The AI Recruitment Playbook Everyone Is Copying Is Backwards

In some of the high-volume BPO campaigns we have seen, the class was full on paper. Offers signed. Start dates confirmed. Then Monday arrived, and close to a third of the expected hires did not show up.

No message. Nothing.

If you lead high-volume recruitment, you know that feeling. But what looks like a ghosting problem is often something deeper: a hiring process that never established a reliable chain of proof. And a broken chain does not only create empty training seats. It can leave you unsure whether the person you assessed, interviewed, and hired was ever the same person at all.

Here is what that looks like at the extreme. Last year, the security company Pindrop was hiring a software engineer. Mid-interview, Pindrop says, its recruiter got an alert: the candidate on screen was flagged as a deepfake. The position "he" was applying for? A seat on Pindrop's deepfake detection team. When the company went on to analyze 300 applicant profiles from one job posting, it reports finding more than 100 fabricated identities, many built with AI-generated resumes and manipulated credentials.

That sounds like an outlier. It is not. In July 2025, the US Department of Justice sentenced an Arizona woman who ran a "laptop farm": company laptops shipped to her address, then operated remotely by North Korean workers using stolen American identities. That one scheme touched 309 US companies and generated over 17 million dollars. Gartner now predicts that by 2028, one in four candidate profiles worldwide will be fake.

Beneath those extreme cases sits a much larger gray area that remote hiring has made difficult to police. Gartner found that 39 percent of candidates used AI during the application process. Among them, 29 percent used it to generate assessment answers, while six percent of all candidates admitted to impersonation or having someone else interview for them. Polishing a resume with AI is not dishonesty; the line is crossed when candidates break stated assessment rules or misrepresent who did the work. And that line is being crossed at scale. Set that against a thousand-application pipeline. Recruiters describe watching a candidate's eyes track text scrolling on a second screen. A more qualified friend takes the online assessment. Interview-assistant tools listen to your questions and feed back polished answers in real time.

The market has responded. Google, Cisco, and McKinsey have reintroduced in-person interviews for parts of their hiring processes, specifically to counter AI-enabled cheating, and in a Gartner survey, 72.4 percent of recruiting leaders said they now conduct in-person interviews to combat fraud.

The irony is hard to miss. We operate seven physical recruitment hubs across the Philippines. For years, that looked old-fashioned. Now some of the world's most sophisticated employers are restoring a version of the checkpoint we never removed: a real person verifying another real person.

So this is a working guide. If you lead talent acquisition for high-volume hiring, especially in BPO and shared services, here is how I think about the problem, the honest trade-offs of your options, and the lessons we have learned recently, some of them the expensive way.

The problem is not ghosting. Ghosting is a symptom.

If you run high-volume hiring, two gaps keep you up at night.

The first is the one this article opened with: the gap between "offers accepted" and "bodies in the training room on day one." The fact that experienced recruitment teams are no longer surprised by heavy day-one fallout tells you how serious the problem has become.

The second is the gap between what the candidate demonstrated in the process and what the candidate can actually do. The assessment score was excellent. The interview was fluent. Then week two of training reveals a person who cannot reproduce any of it.

Before I diagnose the candidates, let me be fair to them. Ghosting runs in both directions: CareerPlug found that while 44 percent of candidates admit ghosting an employer, 53 percent say they have also been ghosted by a prospective employer. Employers helped normalize the same behavior they now complain about.

Now, the diagnosis. It is tempting to say both gaps come down to weak candidate intent, but that is not quite right. A fraudster can be extremely intent on getting hired. What ghosting, proxy assessments, and deepfakes actually have in common is a broken chain of proof. A hiring process has to establish three separate things: identity (this is a real person, and the same person at every stage), capability (the candidate can perform to the required standard under the assessment conditions you set), and intent (the candidate understands the actual conditions and is presently prepared to report on day one and stay). Ghosting is a failure to establish intent. Prohibited or undisclosed AI assistance and proxy test-taking are failures to establish capability. Deepfakes and stand-ins are a failure to establish identity. A fully remote, frictionless funnel can fail all three without ever noticing, because it never asks the candidate to prove anything that costs effort. An application is interest, not intent; a confirmed interview slot is still a free click. When every signal is cheap, committed candidates become harder to distinguish from mass applications, backup options, and deliberate fraud.

Your three real options, honestly weighed

To be precise about the choice: where the process happens and who exercises judgment are two different decisions. A remote process can use live human interviews and identity technology; an onsite process can run on AI behind the scenes. In practice, though, high-volume operations tend to land on one of three configurations.

Remote-first hiring with digital verification

Sourcing, screening, assessments, and interviews all happen online, with AI carrying much of the load, and fraud is managed with technology: liveness detection, ID document matching, proctoring software.

The real advantage. Reach, speed, and accessibility. Tens of thousands of applications with a small team, low cost per candidate, and a fair path for people in far provinces or with mobility and caregiving constraints. Anyone who tells you this model has no merits is selling you something.

The real trade-off. You are betting your funnel on detection technology winning an arms race against generation technology, and every fraud case in this article's opening exploited a remote process. Contextual signals are weak: a form cannot ask why a candidate hesitated. And if AI becomes the visible judge, many candidates leave. Greenhouse surveyed 2,950 active job seekers: 38 percent had already withdrawn from a process because it included an AI interview, and another 12 percent said they would. And those numbers only count the withdrawals someone measured. Most screening stacks, especially third-party tools, rarely report midway abandonment, so funnel data contains only the candidates who endured the process. Conclusions drawn from that survivors-only sample, such as "the ones we lost lacked intent," are built on precisely the people the process did not drive away; the lost candidates' intent was never tested, it was exhausted by front-loaded forms and assessments. Pew found 66 percent of US adults would not want to apply where AI helps make the decision, and Gartner found only 26 percent of candidates trust AI to evaluate them fairly. A 2025 peer-reviewed experiment found that AI-led interviews reduced application intention through perceived unfairness; the effect appeared in the study's publishing-industry scenario but was not statistically significant in its high-tech scenario, which suggests industry context matters. BPO was not tested, but in our own candidate conversations the pattern is consistent: in a market where the interview has always been a relationship moment, an AI-only interview feels impersonal. Candidates describe feeling like part of the herd, a data point being collected rather than a person being considered.

Centralized onsite recruitment

Every candidate travels to your site. Every interview is face to face, every assessment proctored in a room.

The real advantage. The strongest verification of all three signals at once, plus real human interaction: questions answered live, the job previewed honestly, confusion caught in the moment.

The real trade-off. Travel burden, infrastructure cost, and limited capacity. Reach shrinks to candidates able to make the trip, and at the volumes a large BPO account demands, hundreds of agents in weeks, a single-site manual process cannot keep pace. Two fairness notes we hold ourselves to: attendance is a behavioral signal, not a measure of talent or character, and a candidate blocked by disability or distance deserves an alternative verified route, not a rejection. And the in-person requirement fits high-volume roles; for scarce, specialized talent, an early travel hurdle loses candidates you cannot afford to lose.

Distributed hybrid: nationwide reach, one verified human checkpoint

This is the model we run, refined over 15 years and more than 50,000 placements. Use AI to remove administrative work. Keep consequential judgment and relationship moments human. Then place one in-person checkpoint where identity, capability, and intent can be verified together.

Sourcing is virtual and nationwide; we process around 700,000 candidates a year, and no set of buildings could be the boundary of that reach. AI supports parsing, matching, and flagging duplicate profiles and identity anomalies for human review, while automation handles scheduling, reminders, and data hygiene. Then, before endorsement, the candidate visits the nearest of our seven hubs: Baliuag in Bulacan, Muñoz and North EDSA in Quezon City, Malate in Manila, Ortigas, Cebu, and Davao. The point of a network is that the trip stays reasonable: for most candidates the nearest hub is a town or a province away, and coverage lines up with where our clients operate their own sites nationwide. We source, screen, verify, and endorse; the client keeps the final hiring decision, made about a person we have actually met.

The visit sits after basic eligibility and before the client spends scarce interview slots, exactly where a false candidate does the most damage. And it has to earn the candidate's trip, so it consolidates the steps otherwise scattered across weeks: screening, a realistic preview of the actual account and shift, a proctored assessment, document checks matched against the person holding them, and an honest conversation about the commute, the pay, and the start date. The candidate leaves with fewer unknowns, a named recruiter, and a clear next step. An onsite visit that just recreates five online forms on office computers adds cost without adding signal.

I used to describe this checkpoint as an intent filter, but the better description is reciprocal commitment. The candidate gives us effort and honesty; we give the candidate the real job, the real site, the real schedule, and a fast decision. Both sides leave with better information. That may be why Gartner found 62 percent of candidates more likely to apply when an organization required in-person interviews: for many candidates, a visible verification standard signals that the employer is running a fair process, one where they are not competing against impersonators and hidden assistance.

The results are worth stating carefully. On a recent 600-hire ramp completed in 30 days, 94 percent of those hires reported on day one, in an industry where substantial day-one fallout has become accepted as normal. Plenty of vendors have good sourcing and good AI. The verified checkpoint is the part of our system we would be least willing to remove, and it is a large part of why 8 of the top 20 global BPOs run ramps through this network today.

The real trade-off. Physical coverage, trained hub recruiters, and excellent handoffs. And here is the honest math on the first two: a hub network only pays for itself when it runs at volume every week of the year. A single employer's hiring comes in waves, a 400-agent ramp this quarter, near silence the next, so a checkpoint built for your own use sits idle most of the time while its cost does not. That is why almost nobody builds one, and why the economics change completely when the infrastructure is shared: our hubs stay busy because they serve many clients' ramps at once, which means each client gets a verified, in-person checkpoint in seven cities without leasing a single square meter or hiring a single hub recruiter. What stays on the client's side either way is handoff discipline: the model fails when the sourcing team, hub recruiter, account manager, and client each hold a different version of the candidate's story. Which brings me to the thing I see go wrong most often.

Before you decide: four things most TA leaders underweight

First: diagnose the exact failure before you buy the cure.

"Ghosting" is not one problem. Is your biggest loss between application and first contact? Confirmed interview and attendance? Signed offer and day one? Is the cause weak intent, employer delay, role mismatch, or identity fraud? Each needs a different control, and a company that responds to fraud by piling slow, unexplained checks onto every candidate will reduce fraud and worsen ghosting at the same time. Map the leak first. Then fix that leak.

Second: candidates are not rejecting AI. They are rejecting AI where judgment should live.

Read the Greenhouse findings closely: what drives candidates out is an algorithm scoring a pre-recorded video with no human present, undisclosed AI use, and AI-led interviews with no alternative. Candidates are fine with the machine confirming their schedule at 9pm. They are not fine with the machine deciding their future. Design your automation around that line and you keep the efficiency without paying the withdrawal tax. Disclosure cuts both ways: state clearly what AI use is acceptable on the candidate's side before the test, not silently after it.

Third: the AI failure pattern I see most often is ignored shadow workflows.

Every recruitment operation runs two workflows. There is the official one, the ATS trail: applied, screened, assessed, endorsed, offered, hired. And there is the shadow one: the recruiter who called the candidate the night before to warm her up, learned she is worried about the night-shift commute, and reassured her about the shuttle. The recruiter who knows a candidate only answers after 6pm because he currently works nights. The recruiter who found out in a side conversation that a top-scoring candidate believed the role would turn remote after training and cannot sustain the onsite schedule.

Now feed only the official workflow into your AI scoring engine. The night-shift candidate "ignored" a reminder for 24 hours, so the model scores him a risk, when he has already visited the hub and arranged his day-one transportation. The misinformed candidate looks perfect on paper, so the model ranks her first, when she has already mentally left. One false negative, one false positive, both from a model working exactly as designed. The model is not broken. It is blind. It makes wrong decisions because the system hands it an incomplete version of every candidate, and garbage input produces garbage output. If your AI ingests the ATS and ignores the shadow workflow, you have automated your least informed view of every candidate and given it authority over your most informed one. Then, when the predictions fail, leadership concludes AI does not work for recruitment. Before blaming the model, inspect the version of reality it was allowed to see. In many failed implementations, the model was starved of context the recruiters already had.

One caution before you fix it: the answer is not to pour every chat thread and recruiter hunch into a model. That creates privacy and bias problems of its own. Capture the job-relevant facts in structured, auditable form: confirmed shift availability, preferred contact time, commute feasibility, clarified work-location expectations, outstanding documents, concerns the candidate stated in their own words. Leave tone judgments and private conversations out. The principle is not more data in; it is relevant, consented, structured context in, with a documented recruiter override, judged against day-one show-up and early retention. The first AI project in recruitment is usually a process-discovery project in disguise.

Fourth: intent is not information you collect. It is behavior you observe, and commitment you build.

Any form can ask whether a candidate can handle a 10pm shift, and almost every candidate ticks yes. A recruiter on the phone does better: the pause before the answer invites the real obstacle, which is usually specific and solvable, a sibling to bring to school in the morning, a parent who will not agree to night work, a transport cost nobody has done the arithmetic on, a competing offer that has not been mentioned. Make those calls. They are cheap and they work.

But understand what a call can and cannot establish. A call collects what a candidate says. A visit collects what a candidate does, and the two diverge exactly where it matters. Saying yes on a call costs a candidate nothing, and neither does disappointing a voice they have never met. Traveling to a hub costs fare, a morning, and often a shift somewhere else. A candidate weighing four prospects will verbally accept all four and visit one or two. You do not have to ask where you rank; attendance answers it. That is revealed preference, and no reminder cadence produces it.

The visit also creates commitment that did not exist beforehand. The candidate has now spent something, so leaving means losing it. They have a named recruiter, a room they sat in, answers to the questions they actually had, and a concrete picture of the account. Abandoning a voice on the phone is frictionless; abandoning a person you spent an hour with is not. And the trip is a rehearsal of the trip they must make on day one, at which point the commute has stopped being hypothetical.

Even so, intent has a shelf life. Committed on Monday can become a better offer on Thursday, so confirm again before day one, with a real conversation rather than another templated reminder. Three minutes of a recruiter's time against the cost of an empty training seat.

What we have learned, the short version

Presence turned out to be a technology, and the building itself does nothing. The hubs looked like legacy overhead until the fraud wave hit; the value was never the walls but the recruiter in the room, the honesty of the job preview, and the speed of the next step. The rule that follows: AI should support the recruiter on stage, not replace them, and the funnel should get more human as it narrows, not less. When we pushed automation into candidate-facing judgment moments, experience and conversion suffered. When we pointed it at operational work, throughput improved and recruiters got hours back for the conversations that actually move candidates. Too many AI implementations run this backwards, automating the candidate's side of the process for the team's convenience while the judgment moments go under-informed.

We also learned to measure intent, not just fit, and to make honesty safe. Fit predicts whether someone can do the job; intent predicts whether they will show up to do it. You only get honest intent signals when a candidate can say "this schedule will not work for me" without being punished for it. An honest no today beats a polite yes followed by silence on day one, because the candidate who trusted you with a no is one you can place somewhere better tomorrow.

And we learned which scoreboard to trust. Pipeline size, application counts, and time-to-endorse can all be inflated by exactly the frictionless processes that produce empty training rooms. Show-up rate and early retention are substantially harder to inflate, which is why they are where we would point any TA leader first.

Remote hiring does not need to disappear, and no one should want it to. But somewhere between application and endorsement, identity, capability, and intent have to be verified together, by someone in a position to actually see the person. For high-volume hiring, we have found that one well-designed human checkpoint surfaces context that automated steps routinely miss. Not because a building is magic. Because that is the moment real context finally enters the process.

And whoever runs that checkpoint for you, your own team or a partner, hold it to the standard this article laid out. It sits after eligibility and before endorsement. It consolidates the process into one visit that earns the candidate's trip. It asks the candidate to do something, not just say something, so intent is observed rather than reported. It verifies documents against the person holding them. It captures structured context that travels with the candidate instead of dying in a chat thread. It re-confirms intent before day one. And it is judged on the two numbers that are hardest to inflate: how many verified people are sitting in seats on Monday morning, and how many are still there at 90 days. Anything less is an office with extra steps.