protection (PDP) Law gives people the right to object to fully automated decisions. The risk is real: in the United States, a federal court on 16 May 2025 preliminarily certified a collective age-discrimination action against Workday's applicant-screening tools, and the case is still ongoing (Source: Davis Wright Tremaine, 2025).

Step 5: run structured interviews. Every candidate gets the same question bank, interviewers have guidance on what a good answer looks like, and ratings are filled in on the scorecard during the interview. The interview score is then blended with the CV score.

Step 6: keep a record you can defend, covering the rubric version, the scores, who decided and when, and the deletion date for rejected applicants' CVs. Systematically scoring applicants is one of the triggers for a data protection impact assessment under Article 34 of the PDP Law, so this record doubles as compliance evidence (Source: Law No. 27 of 2022, Article 34).

What does this look like in a real system?

We built part of this pattern into TalentLens: a weighted rubric per role, the evidence behind every score, and a structured interview. A single intake point, data masking, rejection sampling and a decision record are the practices we recommend around it. TalentLens surfaces the strongest five candidates within minutes, each with their gaps listed. For the Digital Marketing Specialist role in its demo, the rubric has seven criteria. Paid-ads experience carries 30% and measurable optimisation results 20%, while education and background are deliberately held to just 5%.

During interviews, TalentLens works as a co-pilot. It transcribes the conversation live and suggests follow-up questions, while the interviewer rates five competencies on a scorecard. Every question in the bank carries a hint, such as "look for concrete before-and-after numbers". The interview score is then combined with the CV score into one final ranking. Because the co-pilot records screen and audio, candidates are told and asked for their consent before the interview begins.

Having built it, we think the honest word for a system like this is bias-resistant, not bias-free. The recruiter still evaluates and decides; the system prepares and suggests. Built by Agentiq for an anonymised Indonesian digital marketing agency; all screens shown use fictional sample data.

What this means for your business

For a business without a dedicated recruiter, AI hands time back to whoever does the screening, and that time goes into better interviews. A bad hire is still expensive, but a rejection you can't explain is now a risk too. The EU even classifies AI for recruitment and selection as high-risk, with the compliance deadline pushed back to 2 December 2027 (Source: Gibson Dunn, 2026). That rule doesn't bind businesses in Indonesia, but the direction of travel is clear.

• Pick the role you hire for most often and write its rubric, with weights totalling 100%.
• Route every application channel, referral CVs included, into a single intake point.
• Decide which fields must be masked before any CV is scored.
• Schedule a review of sampled rejections in every round.
• Set a retention period for rejected applicants' CVs, then delete them on schedule.

If you'd like to see what AI-powered candidate screening could look like for your business, walk us through your hiring process on our contact page. The Agentiq team will reply with concrete recommendations, not a template.