Nearly every applicant tracking system sold today has AI on the feature list, and most of it does the same three things. It parses a resume, scores that resume against a job description, and drafts an email. All three are useful. None of them is what the pitch implies, and the distance between the two is where a hiring team loses a quarter and a budget.
The useful question is not whether to use AI in recruitment. That decision has mostly been made for you by the tools you already pay for. The question is which parts of your funnel are safe to hand over, which parts need a human at the end of them, and what you now have to be able to prove about both if a candidate, a client or a regulator asks.
This piece goes through the recruitment funnel in order. Sourcing, screening, scheduling, interviewing, assessment, communication, offer and measurement. For each one: what AI does well enough to rely on, where it fails, and what you have to keep human. Then the disclosure and data rules that now sit around all of it, and the ways these projects usually go wrong.
One thing up front. The single most important distinction in this whole subject is between steps that add candidates and steps that remove them. AI that adds is cheap to get wrong, because a bad suggestion costs you a minute of review. AI that removes is expensive to get wrong, because a bad rejection costs you a hire you will never know you missed, and in some jurisdictions a claim you will find hard to defend. Almost every recommendation below comes out of that one line. If you want the wider view of AI across the HR function rather than hiring alone, our piece on AI in HR covers the rest of the employee lifecycle.
What follows is shaped by the hiring funnels we configure. Engage runs recruitment for more than 200 companies hiring a few thousand roles a year between them, and the pattern that repeats is that the AI budget goes to screening while the bottleneck sits in scheduling and feedback. Where something below is flagged as a common failure, that is what we have seen in real funnels rather than a guess about what might be in them.
This is general guidance for employers, current as at the date above. It is not legal advice. AI and employment rules are moving quickly in the EU, in several US states and in India, and dates in particular have been amended more than once, so confirm anything jurisdiction-specific with your advisor before you act on it.
Where AI Helps, and Where It Doesn't
The honest summary is that AI in recruitment is very good at volume and unreliable at judgement, and most disappointment comes from buying it for the second thing.
Volume work is anything where the task is the same every time and the cost of a small error is a minute of somebody's attention. Reading 900 resumes into structured fields. Finding the eleven people in your own database who once reached a final round for a similar role. Working out which of four interviewers has a free hour that also suits the candidate. Writing the first draft of a job description. None of that is a decision about a person, and all of it is work your recruiters currently do instead of talking to candidates.
Judgement work is anything where the output is a conclusion about somebody's suitability. Scoring, ranking, rejecting, inferring personality, predicting performance. AI can contribute to those, and it is genuinely useful in a supporting role, but it cannot be the last step, and the reasons are practical before they are legal. A model trained on your past hiring reproduces your past hiring. If your last three years of engineering hires came from the same six colleges, a model that learns what a good candidate looks like has learned what those six colleges look like.
The split runs task by task, and it decides who has to be able to explain the outcome.
| Recruitment task | What AI is genuinely good at | Where it fails | Who decides |
|---|---|---|---|
| Sourcing and rediscovery | Semantic search across resumes, matching on skills rather than job titles, surfacing past applicants | Reproduces the profile of who you hired before, so it narrows the pool while appearing to widen it | Recruiter, on an expanded list |
| Resume parsing | Turning unstructured resumes into consistent fields, deduplicating applicants | Non-standard formats, Indian resumes with tables and photos, multi-page academic CVs | Nobody; verify a sample |
| Screening and ranking | Ordering a review queue, applying explicit knockout rules, flagging missing information | Scoring on proxies for college, gender, age and career gaps; unexplainable rejections | Recruiter, on every rejection |
| Scheduling and coordination | Panel availability, reschedules, reminders, interviewer load balancing | Almost nothing, which is why this is the safest place to start | Nobody |
| Assessments and interviews | Structured note-taking, transcription, mapping answers to a rubric, work-sample scoring | Personality and emotion inference, video-based scoring, anything without a job-related rubric | Interviewer, on the record |
| Candidate communication | Drafting outreach, answering process questions from your own content, rejection at volume | Fabricated policy answers, obviously templated personalisation | Recruiter, on anything committal |
| Offer and documentation | Letter generation, CTC breakups, document chasing, verification packs | Pay recommendations trained on your historical pay | HR and hiring manager |
| Analytics | Funnel drop-off, source performance, stage-wise selection rates | Explaining causes it cannot see, small sample noise | Whoever owns the hiring plan |
Read the last column first. Where it says nobody, automate freely. Where it names a person, that person has to be able to say why, in words, without pointing at a score.
Sourcing and Candidate Rediscovery
Sourcing is where AI earns its keep with the least risk attached, because sourcing adds people to the pool rather than taking them out.
The first use case is semantic search over job boards and profiles. Boolean strings match words; a language model matches meaning, which matters more than it sounds in Indian hiring where titles are close to meaningless. The same job is a Software Engineer III at one company, a Member of Technical Staff at the next and a Senior Associate at the third. Searching for the title misses two of them. Searching for what the person actually did finds all three.
The second use case is the one most teams own already and almost nobody uses, which is rediscovery inside your own database. Every applicant tracking system with a few years of history contains people who reached your final round for a similar role and lost to somebody marginally better, people who declined an offer for reasons that have since changed, and people who applied for the wrong opening. They have already been assessed by your team and they already know your company. Searching that database properly is the highest return sourcing use of AI available, and it costs nothing in candidate acquisition. When we run it against a customer database for the first time, a search on a single open role typically surfaces a dozen or more previously assessed candidates who were never re-contacted, and several roles have since closed from that list rather than from new sourcing. Before you build it, read section 8, because how long you are entitled to keep those resumes is a live question under the Digital Personal Data Protection Act and the answer constrains the whole idea.
The third is skill adjacency. Given a role, a model can suggest the roles people usually move from into it, which is how teams find the support engineer who should be looked at for a solutions role, or the person from a different industry with the same underlying skill. This is genuinely useful for hard-to-fill roles, and it is the direct opposite of what a keyword filter does.
Now the failure. Similar to this profile is the most dangerous button in a sourcing tool. Similarity is measured against people you already hired or already liked, so it points backwards. If you use it, use it alongside a search built from the job requirements themselves rather than instead of one, and check what the two lists have in common. Where they diverge is usually where the interesting candidates are. Our piece on recruitment software and diverse hiring goes further into how the pool composition decides the outcome long before anyone runs an interview.
The operating rule for this section: use AI to expand the pool, never to narrow it. Expansion errors cost review time. Narrowing errors cost hires.
Screening and Ranking Applications
This is the use case everybody buys, the one that produces the visible saving, and the one under legal scrutiny in three jurisdictions. It can be done properly. It requires separating two things that vendors sell as one.
Knockout criteria are rules. Does the candidate have the licence or certification the role legally requires, are they willing to relocate to the posting location, is their notice period inside the window, do they have work authorisation. These are objective, verifiable and binary, and they should be implemented as rules rather than as model scores, because a rule can be written down, shown to a candidate and audited afterwards. Be careful about what you put in this list, though: notice period and current CTC are common knockouts in India and both correlate with things you probably do not intend to select on. A CTC floor filters out people who took a pay cut for caregiving, and returners are disproportionately women.
Ranking is a model score. Its correct use is to order the review queue, so that the resumes most likely to be relevant get read first by a human who is fresher at resume forty than at resume four hundred. Its incorrect use is to end the process for everyone below a cutoff. The distinction sounds academic until you look at what it costs to defend. If a rejected candidate asks why, ordering a queue produces a human answer. A cutoff produces a number, and a number is not a reason.
There is a third use, less discussed and better than either, which is using AI to remove information rather than to judge it. Indian resumes routinely carry a photograph, date of birth, marital status, father's name, religion, and a surname that carries caste information, none of which is relevant to any role and all of which reaches the reviewer's eye before the experience section does. A parser can strip those fields before a human ever sees the file, and can hold them separately for the diversity reporting you may need later. This is a genuinely useful application of the technology to a real problem, and it is one of the few AI screening changes that reliably improves the decision rather than accelerating it.
Two things to build in from the start.
Every rejection needs a reason category selected by a person, recorded against the candidate, and consistent across the team. Not a score. Six or eight categories are enough. This costs a click and it is the single most valuable record you will have if anyone ever queries your process, including your own leadership.
And you should measure selection rates by stage and by group. The standard test in US practice is the four-fifths rule: if the selection rate for any group is less than 80% of the rate for the highest-selected group, that is treated as evidence of adverse impact worth investigating. You do not need to be subject to US law for the arithmetic to be useful. It is the fastest way to find out that your new screen quietly changed who gets through.
The cautionary history is short and worth knowing. Amazon abandoned an internal resume-screening model in 2018 after finding it had learned to downgrade resumes containing the word women's, from a training set of ten years of mostly male technical hires. More recently, the litigation against Workday in the United States over age discrimination in algorithmic screening has proceeded as a collective action, on the argument that a vendor supplying the screening tool can itself be liable as an agent of the employers using it. In March 2026 the court also confirmed that applicants aged 40 and over can bring these claims at all, rejecting the argument that the protection runs only to people already employed. Whichever way that ends, it does not move the employer's own exposure anywhere. If your tool rejects someone, you rejected them.
Scheduling and the Boring Wins
If you have one quarter to show that AI in recruitment was worth funding, spend it here.
Interview coordination is pure volume work with no selection decision anywhere in it. A scheduling agent that reads panel calendars, offers the candidate real slots, handles reschedules, sends the interview kit and the joining link, chases confirmations and re-books no-shows carries no adverse impact risk at all, because it never expresses a view about anybody. It also removes the task recruiters most reliably hate and most reliably get blamed for.
The related wins in this category are the same shape. Automatic interviewer load balancing, so the same three senior engineers are not carrying every panel. Interview kits generated per role and per stage, so the interviewer arrives with the rubric rather than improvising. Reminder sequences to both sides. Automatic follow-up when feedback has not been submitted within two days, which is the real reason your time-to-hire looks bad.
Track three numbers before and after: median hours from screen-pass to interview scheduled, candidate no-show rate, and the spread of interview load across your panel. All three move quickly, all three are unambiguous, and none of them requires anybody to argue about model quality.
Two of them move first in the funnels we migrate. The gap between screen-pass and interview scheduled usually starts at a week or more and lands inside 24 to 48 hours once coordination is automated. Interviewer feedback submitted within two days starts below 40% and settles between 75% and 90% once the chasing runs itself. Record both before you change anything, because they are the numbers that will justify the rest of the programme.
Interviews, Assessments and What Not to Automate
Assessment is where the technology is most aggressively marketed and least well evidenced, so it is worth being specific about which parts hold up.
What works is structured support for a human interviewer. Recording and transcribing, with consent, and mapping what the candidate said onto a competency rubric so the scorecard is filled from evidence rather than from memory an hour later. Flagging that a required question was never asked. Producing a summary the next interviewer reads before the loop, so the candidate is not asked the same three questions four times. This raises the quality of interviews mostly by making them consistent, and consistency is the part of interviewing that reliably predicts performance.
What works less well but is defensible is scoring on job simulations and work samples, where the task looks like the job and the scoring rubric is written down in advance. A grading model applied to a structured exercise with a published rubric is arguable in a way a personality inference is not.
What does not work, and in one jurisdiction is banned outright, is inferring character from a face or a voice. Emotion recognition in the workplace is a prohibited practice under Article 5(1)(f) of the EU AI Act, and that prohibition has applied since February 2025. It is not a high-risk-with-conditions category, it is on the prohibited list, and recruitment falls inside the workplace scope. The exception for medical and safety purposes is narrow enough to cover driver fatigue monitoring and not much else, and the AI Omnibus that deferred the Act's high-risk deadlines in July 2026 left the prohibition exactly where it was. Even outside the EU the underlying claim is weak: scoring candidates on facial expression, speech cadence or word choice systematically penalises people with accents, speech differences, non-native fluency and several disabilities, which is a discrimination exposure attached to a signal that was never worth much anyway.
There is a newer problem in this section that has nothing to do with your tools. Candidates use AI too, and the traditional take-home assignment has stopped carrying information. A polished submission now tells you almost nothing. The workable responses are to move the exercise into a live conversation, or to keep the take-home and then spend twenty minutes asking the candidate to explain, extend and debug their own submission. Someone who did the work can do that immediately. What you should not do is buy an AI-detection tool and act on its output. Detection accuracy on text is poor, the false positives are not randomly distributed, and accusing a candidate of cheating on the word of a classifier is a far worse outcome than a bad hiring signal.
Job Descriptions and Candidate Communication
Job description drafting is the most common first use of AI in recruitment and the one that needs the most editing, for a reason nobody warns you about. Generated job descriptions inflate. They add degree requirements, extra years of experience and long responsibility lists because that is what the average posting they learned from contains, and every one of those additions shrinks your applicant pool before you have met anybody. Use the draft for structure and then cut it. A useful discipline is to require a hiring manager to justify each stated requirement as something the person must have on day one, and to move everything else into a preferred list or delete it.
The same tools are good at the things you should check for anyway. Age-coded and gender-coded language, and the phrases that are common in Indian postings and are straightforwardly discriminatory, including requests for age, marital status or a photograph, and roles advertised as suiting young and dynamic candidates. Getting those out of the posting is a small use case with an outsized effect on who applies.
For outreach, personalisation at volume is real but has a ceiling. A message that references the candidate's actual work outperforms a template, and a message that is obviously a template with a variable slotted in performs worse than a short honest one. Set the bar at: would I send this if I had written it myself.
Candidate-facing chat is worth doing only if it is grounded in your own material. A bot answering process questions from your careers content, your interview process page and your policy documents is useful and mostly accurate. A general model asked about your notice period, your work-from-home policy or your compensation bands will invent an answer, and a candidate will hold you to it. Ground it, restrict it to process questions, and give it a route to a human as the first fallback rather than the last.
The largest single win in this section is the least technical. Most companies do not respond to most applicants, because the volume makes it impossible by hand. It is no longer impossible. Sending every applicant an actual outcome, with a reason category rather than silence, is the highest-value thing AI has made feasible in recruitment communication, and it is the one candidates notice.
Offer, Documentation and the Handover
The last stretch of the funnel is mostly document work, which is exactly what this technology is for.
Offer generation from an approved template, with the CTC breakup computed and the components mapped to your salary structure, removes a slow manual step and a common source of error. Our offer letter generator does this part, and the structure underneath it, which decides everything from PF liability to leave encashment values, is worked through in our guide to employee salary structure in India.
Background verification and document collection is the other obvious candidate. Chasing missing documents, reading uploaded ones for completeness, flagging a mismatch between the resume and the verification report, and keeping candidates informed while it runs. It is entirely coordination work.
The handover to onboarding is where most of the value quietly sits, because the data collected during hiring is the data payroll and HR need next, and it is usually retyped. If the candidate record becomes the employee record without re-entry, you have removed a day of work and a class of errors at the same time. That is a system design question more than an AI one, and it is covered in our piece on the employee lifecycle in HR software.
One thing not to automate here. Pay recommendation models trained on your own historical offers will reproduce your historical pay gaps, and will do it with a confident number attached. Benchmark against external market data and your published bands, and keep the decision with a human who can be asked about it.
What You Owe Candidates
Two years ago this section would have been short. It is not any more, and the position differs sharply between where you hire and where your candidates are.
| Where you hire | What applies | What it requires in practice |
|---|---|---|
| European Union | EU AI Act; GDPR Article 22 | Emotion recognition in the workplace has been prohibited under Article 5(1)(f) since February 2025, and the AI Omnibus of July 2026 left that prohibition untouched. Recruitment and selection systems are listed as high risk in Annex III, bringing risk management, data governance, logging, human oversight and transparency duties, and those obligations were deferred from 2 August 2026 to 2 December 2027 by Regulation (EU) 2026/1744. That is preparation time rather than a reprieve. Separately, GDPR Article 22 gives candidates a right not to be subject to a solely automated decision with significant effects, and a right to human intervention. |
| New York City | Local Law 144 | Automated employment decision tools require an independent bias audit within the previous year, publication of the audit summary, and at least ten business days' notice to candidates that the tool is being used and what it assesses. |
| Illinois | AI Video Interview Act; amendment to the Human Rights Act effective 1 January 2026 | Consent, explanation and deletion rights for AI-analysed video interviews, and a general prohibition on using AI that discriminates in employment decisions, including on proxies for protected characteristics. |
| Colorado and other US states | State AI acts, various | Colorado's AI Act, once the strictest state law in this area, was delayed twice and then rewritten. SB 189, signed in May 2026, moved the effective date to 1 January 2027 and removed the duty of reasonable care against algorithmic discrimination along with the deployer risk-management and impact-assessment obligations that gave the original law its force. Other states continue to legislate. Treat both the dates and the substance in this group as unsettled. |
| United States, federally | Title VII, ADEA, ADA | No AI-specific statute. Existing discrimination law applies to the outcome regardless of how the decision was produced, and the four-fifths rule is the usual first test of adverse impact. |
| India | Digital Personal Data Protection Act, 2023; Rights of Persons with Disabilities Act, 2016 | No AI-specific employment statute. The binding constraints are data protection ones: notice and consent for collecting candidate data, use limited to the stated purpose, retention no longer than needed, and erasure on request. The DPDP Rules were notified in November 2025 on a phased timeline, and the notice, consent, data principal rights and grievance provisions land at the end of an 18-month transition in May 2027. The RPwD Act requires an equal opportunity policy for covered employers, which your assessment process has to be consistent with. The wider obligations sit in our labour law compliance checklist. |
The Indian row deserves a paragraph, because it is the one most readers are actually in and the one written about least honestly.
India has no general anti-discrimination statute covering private employment the way Title VII does in the United States. That is a real gap and it means the legal pressure on AI screening here is weaker than the headlines suggest. What does bind you is data protection. Candidate resumes are personal data, and the Digital Personal Data Protection Act requires a clear notice of what you collect and why, consent for that purpose, use confined to it, retention only as long as the purpose lasts, and deletion when it does not. Three consequences follow directly and most hiring teams have addressed none of them.
Your resume database has a retention question attached to it. The rediscovery use case in section 2 depends on holding candidate data for years after the application it was collected for. That is defensible only if you asked for it, told the candidate you would, and let them ask you to stop. Get the consent language right at the point of application and the whole use case becomes clean.
Sending candidate resumes to a third-party model is processing by another party. If your screening runs through an external API, that vendor is handling personal data on your behalf and the contract has to say so, including what they may do with it. Assume by default that anything sent to a consumer AI tool has left your control, and do not let recruiters paste resumes into one.
Erasure requests have to reach the model layer, not only the ATS. If a candidate asks to be deleted and your vector index or your enrichment tool still holds their profile, you have not deleted them.
Beyond the law, there is a straightforward practice position that most Indian employers can adopt without waiting for anybody. Tell candidates where AI is used in your process. Do not let it reject anyone by itself. Give a real route to a human. If you hire for or alongside multinational clients, expect this to be asked of you contractually well before it is asked of you legally.
Measuring Whether It Worked
Most AI recruitment pilots cannot be evaluated, for the same reason most HR projects cannot: nobody wrote down the before. Spend the first two weeks recording the baseline and the rest of the exercise gets easier.
| Metric | How to define it | What it tells you |
|---|---|---|
| Time to hire | Days from application to offer acceptance, per role family | The candidate's experience of your speed. Moves with scheduling automation |
| Time to fill | Days from requisition approval to acceptance | The business's experience of your speed. Moves with sourcing |
| Screens per recruiter per week | Completed human reviews, not resumes ingested | Whether the tooling created capacity or just created a queue |
| Interview to offer ratio | Onsite interviews per offer made | Whether screening quality improved or only screening speed |
| Offer acceptance rate | Offers accepted over offers made | Catches a fast process that is selecting the wrong people |
| Quality of hire at six months | Manager rating plus retention at six months, tracked by source. Needs performance data to be usable | The only measure that says whether the screen works. Slow, and worth the wait |
| Stage-wise selection rate by group | Passes over entries at each stage, split by the groups you can lawfully measure | Adverse impact. Apply the four-fifths test stage by stage, not only at offer |
| Rejected-resume audit | A random sample of AI-deprioritised applications, read by a human blind to the score | False negatives, which no other metric in this list will ever show you |
The last row is the one nobody runs and the one that matters most. Every other metric measures the people who got through. A screening model can look excellent on all of them while quietly discarding people it should not, and the only way to detect that is to go and read what it threw away. Pull fifty rejected applications a month, have someone read them without seeing the score, and record how many they would have advanced. If the answer is consistently more than a handful, your cutoff is wrong, and you have found it in a month rather than a year.
The broader point about running the function on numbers instead of impressions is in our piece on data-driven HR.
Where AI Hiring Projects Fail
These are roughly in the order they turn up, not in order of severity. The most expensive one is fifth.
No baseline. Close to universal. If you did not record time to hire, screens per recruiter and offer acceptance before the pilot, you will not be able to say whether it worked, and the conversation about renewal will be decided by whoever is most confident rather than by whoever is right.
Resume data leaving without a contract. Recruiters pasting candidate resumes into consumer chat tools is the most common data protection failure in hiring right now, and it is happening in most companies whether or not anybody has approved it. It is a training problem more than a technology one, and it is fixed by giving people a sanctioned tool that does the same job.
Buying a score you cannot explain. If the vendor cannot tell you what the model uses, you cannot answer a candidate, an auditor or your own general counsel. Ask what features the score is built from, and ask what happens to the score when the same resume is submitted with a different name. Ask before you buy, because you will not get an answer after.
Two systems holding different truths. When the AI layer and the ATS disagree about a candidate's status, candidates get two rejections, or an interview invitation after a rejection, or nothing at all. It is embarrassing in a way that outlasts the tool. This is the case for one system of record rather than a stack of point solutions, and the same argument as replacing spreadsheets with a single system.
Automating the decision instead of the queue. Less common than the four above and more expensive than all of them. Ordering the work is nearly all of the benefit and almost none of the risk. Ending the process is the reverse.
Optimising the funnel while ignoring the offer. Faster screening does not fix a low acceptance rate. If candidates are declining, the constraint is your compensation, your process length or your interviewers, and no amount of AI in the top half will move it.
No route back to a human. Every automated step needs an exit. A candidate who cannot reach a person, or whose unusual background does not fit the parser, is currently being lost silently.
Training on your best hires. Rarer, because most companies buy a model rather than build one, and serious where it happens. It is the natural instinct and it is how you automate your existing bias with a confidence interval attached. Where you use historical data at all, validate against outcomes rather than against who got hired.
Treating the vendor's bias audit as yours. Last on this list mostly because few employers have any bias audit to misread. The vendor tested their model, usually on their data. Your obligation, where one applies, attaches to your use of the tool with your candidates. Run your own selection-rate analysis on your own funnel.
Questions People Ask
Can AI reject candidates automatically?
Technically yes, and in most jurisdictions it is not directly prohibited, but it is the highest-risk configuration available to you and there is little upside in it. In the EU, a decision with significant effects made solely by automated means engages a candidate's right to human intervention under GDPR Article 22. Elsewhere, an unexplainable rejection is simply hard to defend to anyone who asks. The practical standard is to let the model order the review queue and to require a person to confirm every rejection with a recorded reason category.
Is AI recruitment legal in India?
Yes. There is no AI-specific employment statute in India, and no general private-sector anti-discrimination law equivalent to Title VII. The binding obligations are data protection ones under the Digital Personal Data Protection Act, 2023: notice, consent, purpose limitation, retention only as long as needed, and erasure on request. The Rules were notified in November 2025 on a phased timeline, with the notice, consent and data principal rights provisions landing at the end of an 18-month transition in May 2027, so the sensible reading is that you have a window to fix your consent language rather than a reason to wait. If you hire in the EU, the UK or the United States, or supply staff to companies that do, their rules will reach you through contract long before Indian law changes.
Does AI reduce bias in hiring or increase it?
Both, depending on where you put it. It reduces bias where it removes information a human should not see, which is why resume anonymisation before review is one of the better uses available. It increases bias where it learns from your past hiring, because your past hiring is what it will reproduce. The deciding factor is not the tool, it is whether you measure stage-wise selection rates by group after deployment. Teams that measure find problems and fix them. Teams that do not find out from a candidate.
Which recruitment tasks give the fastest return from AI?
Interview scheduling and coordination, then resume parsing and deduplication, then rediscovery of past applicants in your own database. All three are volume tasks with no selection decision in them, they show a measurable result inside a quarter, and none of them carries adverse impact risk. Screening and assessment produce bigger headline savings and take much longer to do safely.
Do we have to tell candidates we are using AI?
In New York City, yes, with at least ten business days' notice before use of an automated employment decision tool, plus a published bias audit. In Illinois, yes, for AI-analysed video interviews. In the EU, transparency duties attach to recruitment systems under the AI Act and to automated decisions under the GDPR. In India there is no specific disclosure requirement, though your DPDP notice has to describe the purpose for which you are processing candidate data. Disclosing anyway is the sensible default, because it costs a paragraph and it is the one thing candidates consistently say they want.
Are AI video interviews with facial or emotion analysis allowed?
Not in the European Union. Emotion recognition in the workplace is a prohibited practice under Article 5(1)(f) of the EU AI Act and has been since February 2025, recruitment sits inside that scope, and the July 2026 amendments that pushed the Act's high-risk deadlines out to December 2027 left the prohibition in place. Illinois requires consent, explanation and deletion rights for AI-analysed video interviews. Beyond the legal position, scoring candidates on facial expression or vocal delivery penalises accents, speech differences and several disabilities, on evidence that was thin to begin with. Structured work samples with a written rubric give you better information with none of the exposure.
How do we handle candidates who use AI to write applications and assignments?
Assume they do, and stop treating polish as a signal. Cover letters and take-home assignments have lost most of their information value. The workable replacements are live problem-solving, or keeping the take-home and spending twenty minutes asking the candidate to explain, extend and debug their own submission, which someone who did the work can do immediately. Do not act on the output of an AI-detection tool. Accuracy on text is poor and the cost of a false accusation is far higher than the cost of a weak signal.
How long can we keep resumes for AI-driven candidate rediscovery?
For as long as the purpose you told the candidate about lasts, and no longer. Under the DPDP Act, personal data has to be erased when the purpose is served, so a database you mine for future roles has to have been collected on a notice that says you will consider applicants for future openings, with a retention period stated and a way to withdraw. Most application forms do not say this, which makes a large part of the rediscovery use case fragile. Fix the notice at the point of application and the problem goes away going forward.
Where This Leaves You
The version of this that works is unglamorous and it is available now. Automate the coordination completely. Use AI to widen the top of the funnel and to strip irrelevant information out of applications. Let it order the review queue. Keep a person on every rejection, with a reason you could read out loud. Measure what you threw away, not only what you kept.
So work in that order. Start with scheduling and interview logistics, because it pays back inside a quarter and risks nothing. Add parsing, deduplication and rediscovery next, once you have fixed the consent language on your application form. Bring in ranking after that, as a queue order rather than a cutoff, with reason codes and a stage-wise selection-rate report from day one. Treat assessment scoring as the last thing you adopt and the first thing you review, and keep facial and emotion analysis out of it entirely. Then run the rejected-resume audit every month, because it is the only part of this that tells you the truth.
The rest of the lifecycle takes over from here. What happens to the candidate record after acceptance sits in our guide to the employee lifecycle in HR software, and the same technology applied to the payroll side is covered in AI in payroll. Smaller teams running hiring without a dedicated recruitment function will find the practical version in our guide for small businesses.
If you would rather your sourcing, screening, interview scheduling and offer generation ran out of the same system that already holds your employee data, that is what our recruitment software is built to do. Book a free demo and bring one open role with its current funnel numbers. It is a much faster conversation when you can see where your own time is actually going.
