AI Resume Screening in India 2026: A Practical Playbook
What AI resume screening actually does to Indian applicants, what the evidence says, what the DPDP Rules require, and a 30-day audit you can run on your own data.
AI resume screening works in India when it ranks and fails when it rejects alone. The evidence, the DPDP timeline and penalties, five fixes and a 30-day audit plan.
The short answer
AI resume screening works in India when it ranks applicants for a human to read and fails when it rejects them on its own. The pressure is real: LinkedIn research reported in January 2026 says applicants per job in India have more than doubled since early 2022, and about 74% of recruiters say qualified talent is harder to find. The legal clock is also running. India's DPDP Rules were notified on 14 November 2025, most obligations apply from about May 2027, and penalties reach Rs 250 crore. Test every screen on your own applicants, mask proxies such as PIN code and school name, and keep a recruiter accountable for each rejection.
Running an AI screen on Indian applicants and unsure what it is filtering out? Book a demo and we will walk through an audit of your funnel with you.
Key findings
- Volume is the problem AI is being hired to solve. LinkedIn research reported by People Matters and Storyboard18 says applicants per job in India have more than doubled since early 2022, 74% of recruiters find qualified talent harder to find, and about 80% of recruiters plan to use more AI in screening.
- Recruiters report value from AI, with a catch. Among recruiters already using AI, 71% said it helped them find candidates they might have overlooked, and half said candidates want more transparency about how AI is used in screening.
- Bias evidence for India is thin, and the strongest study points at language, not surnames. In a 2025 AIES paper from the University of Lausanne and IIIT Bangalore, GPT-4o and Gemini Flash 1.5 scored Indian interview transcripts lower than UK ones. Swapping names that signal caste or region did not significantly change scores.
- Human hiring bias in India is documented. A 2007 audit study by Thorat and Attewell found applicants with Dalit surnames received callbacks at roughly two thirds the rate of applicants with upper-caste surnames.
- The law is arriving in phases. DPDP notice, safeguard, breach and rights rules apply 18 months after publication, and the penalty ceiling is Rs 250 crore.
Scope and method
This guide covers AI tools that parse, score or rank resumes for Indian hiring teams. It does not rank vendors; for that see our evidence review of AI candidate screening tools and our guide to AI recruiting software in India. Every figure below names its source. The LinkedIn figures come from press reports of LinkedIn research that did not publish the sample size or method. Earlier versions of this page said Indian rollouts silently reject 35 to 45 percent of the best candidates, and cited 3,000 plus hiring projects. We could not trace either figure to a source, so we removed them. Updated 15 October 2026. This is not legal advice.
What Indian hiring teams are dealing with
LinkedIn's research, as reported in January 2026, describes a market where both sides are stretched. Applicants per job have more than doubled since early 2022. Recruiter hiring activity is about 40% above pre-pandemic levels. Nearly half of recruiters say separating genuine applicants from low-quality or misleading ones slows hiring, and more than half of those who say hiring got harder blame the rise in AI-generated applications. On the job seeker side, 84% say they feel unprepared and 94% plan to use AI in their search.
That is the demand for AI screening. It explains why teams adopt it. It does not show the tools pick the right people, which is the question the rest of this guide deals with.
Where AI screening goes wrong for Indian applicants
Language and style can pull scores down
The best-designed study we found tested large language models on 200 transcribed job interviews, 100 from Indian students and 100 from UK participants. Indian transcripts scored significantly lower on most measures, and region stayed a significant predictor after controlling for sentence length and readability. Median scores were about 2 to 2.75 for Indian transcripts and 2.75 to 3.25 for UK transcripts. The same study found that substituting names signalling caste or region in 50 transcripts did not significantly change scores. The authors concluded that names alone did not reliably move the result.
Two cautions. This was interviews, not resumes, with student participants and two models. So treat it as a warning about style and phrasing, not a measured error rate for resume screening.
Proxies can stand in for protected traits
Removing a sensitive field does not remove the signal. A 2022 European Union Agency for Fundamental Rights study, as summarised by The India Forum, reportedly found systems fall back on proxies such as neighbourhood and institutional background when a sensitive variable is removed. We found no published study that measures caste inference from PIN code, school or first name in Indian resume screening. That is an evidence gap, not proof of safety, and it is why an audit on your own data matters.
"We do not know how much your tool relies on PIN code or school name. You can find out in an afternoon by masking them and re-running last quarter's applicants."
What the DPDP Act means for resume screening
Chart 1 shows the penalty ceilings under the Digital Personal Data Protection Act, as summarised by EY India. Chart 2 shows when the Rules take effect.
What this means in practice:
- Employment ground. The Act allows processing for employment purposes or to safeguard an employer from loss or liability. Commentary from ELP Law says the wording lacks an explicit necessity test, and expects courts and regulators to apply proportionality. The commentary we reviewed does not settle whether the ground covers job applicants, so a clear notice and, where you can, consent is the safer route.
- Accuracy and explanation. Section 8 of the Act is the provision The People's Board guide ties to explaining a rejection in plain language. A score you cannot explain is a liability.
- Breaches. Rules call for notifying the Board within 72 hours of becoming aware of a breach, per EY. A resume database is personal data.
- Retention. Set a deletion timeline for rejected applicants, and keep processing logs for at least one year, per EY.
Five fixes that work
These are our recommendations, based on the evidence above, not vendor claims.
- Rank, do not reject. Use scores to decide who a recruiter reads first. Automate rejection only for legal must-haves such as a required licence or work authorisation.
- Mask and test proxies. Re-run a quarter of applicants with PIN code, school name and graduation year hidden. A large shift in who is shortlisted means the model leans on them. The People's Board audit guide recommends the same test and does not define what counts as a meaningful shift, so set your own threshold in advance.
- Check selection rates by group. Compare shortlist rates and flag any group below 80% of the highest group's rate. This four-fifths test comes from US employment practice. We found no Indian statutory equivalent, so treat it as an internal alarm, not a legal safe harbour.
- Read the tail. Every week, have a recruiter read 20 to 30 applications from the bottom of the ranking and log how many they would have advanced. Our guide to AI screening false negatives explains how to read the result.
- Tell candidates and keep records. Say plainly that AI helps screen, who reviews the output, and how to ask for human review. Half of recruiters in LinkedIn's research say candidates want that transparency.
A 30-day audit plan
- Days 1 to 5: inventory. List every tool that scores, ranks or filters candidates, and note whether each makes the final call or only assists.
- Days 6 to 12: baseline. Pull applicant, shortlist and hire counts by the groups you can lawfully measure for the last two quarters.
- Days 13 to 20: proxy test. Mask PIN code, school and graduation year, re-run, and record the shift.
- Days 21 to 26: tail review. Read the weekly bottom-of-list sample and log missed candidates.
- Days 27 to 30: decide. Fix settings, retrain, add human review for borderline scores, or replace the tool. Put the next audit date in the calendar.
Our argument in hiring intelligence, not more tools explains why the audit comes before the purchase.
Questions to ask any vendor
- Can I run your tool on 200 of my own past applications before I sign?
- Which fields does the model use, and can I mask PIN code, school and graduation year?
- Can a recruiter see why a candidate ranked where they did, and override it?
- Where is applicant data stored, how long is it kept, and what happens within 72 hours of a breach?
- Will you share bias audit results and the date of the last one?
See how TheHireHub compares on these points, or read about our AI recruitment software and agentic AI recruiting.
The one thing to take from this
Speed is easy to measure and bias is not, so teams measure speed. The fix is a habit, not a purchase: test on your own applicants, mask what should not matter, and keep a person accountable for every rejection. If you would like help running that audit, book a demo or try the platform on one open role.
Audit your screening funnel
Share one open role and we will show you who your current screen ranks first, who it ranks last, and why.
Frequently Asked Questions
It works for ordering a queue, and LinkedIn research reports that 71% of recruiters using AI say it helped them find candidates they might have overlooked. It is not proven as an automatic rejector. Test it on your own applicants and keep a recruiter accountable for every rejection.
We could not find a published study that measures this for resumes. The closest evidence is a 2025 AIES paper in which GPT-4o and Gemini Flash 1.5 scored Indian interview transcripts lower than UK ones, even after controlling for sentence length and readability. That points to phrasing and style effects, so run a test on your own applicants.
The evidence is mixed. A 2007 audit study found human employers called back applicants with Dalit surnames at roughly two thirds the rate of upper-caste surnames. In the 2025 AIES study, swapping names that signalled caste or region did not significantly change LLM scores. Proxies such as PIN code or school are untested in published work, so mask them and test.
The DPDP Rules were notified on 14 November 2025. Notice, safeguard, breach and rights rules apply 18 months later, about May 2027, and penalties reach Rs 250 crore. The Act allows processing for employment purposes, but commentary expects a proportionality test and does not settle whether that covers applicants, so give a clear notice.
Inventory every tool that scores or filters candidates, pull shortlist rates by group for two quarters, and flag any group below 80% of the highest group's rate. Then mask PIN code, school and graduation year, re-run the same applicants, and compare who is shortlisted. Finish with a weekly human read of the bottom of the ranking.
We found no Indian statutory equivalent. It is a US guideline under which a group selection rate below 80% of the highest group's rate is generally treated as evidence of adverse impact. Indian teams use it as an internal alarm, not as a legal safe harbour.
Not before you audit. Many problems come from settings, proxy fields and how scores are used, not from the vendor. Run the 30-day audit first, and if a vendor will not let you test on your own data or explain a ranking, that is a good reason to switch.
Set a defined deletion timeline for each purpose rather than keeping resumes indefinitely. Under the DPDP Rules, EY notes that erasure timelines are purpose-specific and that processing logs must be retained for at least one year. Confirm your exact schedule with counsel.
