AI Candidate Screening in India: What Works, What Breaks

We build AI screening tools, so treat what follows with the scepticism that deserves. It is still worth writing, because the category is sold with a confidence the technology does not support, and Indian employers are buying it faster than they are evaluating it.

The short version: AI is genuinely good at ordering a long list and explaining its reasoning, and genuinely bad at being trusted to close a candidate's file. Most of the harm in this category comes from giving it the second job because it did the first one well.

What AI screening actually does today

Three quite different things get sold under one label, and they carry different risks:

The third category is where most of the reputational damage in this industry has come from, internationally and now here. Claims about inferring personality or performance from a recorded interview are, at best, contested. Treat any vendor demo that reads traits off a face or a voice as a claim requiring evidence, not a feature. [VERIFY: independent validation studies for inferred-trait hiring assessments are limited and mostly vendor-funded — ask for peer-reviewed, third-party validity evidence before buying]

The failure modes, stated plainly

1. Proxy discrimination

A model trained on your past hiring learns the pattern in that history, including the parts you did not intend. It rarely needs a protected attribute to do this. In India the strong proxies are specific and well known: college tier, English fluency in written text, city of residence, surname, and career gaps. Each correlates with social background more than with ability to do the job.

Career gaps deserve particular attention here, because in India they are strongly gendered. A screening rule that treats continuous employment as a positive signal is, in practice, a rule that penalises women who took time out. Nobody wrote that rule. It emerges.

2. The confident wrong answer

Language models produce fluent justifications for incorrect conclusions. A screening output that says "limited exposure to payments systems" about a candidate who spent three years on UPI reconciliation is not obviously wrong on the page — it reads exactly like a correct assessment. The only defence is that a human can see the claim and check it against the resume, which requires the reasoning to be shown rather than a bare score.

3. Optimisation against the screener

Candidates adapt. Once it is known that a model reads the resume, resumes get written for the model — and the candidates who adapt fastest are the ones with the most access to advice, not the ones best suited to the job. Any screening signal that can be gamed by rewording will eventually be gamed by rewording. Screen for what somebody did and what it produced, not for the presence of terminology.

4. Scale without review

A human screener who is wrong is wrong a few dozen times. A model that is wrong is wrong consistently, at volume, in the same direction, and silently. This is the actual difference in kind, and it is why the rejection decision is the one to keep away from automation.

What the DPDP Act 2023 asks of you

India has no statute specifically governing automated hiring decisions as of 2026. It does have the Digital Personal Data Protection Act 2023, and hiring is squarely within it. The obligations that matter most to a recruiting team:

• Consent must be free, specific, informed, unconditional and for a stated purpose. "They uploaded a resume to a job board" is not consent to being surfaced to your company for an unrelated role.
• Withdrawal must be as easy as granting, and must actually take effect. If a candidate withdraws, they should stop appearing — not stop appearing on the next reindex.
• Purpose limitation. Data collected to assess a candidate for one role is not a general-purpose talent database unless the candidate agreed to that.
• Retention. Personal data should not be kept once the stated purpose is served. Most Indian ATS instances keep everything forever by default.

None of this is specific to AI. All of it becomes sharper with AI, because the reason to hoard candidate data is that a model can do something with it. [VERIFY: DPDP Act rules and enforcement timelines have been phased; confirm the current position with counsel before relying on any specific compliance date]

A practical evaluation checklist

Questions worth asking any vendor in this category, including us:

1. Can I see the reasoning behind every score, or only the score? A number without reasoning cannot be checked, argued with, or defended to a rejected candidate.
2. Does the system ever reject without a human? If yes, ask what the audit trail looks like and who reviews it.
3. What does it do with candidate data after the role closes? Ask for the retention period in writing.
4. Can you show rejection rates broken down by college tier, gender and city? If the vendor has never computed this, they do not know whether their product discriminates. Neither will you.
5. What consent did the candidate give, to whom, and can they withdraw it? If the answer involves a resume scraped from a database, that is your compliance exposure, not the vendor's.
6. How does it fail? Any honest answer to this is more informative than the entire feature list.

How Scout approaches it

Our position follows from the rule above, and it is worth stating what we do not do.

Scout scores every application from 0 to 100 against your specific job description and shows the written reasoning with it — which requirements are met, which are missing, what a recruiter should probe in a screen. The score orders your list. It does not close anybody's file: shortlist and reject are actions a person in your team takes, and when they take one the candidate is told. Silence is what teaches good candidates to stop applying to you, so on Scout's employer side silence is not one of the options.

On data: a candidate is visible to employers only after they have explicitly consented to be, and withdrawal takes effect immediately and unconditionally — a withdrawal is not compared against anything, it simply removes visibility. That is a deliberate design choice rather than a policy statement, because a rule that depends on comparing two timestamps eventually shows a withdrawn candidate to somebody.

And the limits, since we are asking you to ask vendors about theirs: Scout's model reads text, so it inherits every weakness of resume text as a signal. It cannot verify a claim. It has the same blind spots about unconventional careers that a hurried human reviewer has, and it applies them consistently, which is worse. That is exactly why the reasoning is shown and the reject button belongs to you.