Diagogue from a real job interview:

And this is exactly the kind of moment that makes me question what we are actually trying to assess in interviews.

The candidate did not say she had no strengths. She did not say she was bad at her job. She simply did not describe herself as a “star”, and within a few seconds that somehow became a judgment about how highly she thought of herself.

That jump is exactly where interviewing stops being assessment and turns into interpretation.

And the problem is much bigger than one strange question.

We very often confuse a strong candidate with a strong employee, even though these are not the same thing at all.

There is a certain type of person who tends to look great in interviews.
They answer quickly, tell polished stories, know how to frame their achievements, speak confidently about their strengths and usually understand quite well what the interviewer wants to hear. They know where to say “I led”, “I built”, “I transformed”, “I delivered”, and they are good at presenting their experience in a way that makes it easy for the interviewer to see value.

There is absolutely nothing wrong with that.

The problem starts when we treat the ability to present yourself well as evidence that you will also perform well in the job.

Because interviewing is a skill of its own.

And if your interview process heavily rewards interview skill, then you will naturally select people who are good at interviews.
This sounds painfully obvious when written down, but companies do it all the time.

Research on impression management has been showing this for years. Self-promotion affects interview ratings, and its impact is stronger in interviews than in actual job-performance ratings. A meta-analysis covering 8,635 participants found exactly that pattern. Another meta-analysis concluded that self-presentation tactics influence interviewer evaluations, especially in less structured interviews.

So when someone leaves an interview and the hiring team says, “She was amazing,” I always want to ask one more question.

Amazing at what exactly?

At doing the job, or at interviewing for the job?

What are we actually measuring?

Imagine two candidates.

Candidate A says:

My strongest quality is strategic thinking. I’m exceptionally good at seeing the bigger picture, influencing stakeholders and turning ambiguity into action.

Candidate B says:

I’m not sure I’d call it exceptional, but I’m usually good at identifying problems early. For example, in my last project…

Candidate A will probably create a stronger first impression.

But based on those two answers alone, we still have no idea who is actually better at strategic thinking or identifying problems.

And yet a weak interview process can already start scoring Candidate A higher, simply because confidence feels like competence, fluency feels like expertise and a polished answer feels like stronger evidence than it actually is.

That is one of the most common mistakes in interviewing.

The interviewer’s job is not to decide whether a candidate sounds like someone who is good at something. The interviewer’s job is to collect enough evidence to understand whether this person has demonstrated the competencies required for this role.

Those are two very different things.

When “culture fit” style interviewing becomes dangerous

The “You don’t think very highly of yourself?” comment bothered me for another reason.

There is no reasonable path from “I wouldn’t describe my qualities as star qualities” to “this person has low self-esteem”.

Maybe she is modest. Maybe she dislikes exaggerated language. Maybe she understands the word “star” differently. Maybe she comes from a culture where openly praising yourself is less socially acceptable. Maybe she simply prefers to talk about what she has done instead of attaching big adjectives to herself.

Research on self-presentation in interviews also shows that people differ in how comfortable they are with self-promotion, and those differences can be cultural as well as individual.

So when an interviewer rewards someone for confidently saying, “Yes, I’m exceptional,” but becomes suspicious when another candidate says, “I’m not sure I would describe myself that way,” they may think they are assessing confidence.

In reality, they may simply be assessing comfort with self-promotion.

And unless self-promotion itself is a requirement of the role, that is not the same thing as assessing job competence.

The interviewer’s job is not to read minds

This is another thing we do surprisingly often in hiring:

• A candidate hesitates, and we decide they lack confidence.

• They give a short answer, and we decide they are not motivated.

• They do not smile enough, and we decide they are not excited about the role.

• They say “we” instead of “I”, and suddenly we suspect they did not personally contribute much.

• They do not call themselves a star, and apparently now they do not think very highly about themselves.

At some point we stop collecting evidence and start writing our own story about the candidate.

The most dangerous part is that these interpretations rarely sound obviously biased in a hiring debrief. They usually sound quite professional.

• “I just didn’t feel enough seniority.”

• “Something was missing.”

• “I’m not sure she has the presence.”

• “He didn’t seem hungry enough.”

• “I didn’t feel leadership energy.”

This is why bias in interviewing is so difficult to catch. It rarely arrives looking like bias. Very often it arrives disguised as intuition, experience or “just a feeling”.

We form an impression first and then try to explain it using business language.

Charisma shouldn’t decide the hire

This is why I keep coming back to structure.

A structured interview does not mean turning the conversation into a robotic questionnaire where the interviewer is not allowed to react or ask follow-up questions.

It means deciding before the interview what exactly you are assessing, what kind of evidence would demonstrate that competency, which questions can actually produce that evidence, and what strong, acceptable and weak answers look like.

It means reducing the amount of space available for random interpretation.

Research on personnel selection consistently places structured interviews among the stronger methods for predicting future job performance.

And this is also why “What is your biggest strength?” usually gives me much less useful information than something like:

“Tell me about a situation where you had to solve X. What was happening, what did you do, why did you choose that approach and what changed as a result?”

Now I have something I can actually assess.

I have behaviour, decisions, context, complexity and outcomes.

Not just an adjective the candidate chose for themselves.

Your quietest candidate might be your strongest employee

And your most confident candidate might genuinely become your strongest employee too.

The point here is not that confidence is bad, or that charismatic candidates should somehow be treated with suspicion. That would just replace one bias with another.

The point is that confidence is not evidence of competence unless confidence itself is relevant to the role and is being deliberately assessed.

Some roles do require a high level of persuasion, social confidence or executive presence. In those cases, those things absolutely matter.

But even then, they should be assessed because they are part of the job, not because the interviewer personally enjoyed the candidate more.

Hiring needs to distinguish between what helps someone perform well during selection and what helps someone perform well after selection.

Sometimes those things overlap.

Sometimes they really do not.

And if your interview process cannot tell the difference, you will keep hiring people who are exceptionally good at getting hired.

So the next time someone feels like a “star candidate”, ask why

• Was it because their examples were strong?

• Because the problems they solved were genuinely complex?

• Because their decisions were good?

Because they showed learning ability, judgment, skill and results against criteria you had defined before meeting them?

Then yes, maybe you really did find a star.

But if the main reasons are that they were charismatic, polished, confident, easy to talk to, knew how to sell themselves and somehow “felt senior”, then I would be much more careful.

Because you may not have found your strongest future employee.

You may have simply found the person who is best at being a candidate.

Whenever we talk about recruiting efficiency, we very often talk about volume: how many candidates we found, how many we contacted, how many entered the process, how many screening calls we ran, how many made it to the hiring manager interview.

And somehow, the more people there are, the more productive the team seems.

But I keep coming back to the same question: where did the actual recruiting go in all of this?

Because for one role, we still hire one person. Even if the pipeline had 100, 500, or 1,000 candidates in it.

A large pipeline on its own is not proof of efficiency. Sometimes it can mean exactly the opposite. Maybe we do not understand the profile well enough. Maybe sourcing is too broad. Maybe we disqualify irrelevant candidates too late. Maybe the team is simply bad at making decisions.

If a recruiter needs 50 screening calls and has to show 20 candidates to a hiring manager to close one role, my first question would definitely not be how to give that recruiter more candidates or more AI screening tools.

I would want to know why they need that many candidates in the first place.

Maybe the kickoff was weak and the team never properly aligned on what they were looking for. Maybe the sourcing strategy is basically “message everyone who looks close enough.” Maybe the assessment process produces interviews but still leaves the team unsure who is actually strong. Maybe the hiring manager changes the requirements after every conversation. Or maybe the team is simply afraid to make a decision and keeps looking because the next candidate might somehow be better.

And this is where I become much more interested not in the absolute size of the pipeline, but in how efficiently that pipeline actually works.

Not just how many people entered it, but how many were relevant, where we lose them, how many interviews we need for one hire, how many candidates have to move through the system before we can make a decision, and how quickly we understand that someone is not the right fit.

If we want to talk about recruiting efficiency, I would rather look at conversion rates, interview-to-hire ratio, candidates per hire, time to disqualify, and stage time.

Those tell me much more about whether recruiting is actually working.

Conversion rates

If 500 candidates enter the funnel but only 20 make it to the first genuinely relevant stage, that does not automatically mean sourcing is strong.

Maybe we just created a huge amount of work for ourselves.

Conversion between stages tells us far more than the absolute number of candidates because it shows us where the pipeline stops working.

If almost everyone drops after screening, I would look at sourcing and the profile.

If the recruiter sends a lot of candidates to the hiring manager and most get rejected, I would look at calibration between the recruiter and the manager.

If many candidates reach the final stages but the team consistently needs ten finalists before making an offer, I would start looking at assessment quality and decision making.

So for me, a good funnel is not one with a lot of people in it. A good funnel is one with conversion rates that make sense, where we understand why people move forward and why they do not.

Interview-to-hire ratio

Another metric I really like in this context is very simple:

How many interviews do we need to make one hire?

You can beautifully optimize scheduling, automate interview notes with AI, and add another interviewer copilot. But if the team still needs 30 interviews to make one hire, I am not convinced the main problem was the lack of automation.

A high number of interviews can mean we are bringing the wrong people into the process. It can mean the assessment criteria are vague. Or it can mean the interviews themselves are not giving the team enough useful information to make a decision.

So efficiency here is not about running more interviews faster.

It is about running enough interviews to make a good decision without dragging dozens of people through a process they should never have entered in the first place.

Candidates per hire

This is probably the simplest way to flip the whole pipeline conversation.

Instead of asking: “How many candidates did we attract?”

ask: “How many candidates did we need to make one good hire?”

Of course, this is not a universal benchmark. High-volume hiring, a niche senior role, and a junior position will all look completely different.

But within similar roles and the same business context, this metric can tell you a lot.

If one hire used to require 70 candidates, and after a better kickoff, sharper sourcing, and stronger calibration with the hiring manager it now takes 25, that is a much more interesting efficiency signal to me than “we increased pipeline by 40%.”

We did not simply do less work. We removed work that was creating no value.

Time to disqualify

There is another metric I think gets far too little attention:

How quickly do we understand that a candidate is not right for the role?

Efficient recruiting is not only about finding the right person quickly. It is also about being able to stop the process once you already have enough information to make a no decision.

If a candidate who clearly does not meet the key criteria reaches the third interview, I would not call that a pipeline problem. That is a process design problem, an assessment criteria problem, or a problem with how consistently the team is using those criteria.

And the later we figure it out, the more expensive that mistake becomes. More recruiter time, more hiring manager time, more interviewer time, more calendar slots, more candidate time, and more work that no longer improves the final decision.

That is why I would not only look at Time to Hire, but also at Time to Disqualify and stage time. They tell us how quickly the process gathers enough information to make a decision, and whether we are moving candidates forward simply because we failed to assess them properly earlier.

For me, that is also part of recruiting efficiency: not just moving candidates faster, but understanding earlier who actually deserves to move forward.

And this is where I get most frustrated with the current conversation about productivity in recruiting.

We now have a huge number of technologies that allow us to work with much larger pipelines. AI sourcing can find more people. Automated outreach can send personalized messages to more candidates. AI screening can process more applications faster.

But being able to process more candidates does not mean we actually need more candidates.

And I think this is exactly where we have started confusing automation with efficiency.

Maybe real recruiting efficiency is not about pushing as many people as possible through the system.

Maybe it is about defining the profile more precisely, understanding the talent market better, finding a smaller and more relevant pool, identifying poor matches earlier, assessing the remaining candidates better, and helping the business make a decision without the endless “let’s see a few more.”

So I increasingly think the more useful conversation about recruiting efficiency is not about how to make the pipeline bigger.

It is about how small we can make it without sacrificing quality.

Because if two recruiting systems produce an equally good result, but one needs 100 candidates to get there and the other needs 30, I know which one I would call more efficient.

The goal of recruiting is not to build a large pipeline. The goal is to make a good hire with as little unnecessary work as possible for everyone involved.

Before that, things were relatively simple. You did more, moved faster, delivered better work, closed things before the deadline, and you had a pretty clear sense that the day went well.

There was something tangible at the end of it.

A role was closed. An interview was done. A candidate list was reviewed. An offer was sent. A task moved to Done.

Then you become a lead, and suddenly you can spend half a day in meetings, help the team make a few important decisions, stop one bad idea before it turns into a real problem, give the business context it was missing — and still sit there in the evening thinking:

What did I actually do today?

I think this is one of the least obvious parts of moving into leadership. The familiar definition of “done” disappears, and with it, a very familiar way of measuring your own value.

The first thing leadership takes away from you is the feeling of “done”

Before leadership, a lot of work is measurable in a very direct way. The vacancy is closed, the interview is finished, the shortlist is ready, the task is complete. Even when the workload is ridiculous, at least you can usually see where one thing ends and another begins.

Leadership work is different.

You cannot really “finish” a team. You do not complete stakeholder management. You do not close the task called “build good processes” and never touch it again. Even something that works perfectly well today may need to change in three months because the business changed, the team changed, hiring volume changed, or the context around you changed.

And when the external measure disappears, your brain tends to invent a new one.

Instead of “the result is done”, the question becomes:

Am I trying hard enough?

The problem is that this question has no finish line.

You can always try a little harder.

So the first strategy is usually obvious: work more

When I first became a Talent Lead, my team was… me.

On the one hand, very convenient. Nobody argued with the lead.

On the other hand, I had absolutely no idea what was actually supposed to change in my work.

More was expected from me now. But more of what?

Nobody handed me a list saying: here are the five things that now make you a good Talent Lead.

So I made what felt like the most logical conclusion: I simply needed to work more.

Close more roles. Reply faster. Control more. Initiate more. Be more involved. Do more of everything.

And honestly, that reaction makes complete sense. If the way you used to work is what got you promoted, the natural assumption is that the next level means doing roughly the same thing, just better.

The problem is that you cannot really do a new role with the old operating system.

I definitely did not understand that at the time.

Then the team appears, and things get even stranger

You would think that once I actually had a team, things would become easier.

They did not.

While I was leading a team of one, I could at least prove my usefulness through the amount of work I personally produced. Once other people joined, that logic stopped working completely.

• I was no longer the person closing every role.

• I was not sourcing every candidate.

• I was not running every interview.

• I was not making every decision myself.

And somewhere around that point, a much more uncomfortable question appeared:

So what exactly is my value now?

• If the team can work without me constantly being involved, does that mean I am doing a good job, or that I am becoming unnecessary?

• If I do not check every detail, is that delegation or am I simply not controlling enough?

• If the team delivered the result, can I even consider that my result?

• And what does a good day for a lead look like if there are not ten finished tasks at the end of it?

It took me much longer than I would like to admit to figure this out. There was burnout, one pretty significant fuck-up, therapy, and a sabbatical somewhere in the process.

The answer ended up being very simple to say and much harder to accept:

A lead is not the best recruiter on the team.

And not the person who can jump into any vacancy and do it faster than everyone else.

The job is to build a system where the team can produce a good, consistent result without the lead having to hold every part of that system together manually.

That shift — from “I do my job well” to “I create the conditions in which other people can do their jobs well” — was probably one of the hardest professional transitions I have made.

Responsibility is very easy to confuse with “I have to carry everything”

I now see the same pattern from the other side, in consulting and with people who join Talent Lead Expert during their first months in a new leadership role.

One student recently described it in a way that immediately felt familiar:

I feel like I am carrying all the responsibility because I don’t even know if I am allowed to require information from other people.

And I understand that logic very well, especially when you are new.

You do not want to be difficult. You do not want to keep bothering the hiring manager with questions. You do not want to go to the CEO again. You do not want to be the new person who arrived yesterday and is already asking everyone for something.

So the easier option is to think: fine, I will figure it out myself.

Then, almost without noticing, you become the person collecting all the information, making all the decisions, following up with everyone, filling every gap and compensating for everything that is missing.

And at some point, it becomes difficult to tell where your responsibility ends and somebody else’s begins.

There is an important difference between “I am asking for help because I cannot handle this myself” and “I am coordinating a process where different people own different parts of the work.”

If a hiring manager owns the role requirements, answering your questions about those requirements is not a favour.

If the business needs to give you context about priorities, you are not “bothering them” by asking for it.

If a decision belongs to someone else, part of your job as a lead is to get that decision — not quietly make it on their behalf because it feels easier.

Leadership without the ability to ask, align, escalate, request decisions and return ownership to the right person simply does not work.

Otherwise you end up with a system where many people are technically responsible, while one person is actually carrying the whole thing.

And eventually, the workday simply stops ending

If your definition of a good day becomes “I did everything I possibly could”, then your workday cannot really end.

There is always one more candidate you could reply to, one more vacancy you could review, one more conversation you could have, one more document you could improve, one more process you could fix, one more problem you could prevent.

This is where it becomes very easy to mistake boundaries for lack of commitment.

If you close your laptop at 7pm and there are still messages in Slack, maybe you did not do enough.

If you do not reply in the evening, maybe you are not responsible enough.

If you do not pick up one more task, maybe you are not trying hard enough.

But a workday that does not last until 10pm is not evidence of poor performance. It is evidence that work has boundaries.

And sometimes a new lead literally has to learn to stop working not when “everything is done”, but when the time allocated for work is over, because “everything” in this role will never be done.

The same applies to the first weeks in a new company.

If nobody gives you proper onboarding, it is very easy to compensate by doubling your workload: launch something quickly, prove value immediately, redesign a process, bring a solution, show everyone that hiring you was a good decision.

But sometimes the most useful thing a new lead can do in the first few weeks is not implement anything.

It is to understand where they actually landed.

Who really makes decisions. Where things hurt. Which agreements only exist verbally. What the business actually expects. What the team has already tried. What is working. What should absolutely not be “fixed” just because it looks unfamiliar.

That is work too.

It just does not give you the same satisfying little hit as moving a task to Done.

So how do you know if you are actually doing a good job?

For me, the answer slowly moved away from “How much did I personally do today?” and toward very different questions.

• Is it easier for the team to make decisions without me?

• Do people understand who owns what?

• Are we seeing problems before they turn into complete chaos?

• Can I explain to the business what is happening and why?

• Does the team have enough context to move without constantly asking, “What do I do here?”

• And maybe the most important one: can the system function for a while without my constant manual control?

This is probably also why we talk so much in Talent Lead Expert not about becoming “more productive” as a lead, but about ownership, decision rights, stakeholder management, metrics, role boundaries and building systems.

Not because leaders need to surround themselves with frameworks and pretty diagrams, but because without external reference points, it is very easy to turn your own anxiety into a management system.

It took me a long time to stop evaluating myself as a lead by the number of things I personally did, and I think this is one of the reasons the transition into leadership feels so strange.

You are often promoted because you are very good at doing the work, and then the new job gradually asks you to stop being the person who does the most of it.

Not to work less. Not to care less. Not to step away from the team.

But to stop being the system itself.

A good result for a lead is not when everything stops without them. It is when things continue to work.

Because a system that only functions because one person remembers everything, pushes everyone, rescues every problem and stays online until midnight is not strong leadership.

It is just very well-organized future burnout.

None of it is malicious. All of it is outdated.

And if you’re a founder, you’re not just hiring the person who survived this playbook. You’re paying for every shortcut baked into it.

Here’s where the cost actually shows up.

1. “Sell the role, don’t mention the risks”

The candidate finds out about the risks in week one. Except now it’s not a conversation, it’s a resignation.

Replacing a mid-level hire costs 6-9 months of their salary, according to most industry benchmarks. That number doesn’t include the three months your team spent onboarding someone who leaves before they’re productive.

What you’re actually paying for: a recruiter who optimized for saying yes, not for fit.

2. “Candidates who negotiate are difficult”

A candidate who states their number and defends it is behaving rationally. A candidate who doesn’t is often the same person who won’t defend your budget in a vendor negotiation eighteen months from now.

If your recruiter treats negotiation as a red flag, you’re filtering for compliance, not for the operators who’ll actually protect your resources later.

3. “Don’t give detailed feedback, legal risk”

This one isn’t about law. It’s about avoiding a screenshot.

The market has gotten smaller than it feels. A candidate you ghost today is a reference check, a LinkedIn comment, or a Glassdoor review tomorrow. Your employer brand is built in the moments nobody’s watching, specifically, the rejections.

Vague, thoughtful feedback given with consent costs a recruiter fifteen minutes. Silence costs you a reputation you’ll need the next time you’re hiring for the same role.

4. “Speed beats everything, close before the counter-offer”

A rushed decision reads as pressure, and pressure sends people back to the safety of what they already have.

More importantly: the candidate has a pipeline too. While you’re evaluating them, they’re evaluating two or three other offers. That’s not disloyalty, that’s a rational person making a decision that will shape the next year of their life.

Founders who pressure candidates to decide fast usually lose the ones with the most options, and keep the ones who had nowhere else to go.

5. “Bigger database means better recruiter”

This is the vanity metric of the entire industry.

A thousand contacts with no accurate status tell you nothing. Fifty contacts with an honest, current answer to “is this person open right now” are worth more than any CRM export.

If your recruiter can’t tell you, in seconds, where your top ten candidates actually stand, the size of the database was never the asset. It was the illusion of one.

6. “Culture fit means people who feel like us”

This is bias with better branding.

A team built on similarity loses the one thing that catches blind spots: difference. If your recruiter’s shorthand for culture fit is “I liked them,” that’s not a filter, that’s a preference dressed up as a process.

The better question isn’t whether a candidate feels familiar. It’s whether they strengthen what your team is currently missing.

7. “The hiring manager’s brief is the final word”

Most bad hires don’t start in the candidate pipeline. They start in the gap between the outcome a hiring manager describes and the profile a recruiter builds from it.

A recruiter who takes the brief at face value is a vendor. A recruiter who pushes back, asks what problem this hire actually needs to solve, and is willing to say the brief is wrong, is a consultant.

That distinction is the entire difference between filling a seat and solving a business problem.

Ось новий блок, у тому ж форматі, спеціально для founder-аудиторії — про те, чому воронкоподібний репорт це не показник здоров’я процесу:

8. “The funnel report proves the process is working”

Founders love a funnel chart. Top of funnel, screened, interviewed, offer, hire, clean shape, easy to present at a board meeting.

The problem: hiring isn’t linear, so the report is fiction dressed up as data.

A candidate who dropped out at screening two months ago can come back when a better-fit role opens. Someone who declined an offer can reach out again when their situation changes. The funnel shows a straight line down. Reality is people moving in and out of the process at different points, for different reasons, on their own timeline.

A funnel-shaped report answers one question: how many people did we move through stages. It doesn’t answer the questions that actually predict whether your next hire works out.

If you want metrics that mean something, ask for these instead:

• Time to fill, but segmented by role level, not averaged across everything, because averaging a senior and a junior hire hides where the actual bottleneck is

• Quality of hire, measured at the 6 and 12-month mark, not at signing, because a great offer acceptance rate tells you nothing about retention

• Candidate experience, specifically response rate to feedback requests and unsolicited referrals, because both are a proxy for whether people trust the process enough to come back or send someone else

• Offer-to-acceptance ratio broken down by why people said no, comp, timeline, counter-offer, because each of those points to a different fix, and lumping them together hides which one is actually costing you hires

A funnel tells you activity happened. These tell you whether the activity produced the right outcome.

If your current hiring reports only show volume moving through stages, that’s not a measurement problem, it’s a visibility gap, and it’s usually the first thing that surfaces in a Talent Strategy Package audit.

None of these habits are dramatic on their own. Stacked together, they’re a system quietly optimized for speed over judgment, and for volume over fit.

The cost doesn’t show up on the requisition. It shows up eight months later, in a resignation, a bad rehire, or a role that’s been open twice.

If you want a straight read on where your own hiring process is bleeding, that’s exactly what a Power Hour is for, one hour, no fluff, just the gap between your brief and your outcomes.

What’s the last hire your team made that looked right on paper and wasn’t?

Few of them anonymize candidate data at any point in that pipeline.

Now the same companies, on the same platforms, are rolling out tools to detect AI-written content and flag it as slop.

Something doesn’t add up.

The pipeline nobody audits

Ask a founder whether their ATS anonymizes resumes before an AI model screens them. Most won’t know. Ask whether the vendor retrains on that data. Fewer will know. Ask who is accountable if the model was trained on biased hiring patterns from five years ago. Silence.

This isn’t a technical oversight. It’s a governance gap that leadership has been comfortable ignoring, because AI adoption looked like efficiency and efficiency looked like a win.

Candidate data moves through screening, ranking, and sometimes rejection, touched by AI at every stage, with no consistent standard for what gets anonymized and what doesn’t.

The same platforms, a different problem

LinkedIn and Substack have both introduced features that flag content as AI-generated. On the surface, this looks like quality control. A way to push back against generic, low-effort posts flooding every feed.

Both platforms also disclose that this same feature trains their own AI models.

So the tool built to fight AI-generated content is, at the same time, teaching AI to sound more human. The people flagging content are training the system that will make flagging harder to do accurately next year.

I tested this myself. A messy, unpunctuated draft I wrote by hand got flagged as AI. The same draft, lightly edited by Claude for grammar, got flagged as AI too. If the model can’t tell hand-written from AI-assisted, what exactly is it optimizing for.

Who actually wins here

Not the recruiter using AI as a drafting tool. Not the candidate whose data moves through an unaudited pipeline. Not the writer opting out of a detection feature to protect their own work.

The platforms win twice. Once when they sell the AI tool as a productivity feature. Again when they use our reaction to that tool, our flags, our opt-outs, our edits, to train the next version of the same model.

We are not passive victims of AI adoption. Every prompt, every flag, every “improve this” click feeds the system. Blaming the model without examining our own participation in training it is not an honest position.

Three questions worth sitting with

1. Where in your hiring pipeline does candidate data touch AI without anonymization, and who signed off on that.

2. If a detection tool can’t distinguish careless human writing from AI-edited writing, what is the actual business case for using it as a filter.

3. Are you building AI into your talent process because it solves a defined problem, or because not doing it felt like falling behind.

None of these have comfortable answers. That’s usually the sign they’re the right ones to ask before the next tool gets rolled out.

A global report on a billion job postings. A Pope Leo xiv document on artificial intelligence. A legal right not to use AI at work and a Apple vs OpenAI hiring lawsuit.

None of these made it onto a hiring dashboard and all three should have!

The global split

PwC’s 2026 Global AI Jobs Barometer analysed more than a billion job ads across 27 countries. The finding that matters most for anyone building a team: the labour market has split into two tracks.

On one track sit roles PwC calls professionalised. AI takes over the routine parts of the job, and what’s left requires more judgement, not less. On the other track sit roles PwC calls democratised, where AI makes the job easier to do with less expertise.

The professionalised track is growing job openings twice as fast, with wages growing 42% faster since 2021.

PwC’s own example of a professionalised role: recruiter.

That is not a coincidence and it is not flattering rhetoric but a structural claim. As AI absorbs sourcing, scheduling, and first-pass screening, what remains in recruiting is judgement: reading a market correctly, structuring an interview process that predicts performance, deciding who actually fits a team that doesn’t exist yet. That work is getting harder to automate, not easier, and the data shows it is getting paid more, not less.

The same report found something sharper at entry level. Junior roles most exposed to AI are now seven times more likely to demand senior skills, things like leadership and independent decision-making, than junior roles with less AI exposure. Since 2019, these “seniorised” entry-level roles have grown 35% while ordinary entry-level roles shrank 10%.

Translation for anyone hiring juniors: the job title says junior. The job description increasingly doesn’t.

A war economy as a hiring case study

Most labour market shifts are driven by economics. Interest rates, funding cycles, demand curves. Ukraine’s market in H1 2026 is a case study in something founders rarely plan for: a market reshaped by a force with nothing to do with economics at all.

DefTech, the defense technology sector, posted a record 1,285 vacancies in June alone, 19% of all IT vacancies on the country’s largest job platform that month. Six months earlier, that share was a fraction of it. Meanwhile Front-end and Mobile development, the categories a “normal” tech market would expect to lead, lost hundreds of open roles over the same half year.

This isn’t a story that only applies to Ukraine. It’s a preview! Any company operating across borders, supply chains, or regulatory zones can hit a moment where a talent market reorganizes itself around a force that has nothing to do with product-market fit, and everything to do with the world around it.

The operational lesson isn’t about defense tech specifically. It’s about detection speed. A hiring system built only to optimize inside a stable market has no mechanism for noticing when the market itself has stopped being stable. The companies that will handle the next shock well are the ones whose hiring strategy already assumes shocks happen.

Hiring just became a legal exposure surface

Two stories broke within days of each other this summer. Neither started as an HR story. Both landed on HR’s desk anyway.

The first is about what happens when you mandate AI use. On May 25, 2026, Pope Leo XIV published his first encyclical, Magnifica Humanitas, a 200-page document warning against concentrating AI power in the hands of a few and calling for stronger public oversight of the technology.

Ten days later, Business Insider reported that a US software engineer, a Unitarian Universalist, had been granted a formal religious exemption from using AI at work, citing the Pope’s position that AI can displace human dignity. Her employer approved it under Title VII of the Civil Rights Act.

One case doesn’t make a policy. But employment lawyers are describing a clear trend line, not a one-off. One Ogletree Deakins attorney told HR trade press he used to see a religious accommodation request roughly once a month. He’s now seeing three or four a week, and AI mandates are a growing share of them. The EEOC’s June 2026 enforcement plan has flagged the space directly.

Here’s the part that should land on every People leader’s desk: under the current legal standard set by the Supreme Court’s Groff v. DeJoy decision, employers can only deny a religious accommodation if it creates a “substantial increased cost,” not mere inconvenience. That bar is high, and it was set before generative AI became a default job requirement almost everywhere.

If your company has a policy that says “use the AI tool” as a hard requirement rather than a default, you may already have a gap. Not a hypothetical one.

The second is about what happens on the other side of the interview table. On July 10, 2026, Apple sued OpenAI, alleging that OpenAI and two former Apple employees misappropriated confidential hardware designs, and that this happened, in part, through the interview process itself. The complaint describes current Apple employees interviewing for OpenAI roles allegedly being drawn into handing over confidential product information during those conversations. OpenAI has pushed back, saying it sees no evidence the claims have merit.

Regardless of how the case resolves, the allegation itself is the operational warning. An interview isn’t only a place where you evaluate a candidate. It’s also a place where confidential information can move in either direction, and most companies have no protocol for that at all. The line between “tell me about a project you’re proud of” and soliciting a competitor’s trade secrets isn’t obvious in the moment. It becomes obvious in a courtroom.

Read together, these two stories point at the same blind spot from opposite directions. One is about what you can legally require a candidate or employee to do. The other is about what you should never let an interview quietly become. Most hiring processes have a policy for neither.

What this means if you’re running hiring, not just watching it

Put the three together and one pattern shows up three times.

• The job market is no longer one market. It’s splitting by how AI touches the role, and the split rewards judgement over task execution, at every level, for every hire you make from now on.

• Structural shocks don’t announce themselves as HR news. They show up as a defense sector absorbing a fifth of a country’s tech vacancies in a month, and a hiring system either notices or it doesn’t.

• And the operational rules you wrote for hiring and onboarding six months ago may already conflict with employment law that moved faster than your handbook did.

None of this is solved by a better job ad or a faster ATS. It’s solved by treating hiring the way you’d treat any other system with real business risk attached to it: something that gets audited, stress-tested, and rebuilt when the environment underneath it changes.

That is the work I do with founders and People leaders who’d rather get ahead of these shifts than clean up after them. If you want a structured look at where your hiring system has gaps right now, that’s exactly what a Power Hour is for. If the gaps are bigger than one hour, the Talent Strategy Package is built for that.

If this was useful, the honest way to say thanks is a paid pledge

here on Substack. A portion of every subscription goes to Ukraine’s defense. Stripe doesn’t have a line item for “helping a democracy survive,” but this is the closest workaround we’ve got.

Pledge

Sources:

• PwC 2026 Global AI Jobs Barometer

• DOU: IT Labour Market Analysis, H1 2026

• TechCrunch: The Pope’s AI Encyclical Isn’t Really About AI

• Mondaq / Ogletree Deakins: Employers Should Prepare for Religious Objections to Workplace AI Use

• HCAMag: Can an Employee Refuse to Use AI for Religious Reasons?

• Fortune: Apple Accuses OpenAI of Stealing Hardware Trade Secrets

• TechCrunch: The Wildest Allegations in Apple’s Trade Secrets Lawsuit Against OpenAI

• TechCrunch: OpenAI Pushes Back on Apple Trade Secret Lawsuit

Nobody checks the other failure point: whether everyone running the process is actually looking for the same person.

Last week I ran an experiment with three recruiting teams during Interview Expert, our third live cohort. Same brief. Same access to me as the client. Here’s what happened.

The case

RIST, a 40-person B2B SaaS company. Churn has been climbing for three months. Customers leave right after their first month. There has never been a Customer Success function, sales has been covering it part-time. My brief as the client: find the first CSM who understands customers, builds onboarding, and cuts churn.

Three teams got the identical brief. They could ask me anything to fill gaps the brief didn’t cover. Each had the same amount of time to build a role profile.

They came back with three different roles.

• Team one optimized for scale. Experience building processes from zero, high account volume, ownership, end-to-end thinking.

• Team two optimized for context. Experience in the US, UK, ideally LatAm markets, handling upset customers, cross-functional work, comfort with AI tools.

• Team three optimized for product. Prior B2B SaaS experience and hands-on CSM process-building were non-negotiable. Everything else was secondary.

All three profiles were defensible. All three were built on the same brief, the same call, the same access to me. And all three described different people.

Why this happens

What I actually said as the client was: understands customers, builds onboarding, cuts churn. That’s three outcomes. Not a profile.

Everything between an outcome and a profile gets filled in by the recruiter, out of their own experience and their own assumptions about what kind of person usually gets there.

There’s research that quantifies exactly how easily this happens. Schmidt and Hunter’s meta-analysis, one of the most cited in personnel selection research, found that even with an identical brief and no ambiguity in the input, different interviewers evaluating the same candidate on separate occasions only agree with each other at about 0.37. That’s disagreement at the evaluation stage, when the role profile is supposedly already settled.

Our case shows a gap that opens earlier. The three teams split before a single resume existed, at the stage of reading the brief itself. If interpretation of the brief already diverges, the eventual disagreement on the candidate will run wider than 0.37 alone suggests.

Why structure isn’t bureaucracy

Structure here has nothing to do with paperwork. It means the team agrees, before the first candidate call, on three things: which competencies are mandatory versus nice-to-have, which questions test each one, and what scale is used to score the answer. All of it locked in once, before anyone forms a personal impression of a specific person.

None of the three RIST teams did this. Each had the same input and the same chance to ask clarifying questions, but none checked with the others on what counted as mandatory versus secondary. Team one decided scale and autonomy were critical and market exposure wasn’t. Team two decided the opposite. Had they compared notes after building their profiles, the gap would have surfaced immediately, before a single candidate entered the pipeline.

That’s the real difference between a structured and an unstructured process. Not how many questions are written down. Whether there’s one system you can hand to a second recruiter, a third interviewer, or the founder, and get the same profile back, instead of a new variant built from someone’s personal idea of the ideal hire.

A role profile isn’t a compliance artifact. It exists to force everyone in the process, from screening to the final round, to build out the brief the same way, before interviews start rather than during them.

What this costs the candidate

For the candidate, a missing shared profile shows up as inconsistency they can’t explain through anything they did.

Screening asks about autonomy and scale. The final round suddenly asks about AI tooling and LatAm experience, neither of which came up before. This is a real case, not a hypothetical. The candidate doesn’t fail because they’re weak. They answered questions for profile A while being scored against profile B.

That’s the actual cost of misalignment. Not just wasted process time bouncing between criteria. A candidate held hostage by whichever person or team happened to run their interview that day.

A candidate can’t force a company to align internally. But they can lower their own risk of falling into the gap between profile A and profile B.

At the first screening call, ask directly who else is involved in the process and what each stage is evaluating against. A vague answer is itself a signal that no shared profile exists inside the company.

Keep track of which competencies each interviewer raised. If new criteria show up at the final stage, name it directly: is this a departure from what was discussed earlier, or an additional layer.

And remember the core point: a rejection after a process with shifting or unclear criteria says more about the company’s lack of a system than about the candidate’s fit.

What this means for the recruiter

A recruiter doesn’t control how precisely a founder or hiring manager articulates a request. They do control what happens the moment that request becomes the team’s work.

After intake, write the role profile down, including what’s mandatory versus optional, and send it to everyone involved before the first candidate call. Not a verbal recap in a meeting. A text everyone can reread and point back to later.

If more than one person is running interviews, agree upfront on who owns which part of the profile, so two interviewers aren’t testing the same thing while a third competency goes unchecked by anyone.

When a hiring manager or client adds a new criterion after several rejections, treat it as a stop signal, not something to quietly fold into the next search. A new criterion means the original profile was incomplete, and it needs updating for everyone, not held as one person’s private addendum.

And the core point: if after a group discussion of the profile the team still senses everyone is picturing a slightly different person, that’s not a detail to defer. That’s the moment to stop and reconcile, before the cost lands on candidates.

The takeaway

Before opening a role, check one simple thing: hand this brief to three different people on your team, and see whether you get three versions of one role, or one.

In the RIST case, the answer was obvious within an hour. In real hiring, nobody checks, because the profile feels clear right up until you put it next to another profile built from the same brief.

The difference is that in a training exercise, three versions of a role can be compared and reconciled before anyone loses real time. In an actual process, those three versions rarely end up in the same room. They pass through the candidate one at a time, and the candidate pays for the gap in their own time and their own uncertainty.

If the honest answer to that opening question is three, the problem isn’t the candidates yet. It’s the brief, and whether anyone on the team took responsibility for turning three outcomes into one profile instead of three.

For People and TA leaders, this isn’t an abstract debate. We’re often in the room when a role gets named, scoped, and benchmarked, sometimes before anyone has decided what it actually involves. And the name we help choose tends to outlive every other decision made that day.

The moment a price gets set

Think about how a brand-new profession gets its price at all.

Five years ago it didn’t exist. So it has no settled title, no settled band, no settled idea of who “naturally” belongs in it. Everything is still soft. And that softness is exactly when the most consequential decisions get made, quietly, often by default.

The same underlying work can be called a “growth engineer” and carry a premium, or keep an older title and sit near the floor.

The gap between those two outcomes is large, and it correlates only weakly with the actual content of the job. What it correlates with is harder to admit.

The clearest example is one that already finished

Programming started as low-status, low-paid work.
In its early decades it was seen as something closer to clerical labor than to engineering, the routine part that came after the “real” work of building the hardware. The word software sits one rung below hardware in that hierarchy for a reason.

Then something happened that usually doesn’t.
As a rule, when a profession shifts from male to female, it loses status and pay, that’s the well-documented direction for teaching, for nursing, for much of care work. Programming ran the other way. The prestige, the word “engineer,” and the money arrived later, and they arrived alongside a change in who held the role.

And this isn’t a feeling. Economists Levanon, England and Allison analyzed fifty years of U.S. census data, profession by profession, using fixed-effects models, comparing each occupation against itself over time rather than against other jobs.

The finding: as the share of women in an occupation rises, its pay falls about a decade later, and not because the work changed. The work done by women gets systematically revalued downward. The mechanism even has a name in the literature: devaluation theory.

Worth underlining, because it’s counterintuitive: it isn’t the individual who gets devalued, it’s the role. When an occupation fills with women, everyone who stays in it earns less over time, men included. Price follows who’s understood to hold the role, not the difficulty of what the role requires.

Why this is a live problem, not a history lesson

The convenient move is to file this under “history.” It happened to programming half a century ago, we’re more aware now, it won’t repeat. I’m not sure we’ve earned that confidence, and here’s why.

The labor market is in a rare state right now: new roles are being created at scale. AI, data, automation, and new infrastructure are producing jobs that didn’t exist five years ago, and none of them yet have a fixed name, a fixed price, or a fixed assumption about who belongs in the seat. Growth roles, hybrids of marketing and code, new blends of analytics and product. In my own field, recruiting, the same thing is happening: recruiting ops, recruiting engineer, roles that wrap the old work in new infrastructure and new titles, and watch the status and the band move with the name.

The mechanism I described works best in exactly these conditions. It doesn’t operate on the settled, it operates on the undefined, because the undefined is easy to define in someone’s favor. While a role’s name, status, and price are still unfixed, that’s precisely when it’s decided whether it becomes “engineering” and expensive or “support” and cheap, and that decision tracks the content of the work only loosely. We’ve seen this script before. We know how it ends. What we rarely get is the chance to recognize it early, rather than reading it off a census a decade later.

No conspiracy, which is the harder part

I want to be precise: I’m not describing a conspiracy. There’s no room where someone decides how to allocate the new professions. If there were, the problem would be easy, you’d just find the room.

What’s actually happening is subtler, and worse for being subtle. The market tends to get more expensive not when the work gets harder, but when the assumption shifts about who does it, and we backfill “difficulty” as the explanation afterward. No one intends it. It’s a blind spot baked into how we decide what counts as “expensive” and “skilled.” And blind spots reproduce themselves, with no author and no malice, until someone looks at them on purpose.

What this means for the people who advise on pay

This is where TA and People leaders actually have leverage, more than we usually admit, and it runs two ways.

First, the obvious one. When a role suddenly costs more after a rename, ask plainly what’s being paid for: new value, or a new label. Sometimes it’s genuine new value, a real technical layer that wasn’t there before, and the premium is earned. Sometimes it’s the label, and the company is paying a market markup on a word. Telling those two apart is one of the few hiring skills that pays for itself directly, and we’re the ones positioned to tell them apart in the room.

Second, the mirror, and the more expensive one. Ask how much real value already sits in your “cheap” roles, the ones you underprice only because they still carry old names. Devaluation works in this direction too. It makes your strongest content, recruiting, or support specialist structurally cheaper than she’s worth, and it won’t register as a problem, because you’re paying “market.” Until a competitor who noticed first renames the role, adds a little scope and a lot more pay, and takes your person for a number that looks inflated right up until she’s gone.

Takeaway

The most expensive thing in hiring isn’t the strong candidate you overpaid for. It’s the value you fail to see for years, because it carries the wrong name, the wrong seat, and the wrong assumption about who provides it.

What makes this moment unusual is that the prices for an entire class of new professions haven’t hardened yet. They’re being set right now. And the logic by which they settle depends, in part, on whether the people closest to those decisions, us, are watching the process while it happens or reading it back afterward, when nothing can be changed.

So here’s the question I’d put to anyone who advises on how roles get scoped and paid: does a role get more expensive because it got harder, or do we only let ourselves see the difficulty once the role has already changed hands?

PwC recently analyzed over a billion job ads across continents, and one finding matters more than it usually gets credit for. As AI use grows, employers increasingly value skills that used to belong to experienced people: critical thinking, decision-making, stakeholder work, adaptability. Interesting on its own, but underneath it sits something bigger.

We’re used to thinking AI takes over routine work, and it does. Almost nobody talks about the other half: that routine work was, for decades, how new people learned the job. A recruiter builds judgment by screening hundreds of imperfect resumes and getting plenty of calls wrong. An analyst learns by cleaning data and writing reports nobody celebrates. It was never the valuable part. It was the part that built the experience the valuable part needs.

Now AI does most of it. So a company no longer needs a person for the simple tasks, but it still expects a specialist with the skills those tasks used to build.

That’s why roles labeled Junior increasingly read like Middle. And this isn’t a feeling, it’s in the data. In PwC’s 2026 report, the junior roles most exposed to AI are seven times more likely to demand traditionally senior skills than roles where AI is barely used. The first rung didn’t disappear. It rose so high it’s hard to reach from the ground.

The surrounding numbers say the same. Entry-level postings in the most AI-exposed sectors have flatlined overall, but within them the “seniorized” junior roles, the ones asking for more than they should, grew 35% since 2019. The market didn’t stop hiring beginners. It stopped hiring them on simple terms. And the bar moves fast enough that even people already in the profession struggle to keep up: the skill set for AI-exposed roles now turns over more than twice as fast as everything else, a gap that widened 75% in a year.

What this means if you’re building a team

For founders and early talent leaders, this lands as a very specific kind of frustration. You open a junior role at a junior budget, and the pipeline is thin. Your instinct says the market is weak, so you raise the bar, look for someone “already strong,” and end up chasing a Middle at the same price. Still nothing.

Here’s the mechanism underneath it. AI removed the simple work that a beginner used to do to pay for themselves in the first months. With that work gone, the economic logic that justified hiring an inexperienced person went with it. The hiring manager isn’t being unreasonable when they want someone immediately useful. There’s genuinely no slack to onboard slowly. But that pressure quietly turns every junior req into a Middle req, and then the company is surprised the role won’t close.

The trap is that this compounds. You skip junior hiring this year because it’s hard, and you do it again next year for the same reason. Two years in, you discover Middle talent is scarce too, because the market, you included, stopped growing it. The shortage you’re complaining about is partly one you’re producing.

What’s actually in your control

Two moves, and they’re really one strategic decision split in half.

The first is cheap and almost everyone skips it: separate what the role truly needs on day one from what got added by inertia. A junior spec is often not a real profile, it’s the stacked wishes of several people who each hedged. Sit with the hiring manager and ask, for every requirement, “what does this person have to do in month one for you to know they’re working out.” Half the list falls away on its own. That’s not cosmetic editing, it’s returning the role to an honest level, and honest roles attract the people they were actually written for.

One example. A client opened a junior analyst role asking for “dashboard-building experience and stakeholder management.” Asked why a junior needed that, it turned out there was one stakeholder, the team lead, and nobody would let a new hire touch dashboards for months. The real need was someone careful with data who asks questions when something doesn’t add up. We rewrote the role around that, and relevant applications jumped immediately, because the right people finally recognized themselves in it.

The second move is harder and more important: put cost back into the conversation. When a hiring manager insists on Middle requirements at a junior budget, that’s not stubbornness, it’s not seeing that the market doesn’t work that way. Your job is to make the economics visible. Here’s what a ready-made Middle costs. Here’s what it costs, and how long it takes, to grow a junior into one. Here’s what your Middle pipeline looks like in two years if you never hire junior at all. That reframes a single req into a question about where your people come from tomorrow.

The strategic choice

This is one of the most serious shifts AI brings to the labor market, and it rarely gets named. We talk endlessly about productivity and automation, and far less about how a specialist is supposed to develop now. If AI absorbs most of the starter tasks, where does a beginner learn the craft? Where do they get to be wrong, get feedback, and slowly take on harder things?

For a company, this isn’t an abstract worry. It’s a build-or-buy decision you’re now making whether you notice it or not. Buy, and you compete for an expensive, shrinking pool of ready talent, paying the premium every time. Build, and you take on the cost and patience of developing people in a world where the old on-ramp is gone, which means designing a new one on purpose: real ownership early, deliberate feedback, AI used to compress months of pattern-exposure into weeks rather than to remove the learning entirely.

Neither path is free. But most companies are defaulting into “buy” without deciding to, and then calling the result a talent shortage. The ones that will have a team in three years are the ones treating this as a choice, and making it on purpose.

The useful question isn't whether the junior market is broken. It's narrower and more uncomfortable: in your company right now, is there any role where a person is allowed to be junior? Where they can do work that isn't immediately valuable, be wrong, and get better on someone else's clock?

If the answer is no, you don't have a sourcing problem. You've quietly decided to stop growing people, and the shortage you'll feel in two years is one you're choosing today.

The message was everywhere: the problem is the people who do the hiring. So if we automate enough of the process, the problem solves itself.

Then AI showed up, and companies automated nearly everything that could be automated: sourcing, finding contacts, message personalization, resume screening, interview notes, scheduling, scorecards, the first rounds of communication. What used to take hours now takes minutes.

And here’s the question I can’t shake: if the problem was the recruiters, why is it still here after all that automation?

• Why do companies still pass on strong candidates?

• Why do they hire people who don’t make it through probation?

• Why do they reopen the same role six months after filling it?

• Why can’t a team agree on who they’re actually looking for?

• Why do two interviewers give opposite verdicts on the same person?

Maybe the problem was never where we were looking.

The uncomfortable truth about hiring

Strip it down and hiring is an attempt to predict how someone will perform a year or two from now. Not roughly — precisely enough to spend tens, sometimes hundreds of thousands of dollars on one person. Because that number isn’t just salary. It’s the cost of search, hiring, and onboarding — an investment that only pays off when the person actually works out, long-term.

How do we make that call? We take someone we’ve never seen do the job, talk to them for a few hours, read a resume, ask some questions, maybe add a test. Then we decide whether they’ll be productive and motivated for years.

Spend enough years in hiring and you stop not noticing it: we make one of the most expensive decisions a company makes, based on a few hours of conversation.

So the hiring problem was never a technology problem. It’s that we still can’t reliably predict how effective a person will be. And everything we automated over the last two or three years happened around that problem — never with the problem itself.

The experiment that should have changed everything

Steve Yegge, former Google Hiring Committee member and Amazon Bar Raiser — recently described a story that confirmed something I’ve been saying for years, both in my course and in conversations with clients about evaluating candidates. (You may have seen it making the rounds: “The Last Technical Interview”)

This was a group of about fifteen people — effectively the final arbiter of all hiring at the company. Co-inventors of major technologies, authors of interviewing books, people who now hold very senior positions. They were confident they knew their craft. And honestly? They did. Their contribution to how that company hired is hard to overstate.

One day, recruiters handed them a separate batch of packets to evaluate. No names, no company names — just resumes, interview results, interviewer notes, and candidate answers. This is called a calibration exercise: everyone involved in hiring evaluates the same candidates to check what each person pays attention to, and whether they even agree on who they’re looking for. In plain terms, the team aligns expectations for the role so they don’t argue later about whether someone fits. The group assumed the packets were from another office — cross-site calibration was routine.

Worth pausing here on how a “blind” packet actually reads. Strip the name and the company off a resume, and what we actually lean on rises to the surface — the part we usually don’t notice behind the brand names:

Look at the screenshot. Every note in the margin is a bias with a respectable name — halo, confirmation, affinity, groupthink, stereotype — and together they're exactly what we like to call interviewer instinct. As long as the packet carries company names and familiar markers, they run quietly in the background and feel like judgment. Take those markers away, and it turns out the decision was resting less on the person than on the labels.

So the result was completely unremarkable. They rejected about two out of three.

The problem: these were their own packets. The documents they themselves had been hired on.

People who’d made it into Google. Who went on to become strong engineers, leaders, architects. Who built this very hiring process with their own hands.

Voted not to hire themselves.

And the worst part isn’t even that. The worst part is that after the experiment, the system barely changed.

Because the experiment didn’t expose a mistake by a few interviewers. It exposed something deeper. Maybe we overestimate how accurate the process is in the first place. And we do it systematically, year after year, with real money behind every decision.

The biggest illusion in recruiting

Most people assume the hiring process is built to pick the best candidate. But look honestly at how companies operate and you see something else: most of the time it’s built not to find the best candidate, but to minimize risk. Those are fundamentally different jobs.

The best candidate might be unconventional — nervous in the interview, thinking differently, not great at packaging their experience. The lowest-risk candidate looks nothing like that: familiar, saying the right words, similar background, not making the team doubt anything.

Between the two, companies almost always pick the second. Not out of incompetence, but by design — a system that rewards familiarity over strength and seniority.

We’ve all heard “not a culture fit” and “overqualified.” And we all know perfectly well: it’s almost never about the candidate. It’s about the company that’s uncomfortable with them.

That’s how the strongest people end up out of the game before the offer.

What AI actually revealed about recruiting

AI did one genuinely useful thing — it stripped away a layer of work we’d spent years confusing with the real value of recruiting.

• We automated sourcing — and saw that sourcing was never the main problem.

• We automated outreach — and roles didn’t start closing instantly.

• We automated screening — and teams didn’t start making better decisions.

• We automated the admin and ops work — and the disagreements between stakeholders didn’t go anywhere.

It turns out the hardest part of hiring isn’t finding the person. The hardest part is agreeing on who you’re looking for, soberly assessing potential, and making a decision when full certainty isn’t available — and won’t be.

This isn’t an operational problem; it’s a problem of thinking and alignment. And no AI has solved it yet, because it’s not about speed. It’s about the quality of judgment.

How to make fewer mistakes

Teams that hire well don’t try to make the process perfect. They try to make it less wrong. Those are different ambitions, and the second one is more honest.

They calibrate interviewers.

Before the hiring starts, they discuss more than the role requirements — they discuss signals.
- What actually points to future success?
- Which criteria are critical, which are just nice-to-have?
- What do we count as evidence of competence, and what just makes a good impression?

Without that conversation, different interviewers evaluate different things and then sincerely wonder why their scores don’t line up.

They check their own predictions.

Six to twelve months later, they go back to the people they hired and compare the prediction to reality.
- Who turned out stronger than expected?
- Who weaker?
- Which signals worked, and which turned out to be noise?

That’s how decision quality is built. Not through a new tool, but through feedback most companies never give themselves.

They separate “I like them” from “they fit.”

Sounds obvious. But this is where a huge share of mistakes is born. People gravitate toward those who think alike, talk alike, share a similar background.

It feels comfortable, like a good decision. But similarity and effectiveness aren’t the same thing — and every hiring cycle built on comfort quietly narrows the team.

So would you hire your top performer today?

This is a question I often put to clients when I help them build their hiring system: if your best people interviewed at your company right now (yes, the ones you’re proud of, the ones you’d hate to lose) — would they get an offer?

Honestly? For most companies, this question becomes a wall. They say “YES, of course” almost without thinking. And then the process audit shows otherwise. That’s what worries me most.

Over the last two or three years, we automated nearly everything around hiring. We learned to find candidates faster, schedule interviews faster, write feedback faster, process information faster. Faster — but not more accurate.

Because the core question is the same as it was five years ago: will this person succeed in this role a year from now? And we answer it roughly the way those Google engineers did when they voted against themselves — confidently, and wrong.

AI removed everything around the hiring decision. But the decision about a person is still made by a person. And until we learn to do that more honestly — checking predictions against reality, separating “one of us” from “strong,” questioning our own confidence — no automation will save us.

The technology got smarter. The question is whether we did.

1. The application stage flooded with automated agents.

2. Resume screening that statistically favors the tool doing the screening.

3. Live interviews where the cheating software is invisible to your recording.

4. Reference checks that can be scripted.

5. Take-home assignments that are functionally unverifiable without a live defense.

If you’ve read the series, you have a clear picture of the mechanics. This final piece isn’t more mechanics. It’s the strategic question that sits above all of them: what does this mean for how you build and run a hiring function in 2026?

The arms race framing is wrong

The dominant narrative around AI fraud in hiring is an arms race: candidates get better tools, companies get better detection, candidates adapt, detection improves. This framing leads to a particular set of decisions — more surveillance, more proctoring, more verification layers.

It’s not wrong. But it’s incomplete. And the part it misses is strategically important.

Only 8% of job seekers describe AI-enabled hiring as fair. At the same time, 70% of hiring managers trust AI to make faster and better hiring decisions.
Both sides are automating. Both sides are dissatisfied with the results…and both sides are responding to the other’s automation with more automation of their own.

The companies investing heavily in detection are solving a real problem. But they’re solving it inside a system that created the conditions for the problem. The deeper question — one that detection tools don’t address — is what kind of hiring process produces reliable signal about whether someone can actually do the work.

What the data says about where this is heading

Three developments from 2025-2026 that matter for organizational strategy:

1. Meta, Google, Canva and Shopify have moved in the opposite direction from detection — they’ve legalized and evaluated AI use during technical interviews. Meta’s AI-enabled coding round, rolled out from October 2025, replaces one of two traditional coding rounds. Candidates work with a multi-file codebase and an integrated AI assistant. The evaluation criteria: problem solving, code quality, verification, and communication — specifically including the ability to catch AI mistakes and defend every line of generated code.

   This is a meaningful signal. The companies with the most sophisticated hiring infrastructure aren’t building better fraud detection. They’re redesigning the interview to make fraud structurally irrelevant — because the skill they’re evaluating is how well someone works with AI, not whether they can solve a problem without it.

2. 72.4% of recruiting leaders are now conducting more in-person interviews specifically to combat fraud. Google, Cisco and McKinsey reintroduced mandatory in-person rounds for sensitive roles in 2025. This is the blunt-force solution: remove the conditions that enable remote fraud. It works. It also sacrifices the candidate pool expansion that made remote hiring valuable, and it concentrates hiring toward candidates who can physically appear — which has its own implications for diversity and access.

3. Cheating rates jumped from 9% in July 2025 to 45% by September 2025 — in two months — driven by viral social media content and the release of stable invisible overlay tools.
   The implication for strategy: candidate behavior can shift dramatically in a short window when the tooling changes. A process that looked adequate in Q2 may be significantly exposed by Q4. This is not a problem you solve once.

The three strategic decisions that actually matter

Everything in this series reduces to three decisions that People leaders need to make explicitly — because the default position on each of them is increasingly untenable.

Decision 1: What are you actually measuring?

Most hiring processes were designed to measure a candidate’s ability to perform under standardized conditions. The problem is that standardized conditions are exactly what AI tooling is optimized for. If your process can be gamed with a $30 monthly subscription, you’re not measuring capability — you’re measuring preparation and tool access.

The alternative isn’t to eliminate standardization. It’s to add moments that can’t be standardized: live defense of specific decisions, dynamic constraint shifts, behavioral depth that requires genuine personal history. These don’t require expensive tools. They require deliberate process design.

Decision 2: Where does human judgment live in your process?

The efficiency case for automating screening is real. Recruiter workload increased to 588 applications per role in Q3 2024 — a 26% increase from the previous year. Nobody is reading 588 resumes. Automation at the top of funnel is not optional at this volume.

But the self-bias finding from the arXiv study changes the calculation: if your screening tool statistically prefers resumes that match its own output style, you’re not selecting the best candidates. You’re selecting the candidates who used the same tool. Human judgment at the final shortlisting stage isn’t a luxury — it’s a correction mechanism for a structural flaw in automated screening.

The question isn’t whether to use AI in your process. It’s where human judgment is non-negotiable, and whether you’ve explicitly protected those moments.

Decision 3: How do you validate the signal your process produces?

This is the least glamorous question and the most important one. Most hiring functions have no systematic way of knowing whether their interview process produces reliable signal. They hire, and then something happens, and if it goes badly they usually attribute it to onboarding or management or fit — not to a process that failed to measure what it claimed to measure.

The only way to know if your process works is to track what it predicts against what actually happens. Document what each interview stage is designed to test. At 90 days, check whether it holds. If candidates who excelled in behavioral rounds consistently underperform on autonomy in the first quarter, your behavioral rounds aren’t measuring what you think. If strong technical screeners consistently struggle in code review, your screening is measuring the wrong thing.

This feedback loop is free to build. Almost nobody builds it.

What good looks like in 2026

It’s not a perfect detection stack. It’s a process deep enough that fraud becomes structurally pointless.

HireVue data shows that ChatGPT and other generative models perform poorly on practical job tryouts and only average on structured assessments — meaning strong genuine candidates still stand out. The implication: a well-designed process with genuine depth still discriminates between candidates accurately. The fraud problem is most acute in shallow processes where a generated answer is indistinguishable from a thoughtful one.

The organizations that will hire better in the next two years aren’t necessarily the ones with the most sophisticated detection. They’re the ones that have made their process harder to fake — not through surveillance, but through depth. Live defense. Constraint shifts. Behavioral drill that requires genuine personal history. 90-day validation loops that close the feedback between assessment and performance.

None of that requires enterprise tooling. All of it requires deliberate process design.

One conclusion

The hiring process is the first signal a candidate receives about how an organization thinks and operates. A process built around surveillance signals distrust. A process built around depth signals rigor.

Both will catch fraud. Only one will attract the candidates you actually want.

This completes the AI Fraud in Hiring series. All five parts:

Part 1: Detection tactics companies are deploying right now

Part 2: What actually happens at each stage of your hiring funnel

Part 3: The dangerous zone — where your process is most exposed

Part 4: What it actually costs to fight AI Fraud in Hiring

Part 5: What this all means (this piece)

None of the tools mentioned across this series represent sponsored content or vendor relationships.

That’s a reasonable place to stop describing the problem. This part answers the practical question: given what we now know, what should you actually spend money on, and how much?

The honest framing upfront: there is no single tool that solves this. The teams handling it well are building layered processes — some free, some paid — where each layer catches what the previous one misses.

The goal of this piece is to help you build that stack without overspending on things that don’t move the needle for your hiring volume.

*All tools mentioned in this piece were selected based on publicly available data and independent research. None of this is sponsored content and no vendor relationship exists with any platform listed.

Before the catalogue: what you get for free

The most effective changes don’t cost anything. This isn’t a consolation — it’s genuinely true, and worth stating clearly before any pricing discussion.

1. One unexpected follow-up question per interview answer costs nothing and catches more AI-assisted candidates than most proctoring software.

2. A 30-minute live defense after every take-home assignment costs one recruiter’s time and eliminates the primary fraud vector for that format.

3. A specific question about your product in the application form costs nothing to add and filters automated agents immediately.

4. VS Code Live Share is free and creates a genuine pair-programming environment where the candidate’s keystrokes are visible in real time — which changes the dynamic of a technical interview without any additional tooling.

If your current process has none of these in place, start here. The paid tools below add detection capability on top of a well-designed process. They are not a substitute for one.

Tier 1: Small teams — up to 50 hires per year
($0–$200/month)

CoderPad — technical interviews

CoderPad pricing: Free tier (2 interviews/month). Starter $70/month ($840/year, 60 interviews). Team $375/month ($4,500/year, 360 interviews).

What it actually does: creates a shared coding environment where both candidate and interviewer work in the same IDE. The value isn’t primarily anti-fraud — it’s that pair programming is inherently harder to cheat in than a screen-share-and-solve format. Code replay lets you review the candidate’s process after the interview.

What it doesn’t do: it won’t detect invisible overlays. A candidate using Cluely in a CoderPad session will still have the overlay visible to them. The value is in format design, not detection.

Best for: engineering teams doing 5-30 technical interviews per month who want a better interview experience and some natural friction against lazy cheating.

TestGorilla — pre-screening assessments for all roles

TestGorilla’s free plan exists but is too limited for real hiring. Core paid plan starts at $135/month (billed annually). The credit model catches some teams off guard — each assessment consumes credits, and volume can add up.

What it actually does: pre-employment skills assessments across a wide range of roles — cognitive ability, role-specific skills, personality. Includes tab-switch tracking and some anti-cheat features. Useful for filtering at the top of funnel before you invest recruiter time.

What it doesn’t do: deep technical assessment for engineering roles. The anti-cheat features are basic compared to dedicated technical platforms. Not designed for live interview detection.

Best for: non-technical roles, mixed hiring, teams that need a first filter before phone screens. Less suitable for senior engineering positions where assessment depth matters.

Tier 2: Mid-size teams — 50–200 hires per year
($200–$800/month)

HackerRank — technical screening and live interviews

HackerRank Starter: $100/month billed annually, 120 candidate attempts per year, $20 per additional attempt. Pro: $450/month, 300 attempts, unlimited user seats, ATS integrations with Greenhouse, Lever, Ashby.

What it actually does: strong technical assessment library (4,000+ problems at Pro), plagiarism detection, code replay, and an AI interviewer feature that generates adaptive follow-up questions based on candidate responses. The plagiarism detection is genuinely useful — it flags code that matches known solutions in its database.

The honest limitation: plagiarism detection catches copied solutions, not invisible overlay assistance. A candidate solving a problem with Cluely, typing their own keystrokes, won’t be flagged. The AI interviewer is the more interesting feature — adaptive follow-ups are structurally harder to prepare for than static questions.

Best for: engineering teams hiring 10-25 engineers per month who need a scalable technical screening layer with ATS integration.

Sherlock — live interview behavioral detection

Sherlock (withsherlock.ai) starts from approximately $2 per interview — which makes it the lowest-cost entry point for actual behavioral detection during live video interviews.

What it actually does: real-time analysis of gaze patterns, speech cadence, and lexical signals during live interviews. Flags behavioral anomalies consistent with reading from a script. Also scores AI fluency — useful if you want to understand a candidate’s genuine relationship with AI tools rather than just detect misuse.

What it doesn’t do: it won’t catch deepfakes without additional liveness verification. And behavioral analysis flags anomalies — it doesn’t tell you definitively that fraud occurred. It’s a signal layer, not a verdict.

Best for: teams doing significant live interview volume across any role type who want a behavioral signal layer without enterprise pricing.

Tier 3: Enterprise — 200+ hires per year
(custom pricing)

FabricHQ — end-to-end AI interview platform with detection

Custom pricing, demo required.

What it actually does: the most comprehensive detection layer available as of 2026. Analyzes 20+ behavioral and technical signals simultaneously. Adaptive AI interviewer that adjusts based on candidate responses. Detection capabilities cover invisible overlay tools — the platform identified behavioral patterns consistent with Cluely and Interview Coder use in 85%+ of confirmed fraud cases in their dataset.

The honest caveat: this is a full platform replacement, not an add-on. Implementing it means changing your interview workflow, not adding a tool to an existing one. That’s a real organizational change, not just a budget question.

Best for: companies doing high-volume technical hiring where fraud detection is a genuine operational priority, not just a concern.

Greenhouse with Real Talent fraud detection

Greenhouse SMB pricing averages $31,198 per year. Enterprise pricing averages $111,300 per year. The fraud detection feature (Real Talent) is only available on Pro — the most expensive tier.

What it actually does: application-stage fraud signals — IP analysis, automated application pattern detection, identity flagging. This catches mass-apply agents and suspicious application patterns before a human recruiter reviews anything.

The honest framing: if you’re already on Greenhouse and haven’t unlocked Real Talent, it’s worth evaluating. If you’re choosing an ATS primarily for fraud detection, this is expensive for that single use case. The value is in having fraud detection integrated into a platform you’re already using, not in the detection capability itself which is narrower than dedicated tools.

Endorsed and Checkr — identity verification

Both are custom enterprise pricing with no public rates. Endorsed focuses on application-stage identity fraud signals. Checkr focuses on post-offer background verification.

These solve different problems: Endorsed is about flagging suspicious applications before you spend interview time on them. Checkr is about verifying the person you’re about to hire is who they say they are. For roles with access to sensitive systems, financial data, or for remote mid-to-senior positions — Checkr or an equivalent is increasingly standard practice, not an optional add-on.

What this all costs in practice:

Let’s discuss three team scenarios.

20-person startup, 15 technical hires per year, no current tooling: CoderPad Starter ($840/year) + process redesign (free) + one specific product question in application form (free). Total: under $1,000/year.
This addresses the most common fraud vectors without enterprise complexity.

150-person company, 80 hires per year across technical and non-technical roles: HackerRank Pro ($5,400/year) for technical + TestGorilla Core ($1,620/year) for non-technical + Sherlock for live behavioral detection (volume-dependent, estimate $2,000-4,000/year). Total: $9,000–11,000/year.
ATS integration included in HackerRank Pro.

500-person company, 300+ hires per year, security-sensitive roles: FabricHQ (custom, estimate $30,000–60,000/year) + Endorsed for application fraud (custom) + Checkr for background verification (per-check, estimate $15,000–25,000/year at volume). Total: $50,000–85,000/year.
At this scale, one prevented fraudulent senior hire at $50,000+ in direct losses justifies the investment without much calculation.

How to choose: decision framework before you spend any $

Before evaluating any tool, answer four questions. The answers tell you which tier you actually need — and sometimes tell you that you need none of it yet.

Question 1: What are you trying to catch?

• Application-stage fraud (mass automated applications, fake identities) → Knockout questions first, Endorsed or Greenhouse Real Talent if volume justifies it.

• Take-home fraud (AI-generated submissions) → Live defense protocol first. No tool required. If you want timing metadata and code replay, CoderPad or HackerRank.

• Live interview fraud (invisible overlays, earpieces) → Process redesign first (follow-up questions, constraint shifts). Sherlock or FabricHQ if you want a behavioral signal layer on top.

• Post-hire substitution (someone else doing the work after joining) → Diagnostic onboarding and 90-day validation. No tool addresses this at the interview stage.

If you can’t answer this question clearly, you’re not ready to evaluate tools. Start with a process audit.

Question 2: What’s your hiring volume?

• Under 30 technical interviews per month: CoderPad Starter or free tier covers you. Don’t buy enterprise.

• 30-100 interviews per month: HackerRank Pro or equivalent. The per-attempt economics work at this volume.

• 100+ interviews per month: custom enterprise conversations make sense. The per-seat and per-attempt pricing at volume justifies dedicated contracts.

• Non-technical roles at any volume: TestGorilla Core as a pre-screen filter. Sherlock for live detection if behavioral fraud is a concern.

Question 3: What does your current process look like?

• If interviews are still static question-and-answer with no follow-up, no constraint shifts, and take-homes with no defense — fix the process first. Adding a detection tool to a shallow process catches some fraud but misses the structural problem. You’ll still hire people who can’t perform the role.

• If your process already has depth — follow-ups, live defense, behavioral drills — then a detection layer genuinely adds signal on top of something that works.

Question 4: What’s the realistic cost of fraud in your context?

• Engineering roles with system access, senior positions, remote-first hiring in security-sensitive industries: the cost of one fraudulent hire ($50,000+ in direct losses, not counting security exposure) justifies significant tooling investment.

• Marketing coordinator, operations analyst, customer success: the fraud risk is real but the consequence of a bad hire is recoverable. Process changes and light tooling are proportionate.

Match your tooling investment to your actual risk profile. Spending $30,000/year on enterprise fraud detection for a 20-person team hiring two engineers per quarter is not a proportionate response.

Implementation sequencing

If you’re starting from zero, this is the order that makes sense:

Month 1: Process audit. Map every stage, identify where you have no follow-up, no defense, no identity verification. This is free and is the foundation everything else builds on.

Month 1-2: Free changes. Add one product-specific question to your application form. Introduce mandatory follow-up questions as a standard for all interviewers. Add live defense to every take-home. Document what each interview stage is actually testing.

Month 2-3: Light tooling. CoderPad or equivalent for technical interviews if you’re doing more than a handful per month. Sherlock for live detection if behavioral fraud is a concern across roles.

Month 3-6: Evaluate whether you need more. Run your 90-day post-hire validation. If the free changes and light tooling are catching issues and your hire quality is improving — you may not need to go further. If you’re still seeing problems, the data from three months of structured hiring will tell you exactly where.

Enterprise tools: only after you’ve exhausted the above and have clear evidence of a gap that paid tooling closes.

Honest conclusion

The tools exist and the pricing, at mid-market tier, is manageable. But the pattern across every team that handles this well is consistent: tooling amplifies a well-designed process. It doesn’t replace one.

If your interviews are still one-question-one-answer with no follow-up, a detection platform won’t fix that. If take-homes have no live defense, plagiarism detection is catching the wrong thing. The process changes in Parts 1 through 3 are prerequisites — not alternatives — to anything listed above.

Sometimes the right answer is a $70/month subscription. Sometimes it’s a process redesign that costs nothing and solves the actual problem. The tool question is secondary to the process question.

If you’d like to think through what your hiring process actually needs before making any purchasing decisions — that’s a different conversation, one about organizational structure and hiring design rather than software. That’s the work we do at RIST. Reach me directly if it’s relevant.

Previous parts in this series:

The stages below are where most hiring teams feel most confident. Live interviews. Defended assessments. Reference calls. The logic being: we’re past the automated filters now, we’re talking to real people, we can tell who knows their work.

That confidence is the problem.

This is precisely where AI assistance is most sophisticated, most invisible, and most economically rational for candidates to use. A $30 monthly subscription to a cheating tool is a negligible cost against a $150,000 engineering offer. The math is obvious. And the tooling has caught up with the math.

Before the stage breakdown — the full picture of what’s actually happening.

The fraud landscape in 2026: it’s not one thing

Most conversations about AI fraud collapse everything into one scenario: candidate reads answers off a second screen. That’s real, but it’s the least sophisticated version of what’s happening.

There are four distinct types, and they require different responses.

Type 1: AI-assisted answers during interviews. The candidate is present and real, but receiving generated responses in real time through an invisible overlay or earpiece. This is the most common type and the one most detection software is built to catch. It’s also detectable with one well-placed follow-up question — no software required.

Type 2: Proxy interviews. Someone else sits the interview entirely. The actual candidate never appears. In the Infosys case from early 2025, an engineering graduate had a friend impersonate him for a video interview, got hired, and was fired within two weeks when his actual work made the substitution obvious — followed by criminal proceedings.

Type 3: Deepfake video and audio. The candidate appears on camera, but the face and voice are synthetically generated. Palo Alto Networks’ Unit 42 demonstrated that a researcher with no image manipulation experience built a convincing synthetic identity for a video interview in 70 minutes on a consumer-grade computer. In 2025, an Indian IT company discovered that three recent hires had used deepfake video throughout their interviews. The fraud surfaced months later during performance reviews.

Type 4: Post-hire substitution. The most dangerous type, and the one no interview proctoring catches. The candidate passes every stage legitimately — and then, after joining, the actual work is done remotely by someone else. This is the mechanism behind the North Korea IT worker scheme. By February 2026, ID.me had suspended more than 130 wallets linked to suspected state-affiliated actors, with wallet creation attempts increasing 200% between March and November 2025.

Type 4 is an organizational security issue, not a hiring process issue. It requires onboarding validation and 90-day performance tracking, not better interview questions.

Stage 5: Assessments and case studies

We covered this in Part 2 with the Anthropic case. One addition relevant at this level.

The prisoner’s dilemma dynamic that’s driving cheating on take-home assignments isn’t primarily about ethics — it’s about perceived competitive pressure. Candidates cheat because they believe their competition already is, and they’re not willing to be the only honest participant in what feels like an unfair process. This is an organizational signal, not just an individual one.

If your assessment format is so standardized that candidates assume AI can handle it — they’re probably right. The solution isn’t more surveillance on the same format. It’s changing what the format actually tests.

The organizational decision: no assessment without live defense. Thirty minutes where the candidate explains their decisions and you change one constraint mid-conversation. This applies to every function, not just technical roles. A product manager defends their PRD. A finance lead defends their model. An operations candidate defends their plan. AI can assist with preparation. Defending reasoning you don’t own — that’s what falls apart in the room.

Stage 6: Behavioral live interviews

Tools like Cluely and Final Round AI transcribe the interviewer’s question and generate a STAR-format response delivered through an invisible overlay or earpiece in real time. This works for every function — HR, sales, product, operations. Behavioral rounds almost never have any form of proctoring.

The organizational implication: your behavioral interview process is producing data that looks like signal but may be noise at scale.

What actually works isn’t a detection tool — it’s interview design. A five-level behavioral drill instead of a standard situational question: facts, reasoning, personal difficulty, reflection, negative consequence. AI produces polished stories. It doesn’t hold up past the third level of follow-up on a specific moment.

Dynamic constraint shifts during the interview do more: “You described that situation — what would you have done if the budget had been cut in half two weeks in?” Scripted answers break on unexpected variables. Genuine experience doesn’t.

The organizational decision: standardize the drill across all interviewers. This isn’t just anti-fraud practice — it’s better interviewing regardless. Inconsistent interview structure is one of the primary sources of hiring bias, independent of AI. Fixing it serves both goals simultaneously.

Stage 7: Technical and coding live interviews

This is where the numbers are hardest to ignore.

FabricHQ’s analysis of 19,368 interviews found that approximately half of tech candidates show signs of AI assistance during live coding. The breakdown of methods matters for understanding the scale of the problem: dedicated cheating tools like Cluely and Interview Coder account for 45% of cases. Voice mode on LLMs — ChatGPT or Claude running in the background — accounts for another 34%. Traditional methods like second screens or tab switching are now a minority at 18%. The top two methods combined — 79% of cases — are specifically designed to be invisible to screen sharing.

The technical reason: these tools use low-level graphics hooks, DirectX on Windows and Metal on macOS, to render a transparent overlay that exists only on the candidate’s local display. When the candidate shares their screen through Zoom or Google Meet, the conferencing software captures what’s beneath the overlay. The overlay itself is invisible to the recording. Screen sharing as a verification mechanism is structurally obsolete against these tools.

Interview Coder 2.0, released in January 2026, added real-time audio processing — the tool now transcribes the interviewer’s spoken question without any manual input from the candidate, generating a response before they’ve finished asking. The company publicly states zero documented cases of detection when the tool is used correctly.

This is not a niche product. It’s a SaaS business with versioning, feature updates, and customer support.

What actually works: live debugging instead of building from scratch. Give a broken piece of code with a non-standard bug. Ask the candidate to find it and explain it line by line. The research shows that AI tools perform significantly worse on debugging tasks and system design with contextual constraints than on standard coding problems — the outputs are often inaccurate when the context is specific enough. A unique bug in someone else’s code is exactly that context.

System design with constraint shifts mid-conversation: 10x load increase, latency cap, budget cut — introduced after the initial design is underway. Add a deliberate error or contradiction in the problem statement and observe whether the candidate catches it.

The organizational decision: if your technical interview format consists primarily of problems that exist verbatim on LeetCode or similar platforms, you are testing preparation and AI access, not engineering capability. Rebuilding the format is a one-time investment. The cost of a fraudulent senior technical hire — FabricHQ estimates $50,000+ in direct losses alone — makes the case without much calculation.

Stage 8: Reference checks

Reference fraud has emerged as a distinct category. AI helps candidates script detailed, credible answers for their references, generate plausible accounts of shared work, and in some cases create entirely synthetic contacts with fabricated employment histories.

The reference check serves a specific function that’s worth being precise about. It is not a primary selection tool. By the time you’re running references, the hiring decision should already be substantively made. What reference checks do: verify the factual record the candidate provided, surface serious concerns that prior rounds didn’t reveal, and occasionally confirm what you already believe. If a reference check is regularly changing your hiring decisions, the earlier stages aren’t providing enough information.

What works at this stage: one unexpected question for the reference — “What’s the most significant mistake they made while working with you?” A scripted reference isn’t prepared for that. A genuine one answers immediately. Verify that the reference is actually listed as a colleague during the relevant period on LinkedIn before the call. For senior roles, back-channel references — people who worked with the candidate that you find independently, not through the candidate’s list — provide substantially better signal.

A note on legal compliance: back-channel references are not permitted in all jurisdictions. Employment law varies significantly by country and region on what you can ask, who you can contact, and how you can use the information. Verify local requirements before building this into your standard process.

Background checks — distinct from reference checks — are mandatory or strongly advisable for roles involving access to financial data, personal data, sensitive systems, or vulnerable populations. The specific requirements vary by jurisdiction and role type. For remote mid-to-senior roles with system access, identity verification before signing is no longer optional risk management. Checkr data indicates 23% of companies have already encountered identity fraud among new hires. Discovering it after signing creates legal, operational, and security complications that discovering it before signing does not.

Stage 9: Offer and post-decision

Fraud at this stage is less common but worth naming.

AI-generated counter-offer arguments have become more sophisticated. A well-structured salary negotiation email with market data and logical framing can be produced in minutes. The issue isn’t the counter itself — negotiation is normal — it’s that the quality of the written argument no longer reliably signals the quality of the candidate’s market knowledge or self-awareness. Respond to the substance: “You’ve benchmarked yourself against senior-level compensation — walk me through what in your experience maps to that level” is a better response than a counter-number.

For candidates who receive rejections: one personal, specific sentence about why the role genuinely interested you is more memorable and more effective than a three-paragraph AI-generated appeal. Recruiters see the pattern immediately and the well-crafted letter produces no impression at all.
My favorite meme about “uno reverse rejection”:

What to do when you don’t have budget for tools

The most important section — and it’s not about technology.

In 2026, a number of organizations are moving back toward human-led assessment after purely automated pipelines. Not because the tools don’t work, but because the realization has set in: the deepest protection is process quality, not detection capability.

Verify identity earlier. Not after the offer — at the first screen. Camera on, candidate introduces themselves, one unexpected question in the first two minutes. This takes no time and eliminates most proxy fraud without any tooling.

Introduce 90-day post-hire validation. Document what the interview was testing. At 90 days, check whether it holds. If a new hire consistently needs support in areas where they demonstrated capability during the process, that’s a signal to review the process, not the person. This is the only mechanism that closes the loop between what interviews measure and what the role actually requires.

Make the first month of onboarding diagnostic. Real tasks with real deadlines in the first two weeks, not just orientation. Post-hire substitution — Type 4 fraud — surfaces here, not in interviews.

Think about process depth, not surveillance. Monitoring and proctoring detect symptoms. Behavioral depth, live defense, and constraint shifts test the actual capability — and make fraud structurally pointless.

Conclusion

AI will not break your hiring process.
A shallow process is already broken.

If interviews can be passed without thinking out loud, if assessments don’t require defense, if references are a formality, and if feedback is vague — AI will outperform your entire process. That’s not a prediction. For a growing number of candidates, it’s already the current state.

2026 is not about fighting AI. It’s about depth. Organizations that build processes where thinking cannot be replaced by generation will hire better. Those that maintain templates will not. This is true on both sides — for hiring teams, and for candidates who believe preparation can be outsourced entirely.

See you next week for Part 4 ;)

One finding sits above everything else in this research:

38.5% of interviews show signs of AI assistance — but most companies have no detection software at all. What actually catches people isn’t an algorithm. It’s one unexpected follow-up question they can’t answer. Everything else is downstream of that.

Stage 1: Job posting and application volume

Application volume is up 51% since AI tools went mainstream. Not because more people are job hunting — because one person can now apply to hundreds of roles in an hour through automated agents. Your recruiters are drowning before they open a single resume.

The tools that address this at scale: Endorsed ($15-30K+/year), Greenhouse Real Talent fraud detection (Pro tier only, $20-50K+/year for mid-size teams) — exist and work.
But most hiring teams don’t have them. The realistic picture for the majority of companies: knockout questions in your ATS form and manual review. That’s it.

Which means the practical lever most teams can pull immediately costs nothing: one specific question about your product in the application form. An automated agent doesn’t know your product. A human who’s done the research does.

Stage 2: Resume screening

There’s a structural contradiction in how most teams approach this stage that almost nobody talks about.

A new study (arXiv, August 2025 — 2,245 resumes, 24 roles) found that LLMs consistently prefer resumes written by the same model that’s doing the screening — by 67-82%. GPT-4o showed 82% self-preference bias. The shortlisting advantage for candidates whose resume matched the evaluator model ranged from 23% in some roles to 60% in sales and accounting. Equip Blog

The implication: if your screening tool is GPT-4 and a candidate wrote their resume with GPT-4, that resume will rank higher than an equally qualified human-written one — purely due to stylistic mirroring. You’re not selecting for the best candidate. You’re selecting for the candidate who uses the same tool you do.

The fix the researchers identified reduces bias by over 50%: don’t rely on a single model for screening.
Combine evaluators, or — more practically — keep a human in the loop on final shortlisting decisions. AI can improve efficiency, but it cannot fully replace the human ability to identify potential, authenticity, and contextual fit.

One more thing worth naming:
the “ATS filters everything” belief that candidates optimize around is largely a myth. Modern ATS systems mostly parse and store. The actual screening happens either manually or through LLM scoring. Candidates are optimizing for a system that doesn’t work the way they think it does.

Stage 3: Phone and video screening

This stage feels safe to most recruiters. You’re talking to a live person. But tools like Final Round AI and Cluely transcribe the interviewer’s question and generate a response through an earpiece or invisible screen overlay in real time. If you don’t go deeper, it’s undetectable.

Most teams catch this through observation, not software. Micro-pauses before every answer. Monotone delivery without natural variation. Perfectly structured sentences with no thinking out loud. Experienced recruiters notice — they often can’t articulate why, but they notice.

What actually works, costs nothing, and makes every interview better regardless of fraud: one unexpected follow-up on every answer.
Not “are you cheating” — just “you mentioned X, where did that not work?” If someone is reading a generated response, there’s no prepared line for that. If they lived it, they answer in seconds.

Make this a standard part of every interview, not an exception. It improves the quality of information you get from every candidate, not just the ones using AI.

Stage 4: Automated filters and instant rejections

Candidates get rejected in 30 seconds and assume the ATS is broken or biased against them. Usually it’s neither.

Instant rejections have two real causes.
The first is knockout questions — location, work authorization, salary expectations, relocation. A “no” on a hard requirement triggers an automatic rejection. This is legitimate, transparent, and built into every ATS for free.
The second cause is where it gets complicated: AI-based screening that filters on criteria unrelated to job performance.

This is a legal exposure, not just an ethical one.
The Mobley v. Workday case was expanded in 2025 — an AI screening tool can be treated as an agent of the employer and held liable for discriminatory outcomes. California has finalized regulations applying anti-discrimination law to AI in hiring. New York already requires annual bias audits for automated employment tools.

If you’re using automated filters: document your criteria, tie them explicitly to job requirements, and review them regularly. In 2026 this isn’t optional risk management — it’s legal hygiene in a growing number of jurisdictions.

Stage 5: Take-home assignments

This is the most broken format in hiring right now.

In January 2026, Anthropic rewrote their technical interview questions.
The reason: candidates were using Claude to cheat on their coding assessments. The company that built one of the most widely used AI tools had to change its own process because of that same tool. If that’s not a precise signal of where things stand, nothing is.

FabricHQ data shows cheating on take-home assignments doubled from 15% in June 2025 to 35% in December 2025 — in six months. Part of this is tooling.
Part of it is what researchers call the prisoner’s dilemma: candidates cheat because they believe their competition already is, and they don’t want to be the only honest participant in an unfair race.

The core problem is structural. Take-home assignments happen in an unobserved environment. Timing metadata and code replay provide some signals — but tools now simulate human typing speed with artificial delays between keystrokes. Even the technical detection signals are degrading.

The solution is straightforward, though it requires changing the process: no take-home without a live defense. Thirty minutes only. The candidate explains their decisions, you change one constraint mid-conversation. AI can be used to prepare. Defending someone else’s solution without understanding it — that’s what falls apart in the room.

This applies to every role, not just technical ones. A product manager defends their PRD. A marketer defends their campaign. An analyst defends their model. The format is the same.

What connects all five stages

The pattern across every stage is the same: AI assistance fails when the environment becomes dynamic and specific. Static questions, standardized formats, unobserved environments — these are where fraud thrives. One unexpected follow-up, one changed constraint, one question that requires genuine context — that’s where it breaks down.

The implication for process design isn’t “add more surveillance.” It’s “make your process more conversational and less predictable.” That’s better hiring practice whether or not fraud is present.

Part 3 covers the second half of the funnel: live interviews, case studies, reference checks, and what happens after the offer. Including one case that surprised me more than anything else in this research.

Subscribe to get it when it’s out.

Subscribe now

That was the tipping point for a lot of teams. What had been an inconvenience became an operational problem. And it’s accelerating.

I’ve spent the last several months researching AI fraud in recruiting — not from a theoretical angle, but from the perspective of hiring teams that are actively dealing with it. This is part one of a four-part series. Today: the detection tactics companies are deploying right now, and why some work better than others.

The data:

• 74% of hiring managers have encountered AI-generated content in applications
  Resume Genius / FNBO, 2025

• 91% of recruiters have detected candidate misrepresentation
  Greenhouse Study, 2025

• 62% of hiring professionals say candidates are now better at faking than HR teams are at detecting
  Checkr, 3,000 managers

• 1 in 4 candidate profiles will be fake by 2028
  Gartner forecast

  ---

Why this happened — and why it’s not slowing down

The mechanics are straightforward. Application volume is up 51% since AI tools became mainstream (Fast Company, 2025). Response rates on job boards hover around 3.1% for most roles (huntr.co, 2026). When the odds are that low, candidates optimize — and AI makes optimization near-effortless.

What started as keyword-stuffed resumes and AI-polished cover letters has evolved into something more structured: automated agents applying to hundreds of roles simultaneously, real-time AI assistants feeding answers during live interviews, and in the most extreme cases, deepfake video to impersonate a different person entirely.

Case in point: Amazon + North Korea

In December 2025, Amazon’s Chief Security Officer Stephen Schmidt disclosed that the company had blocked over 1,800 job applications from suspected North Korean state-affiliated actors since April 2024 — with attempts increasing 27% quarter over quarter. The method: stolen or synthetic identities, AI-generated resumes, deepfake video in interviews, and post-hire data exfiltration. This is the same technical infrastructure used by ordinary applicants, deployed by a state-level threat actor. The Register.

The implication for hiring teams isn’t that every candidate is a state actor. It’s that the tooling is now commoditized — and your detection methods were built for a different threat environment.

What companies are actually doing

There’s a meaningful gap between what companies say they do and what actually moves the needle. Here’s what the evidence shows, with an honest assessment of effectiveness for each.

🍌 The “banana trap” — hidden instructions in job postings

Detection method · Low cost · High signal

The most viral example: a company drowning in automated applications added a single line to their job posting — “If you are a human reading this, start your cover letter with the word BANANA.” AI agents processing the posting at scale skipped it. Human applicants didn’t.

What started as a one-off workaround has become a category. Variations now include instructions buried in the second paragraph, embedded in small print at the footer, or phrased as casual asides. Some companies use unexpected specific questions (”What did you notice about our checkout flow that most people miss?”) — anything that requires genuine engagement with the posting, not just parsing it. More on this tactic.

The underlying logic is sound: it creates a low-cost filter that exploits the difference between reading and processing. The limitation is that it’s becoming known — which means more sophisticated agents will eventually be trained to handle it.

→ What to consider: effective now, but needs rotation. Treat it like a CAPTCHA — the moment the answer becomes predictable, the signal degrades.

⬜ Prompt injection — in both directions

Detection method · ATS-level · Mixed effectiveness

Candidates have been hiding white-text keywords in resumes since 2024 to manipulate ATS ranking. ManpowerGroup detects this in roughly 10% of resumes it scans. Greenhouse, processing around 300 million resumes annually, puts the number closer to 1%. The discrepancy suggests most attempts are crude and the detection isn’t the hard part — the problem is that it still works often enough to keep happening.

The more interesting development is the reverse: according to OWASP’s 2025 report, some companies now embed hidden prompt instructions in job descriptions — essentially injecting instructions that tell an AI processing the application to flag it as AI-generated. Built In.

→ What to consider: useful as a passive signal layer, but not reliable as a primary filter. An AI agent sophisticated enough to auto-apply at scale is usually sophisticated enough to avoid crude white-texting.

📹 Liveness verification — the “turn your head” test

Detection method · Video interviews · High effectiveness vs deepfakes

Google and McKinsey reintroduced mandatory in-person rounds for sensitive roles in 2025 specifically to counter this. For teams that can’t do that, a simpler approach has emerged: spontaneous physical requests during video calls. “Hold up your ID,” “raise your hand,” “turn your head left.” Current real-time deepfake filters struggle with sudden, unpredictable movement — the lip-sync and facial rendering breaks down.

Pindrop, a voice security company, used exactly this to catch a candidate they called “Ivan X” during an engineering interview. Resume strong. Answers confident. But micro-expressions were out of sync, the IP address pointed to a different country, and when an unexpected technical question came in — a pause that revealed processing latency. Eight days later, Ivan X reapplied through a different recruiter with a visually different face but the same credentials. Full Pindrop case study.

→ What to consider: highly effective against current deepfake tooling. Build spontaneous verification into your standard video interview protocol — not as a one-off, but as a consistent step.

🔀 Context-specific application questions

Detection method · Application form · High signal for all roles

This is the tactic with the highest signal-to-noise ratio and the lowest implementation cost. The logic: add one question to your application form that requires genuine, specific engagement with your product, company, or role — something a generic AI agent can’t answer from the job description alone.

Examples that work: “What’s one thing about our onboarding flow that you’d change after using it for 10 minutes?” or “Which of our recent product decisions do you disagree with and why?” A generic LLM produces a generic answer. A person who’s actually done the research doesn’t.

Anthropic made this public in 2025 — requiring cover letters written without AI assistance to assess “unassisted communication skills.” It’s now being adopted more broadly as a standard application layer, not just a compliance statement.

→ What to consider: works across all role types, not just technical. Rotate the question per role and per hiring cycle to prevent it from becoming predictable. The goal is genuine engagement, not a trick.

The bigger pattern

Each of these tactics works on the same principle: create a moment that requires genuine human engagement — something that can’t be automated away at the application layer. The banana trap requires reading. Liveness verification requires being physically present. Context questions require having actually used the product.

None of them are perfect. All of them degrade over time as the tooling catches up. The teams handling this well aren’t looking for a single solution — they’re building layered pipelines where each stage catches what the previous one misses.

What’s coming in Part 2

The application stage is one layer. The deeper problem is what happens further down the funnel — in phone screens, take-homes, behavioral rounds, and reference checks. Each stage has its own vulnerability profile. Part 2 breaks down the full funnel, stage by stage, with the detection approaches that have evidence behind them.

I write about hiring strategy, process design, and what the data actually says — for people who build and run recruiting functions. No sponsored content.

Let’s ask the harder question: are we even measuring the right things?

I’ve run recruiting audits across dozens of companies. The picture is almost always the same. Metrics exist. Dashboards exist. Reports go out every week. But decisions get made based on numbers that are either incomplete, misleading, or completely disconnected from what the business actually needs.

And that gap has a price tag.

If you want to skip straight to the data — the full benchmark report with formulas, benchmarks and regional breakdown is linked at the end of this article.

Subscribe now

The market shifted. The metrics didn’t.

Hiring is growing again — up 8% after two years of decline. Sounds like good news. Even Gini is reporting increases in job openings and hires across most categories this year.

But the baseline is still 30% below 2021. Teams are 14% smaller. Recruiters are handling 93% more applications. And actual hires per recruiter are down 43%.

More work, fewer people, less output. This isn’t a productivity crisis. It’s a signal that something is broken in how we measure effectiveness — and in how we’re implementing AI without knowing what we’re actually trying to improve.

What stopped telling the truth

1. Resume Screen Volume. AI screens entire pipelines in seconds. This metric no longer says anything about effort or quality of work. We need something that measures outcomes, not activity.

2. Outreach Volume — calls and messages. I’m old enough to remember writing every message manually, and I know exactly why this metric mattered fifteen years ago. But when a tool sends 500 messages on your behalf, “how many did you send” no longer equals “how much useful work did you do.”

3. Application volume as a pipeline quality indicator. GenAI lets candidates apply to hundreds of roles in one click. In 2025, recruiters receive 93% more CVs than four years ago. Quality hasn’t grown proportionally. A bigger flow is not a better pipeline. And we’re already seeing what Gartner predicted: by 2028, one in four candidate profiles will be fake.

4. Time-to-Fill without context. Closing a role in 15 days sounds great. But if that person left on day 60, was it a success? Speed without what happened next is just a number.

5. Candidate Experience as a checkbox. This one goes deeper than whether the recruiter replied on time. It’s about whether the candidate felt the process was fair. Was it easy to get to the interview? Did the information available online actually reflect the company? Did the candidate understand how the decision was made? If we’re not asking this systematically, we don’t know where we’re losing people — or why.

What strong teams actually measure

1. Employer Brand Influence on Offer Acceptance. The correlation between your Glassdoor or LinkedIn rating and how often candidates accept offers in specific roles. According to MSH 2026, strong employer branding drives a 50% reduction in cost-per-hire. This isn’t about reputation anymore. It’s a direct financial lever.

2. Quality of Hire + Skills Match Rate. Everyone has been talking about QoH since 2025. And yet only 28% of companies officially measure it (Gem, SHRM). But even QoH alone is incomplete. It needs to sit alongside Skills Match Rate: the percentage of candidates who passed a skills assessment and actually meet the role requirements. Critical anywhere skills-based hiring is replacing credential screening — which is already happening.

3. 90-day retention separately from annual. McKinsey found that 18% of new hires leave during or immediately after probation. If you’re only looking at annual retention, you won’t see this problem until the next reporting cycle. That’s often a full year of lost time.
   Early attrition is also rarely random. It usually points to one of three things: a hiring decision that prioritised speed over fit, an onboarding process that didn't set the person up properly, or a gap between what was sold in the interview and what the role actually looked like on day one. All three are measurable. All three are fixable. But only if you're tracking the right moment — not waiting until the annual review to notice the pattern.

4. Talent Rediscovery Rate. How many of your hires came from your own database. In 2025 that’s already 46% of sourced hires — up from 26% in 2021. Your existing base is your best sourcing channel. Most teams are still paying for new contacts instead of going back to the ones they already have.

5. Pipeline Velocity by stage. Not overall TTF, but where exactly the process is getting stuck. Where candidates wait longest. Where the most drop-off happens. This lets you fix the specific problem instead of optimizing the entire funnel when only one stage needs attention.

6. Internal Mobility Rate. LinkedIn calls this one of the top TA priorities right now. Lower cost-per-hire, higher retention, stronger culture — all true. But I wouldn’t put it at the top of the list. It matters, but it’s secondary to what’s happening inside the recruiting function itself.

7. Recruiter Augmentation Index. This is, in my view, the most important metric right now — and almost no one is tracking it. How much of a recruiter’s time goes to high-value activities: conversations, advisory, closing — versus admin tasks? SHRM reports that AI is now used in 43% of HR tasks, up from 26% in 2024. But the question isn’t how much AI you’ve adopted. The question is whether it’s actually freeing up human time for the work only a human can do. If not, you’ve just added new tools on top of old processes.

What this costs — a real example

A client came to me with a clear goal: Time-to-Fill of 17 working days across all roles.

Sounds specific. Feels measurable. But the audit told a different story.

Looking at 14 months of data across the full recruiting cycle, the actual TTF ranged from 8 days for a Business Analyst to 41 days for a .NET Engineer. A single target across all roles isn’t a strategy. It’s an illusion of control.

But the more interesting finding was in the pipeline conversion data. Offer acceptance was strong — 83%. Meaning when candidates got to an offer, they took it. The loss was happening earlier: between HR interview and final decision. For Java Developer roles, stage conversion dropped to 9%. For AI Architect, 25%.

The problem wasn’t the recruiters. It wasn’t the market. It was that Sales was passing incomplete briefs to recruiting. Recruiters were starting searches based on guesses. Technical interviewers had no aligned criteria. And nobody had baseline data to even see where time was being lost.

When the client said they wanted to cut recruiting operations by 20% through AI — the first question wasn’t which tool to implement. It was: do you know your Pipeline Velocity baseline right now? Because without it, there’s nothing to measure the 20% against.

That’s almost always where it starts.

Where to start

Not with implementing everything at once. That’s the most common mistake I see.

Step 1. List what you’re actually tracking now. Everything in your weekly and monthly reports. No judgment, just inventory.

Step 2. Check if you have a baseline. For each metric: do you know what the number was six months ago? A year ago? If not, it’s not a metric. It’s a snapshot.

Step 3. Find where the business can’t get answers. Where do your stakeholders ask questions you can’t answer with data? Start there. Not with what’s trendy to measure, but with the actual gap between what’s being asked and what you have.

Step 4. Add one metric at a time. One new metric, with a clear definition, formula, and owner. Otherwise in a month it’s just another column in a spreadsheet nobody reads.

Why this matters

We’re at a point where the tools have changed but the metrics haven’t.

AI has transformed parts of the recruiting process beyond recognition. But if we’re still reporting on the same KPIs, we can’t see whether that transformation is creating value — or just creating more movement without more output.

Recruiting has always been about people. But to demonstrate the value of that work at the business level, you need a language of numbers that reflects real impact. Not activity. Impact.

So — which of these metrics are you already tracking? What are you trying to implement but not sure where to start? And is there something you’re measuring that I didn’t mention?

Drop it in the comments. Genuinely curious what the real picture looks like across different teams.

Leave a comment

You can find the full benchmark report with formulas, benchmarks and regional breakdown here.

"$15,000 referral bonus — is that a lot, a little, or does no one actually know where that number comes from?"
The first reaction is usually "wow." The second is "maybe it's fair?" The third is "do I even know what my own hire costs?"

A lot gets written about referral programs: what share of hires they produce, that conversion is higher, retention is better, candidates are “higher quality.” Almost nobody shows the economics. So let’s do the math.

What does hiring actually cost in Ukrainian IT?

Before talking about the referral bonus, you need to know what hiring through standard channels actually costs. Here are real market figures.

Full cost of a single hire: real numbers

Two typical scenarios — Middle Backend Developer ($2,500/mo) vs Senior ($4,500/mo).

CAC (Cost to Acquire a Candidate) — borrowed from growth: how much it costs to bring one person in. Same logic, applied to hiring.

What this means for your bonus. Middle: standard CAC = $3,660 → a bonus of $1,000–1,800 is fully justified and you still save $2,000+. Senior: CAC = $8,280 → a bonus of $2,500–3,000 saves you $5,000+ per hire.

How to determine the right Referral Bonus

There are three rational ways to calculate it. Not “this is what the market does” — but based on financial logic.

1. Based on Your Real Cost of Hiring (Most Accurate)

First, calculate how much it costs you to close the role without a referral — via recruiter, job platforms, or agency.
That’s your standard CAC (Cost of Acquisition).

Then take 30–50% of that number.

Why 30–50%?
Because a referral replaces an expensive sourcing channel — but it doesn’t eliminate all costs. You still spend time on interviews, recruiter effort, onboarding, and you still carry hiring risk. So you don’t share the full savings — typically one-third to one-half.

Logic: a referral saves you money — you share part of that saving.

Example:

• Middle Dev: CAC $3,660 → bonus ~$1,400

• Senior Dev: CAC $8,280 → bonus ~$3,000

2. Based on Monthly Salary

If you don’t have a precise CAC, anchor to salary:

• Junior — 0.5 monthly salary

• Middle — 0.75

• Senior — 1.0

• Lead — 1.5

Example:

• Middle at $2,500 → ~$1,875

• Senior at $4,500 → ~$4,500

Simple method — but sometimes overpriced.

3. Compared to Agency Fees

Agencies typically charge 15–20% of annual salary.
A referral bonus logically sits at 3–5% of annual salary.

Example:
Senior at $4,500/month:
$4,500 × 12 × 4% ≈ $2,160

Logic: referral must be significantly cheaper than an agency — but valuable enough to motivate people.

When and how to pay: 3 common structures

Split model (most common):

• 50% after the candidate starts

• 50% after probation (2–3 months)

Probation-based payout:
100% after successful probation.

3-month split:
33% / 33% / 34% monthly — used in long onboarding cycles.

Always document:

• payout conditions

• what happens if the candidate leaves early

• who can refer (internal only or external as well)

• maximum bonus per person per quarter

The real question: ROI

The bonus is not the main variable.
The real question is: when does the hire pay back?

Referral hiring is cheaper than agency hiring.
But cheaper ≠ automatically profitable.

ROI doesn’t depend on the bonus.
ROI depends on the value the person generates.

So instead of asking “Is the bonus expensive?”
Ask: What is the full hiring economics?

Full hiring cost formula

Most companies calculate only CAC. That’s incomplete.
Costs don’t stop when the offer is signed.

Total Hiring Cost = CAC + (Salary × burden × ramp-up months × 0.5)

Step 1 — Monthly Real Cost

Monthly cost = Salary × burden
(Burden reflects taxes and benefits — e.g., 1.22 for contractors, 1.47 for full employment.)

Step 2 — Ramp-Up Cost

Ramp-up cost = Monthly cost × months × 0.5
(0.5 assumes ~50% productivity during ramp-up.)

Step 3 — Add Acquisition Cost

Total Hiring Cost = CAC + Ramp-up Cost
This shows not just “cost to close a vacancy” — but the full investment until full productivity.

Real client example: Middle Developer

One client hired a Middle Backend Developer.

Input data:

• Salary: $2,500/month

• Format: contractor (burden ×1.22)

• CAC: $3,660

• Onboarding: $800

• Ramp-up: 3 months

• Expected revenue: ~$8,000/month

Step 1 — Monthly Real Cost

$2,500 × 1.22 = $3,050

Step 2 — Ramp-Up Cost

$3,050 × 3 × 0.5 = $4,575

Step 3 — Full Investment

$3,660 + $800 + $4,575 = $9,035

That’s the real investment before full productivity.

Once fully ramped:
$8,000 − $3,050 = $4,950 net monthly contribution

The $9,035 investment pays back in roughly two months after ramp-up.

At that scale, the difference between a $900 and $1,200 referral bonus doesn’t change the decision.

The key question is not “Is the bonus expensive?”
It’s “When does the hire break even?”

How to calculate ROI for non-revenue roles

Not every role generates direct revenue. But economics still apply.

1. Cost Avoidance

HR reduces turnover. If two exits per year cost $8,000 each, that’s $16,000 saved.

2. Time Saving

A recruiter frees 5 hours per week for four managers earning $60/hour.
That’s $4,800/month of productivity recovered.

3. Capability Multiplier

An Operations Manager enables 80 clients instead of 50.
Extra 30 clients × $500 margin = $15,000/month.

4. Risk Mitigation

Finance or legal prevents a $20,000 penalty.
One avoided incident can cover annual compensation.

5. Unit Economics Impact

A Product Manager increases LTV from $1,000 to $1,200.
At 500 clients/year → +$100,000.
Even 20% attributed impact = $20,000 in added value.

If a role doesn’t generate revenue directly, it either:

• reduces losses

• frees expensive time

• enables scaling

• lowers risk

• improves unit economics

All measurable in money.

Why most companies don’t show this

Because it requires:

• knowing real CAC

• knowing revenue per employee

• calculating burden

• factoring ramp-up

Most companies never combine this into one financial model.

Conclusion

The question isn’t whether $3,000 or $15,000 is expensive.

The question is: when does it pay back?

Calculate the economics once — and your referral program stops being an HR initiative.

It becomes a financial instrument with predictable ROI.

A bonus should be based on:

• cost of vacancy delay

• time-to-hire

• alternative sourcing cost (agency, ads, sourcing)

• role margin contribution

👉 How do you define referral bonus size in your company?

There’s a lot of noise right now about AI, automation, and doing things “faster.”
And not enough conversation about the reality that meaningful change in business is still hard — even with better tools.

So I want to talk more about real companies.
Real transformations.
And the friction teams face when moving from “how we’ve always done it” to a systematic approach.

That’s why I’m launching a new #CaseStudy series.

Here’s the first one.

The Client Case:

Why transitioning to competency-based hiring is harder than it seems

The client is shifting from evaluating technical and soft skills separately to implementing a structured competency model.

We:

• delivered training

• conducted consulting sessions

• worked through practical cases

The team understands the framework.

Implementation is still challenging.

Not because of a lack of knowledge.
But because it requires a shift in how people think.

And that’s normal. Competencies aren’t just a new interview template.
They represent a different evaluation logic.

Why competencies aren’t theory — they work in practice

Competencies weren’t created to complicate hiring.

They emerged because:

• isolated skills don’t reliably predict performance

• knowledge does not equal execution

• results are driven by behavior, not by a list of tools

Decades of research in organizational psychology show that structured interviews with behavioral indicators have significantly higher predictive validity than unstructured, intuition-driven evaluations.

A competency is not “works well in a team.”

It’s a combination of:

• knowledge

• skills

• mindset

• observable behavior in context

And that combination is what drives business outcomes.

What proved difficult in this case

1️⃣ Mistake #1 — Applying the high-performer model directly to hiring

The company defined competencies based on their strongest internal employees.

For example, “product thinking” internally meant:

• initiating roadmap changes

• influencing strategic priorities

• proposing business model optimizations

Then they began expecting the same behavioral evidence from candidates.

But candidates:

• don’t know internal processes

• don’t have access to your data

• haven’t worked in your context

So hiring-stage indicators must look different:

• thinks through the user lens

• explains how decisions impacted metrics

• articulates trade-offs in prioritization

If you don’t separate these, interviews become unrealistic.

2️⃣ If you’re used to evaluating skills — start by integrating them

You don’t need to break your process overnight.

For example, you may currently assess:

• interview methodology knowledge

• ATS usage

• decision-making

• ability to operate in ambiguity

• speed

Evaluated separately, this often leads to shallow questions.

Instead, combine them into one competency block like “Quality of Assessment and Decision-Making” and use a deep behavioral case:

“Tell me about the most complex role you worked on with high uncertainty. How did you make decisions? What did you rely on? Where did you get it wrong?”

Within a single example, you’ll see:

• how the candidate structures information

• how they manage risk

• how they justify decisions

• how they adapt

You stop evaluating isolated skills.
You start evaluating a behavioral system.

3️⃣ Competency-based interviews are not a checklist

You don’t need to “cover every competency.”

The goal isn’t the number of questions.

The real question is:

• Do you have enough behavioral evidence?

• Do you understand how the person acts in critical situations?

• Have you reduced the key risks of the role?

If a role requires autonomy, but every example shows dependency on a manager, that’s a risk — even if technical skills are strong.

4️⃣ A skills-based approach doesn’t mean breaking everything into 25 items

A strong skills-based philosophy means:

• moving away from degrees as the primary filter

• assessing real ability to perform

• focusing on capability over titles

It does not mean measuring every micro-skill independently.

A mature approach requires understanding:

• which combination of skills enables solving business problems

• which behaviors demonstrate real application

A skill is a component.
A competency is a system.

Why implementation feels difficult

Because the focus shifts:

from “What does this person know?”
to “How does this person think and act?”

And that’s harder to assess.

But that’s what predicts results.

Where the client is now

We completed the foundational phase of the transformation:

• defined key competencies for roles

• clarified behavioral indicators

• delivered detailed training

• documented the updated framework

• redesigned the hiring manager briefing form

• introduced competency profiling at the start of each search

Now, before opening a role, the team defines:

• which competencies are critical

• which can be developed

• which represent risk in this specific context

The focus shifts from “Who are we looking for?”
to “What behavioral model does this business challenge require?”

The transformation is still ongoing.

The team is building a new cognitive habit.
Learning not to revert to intuition-driven evaluation.

That doesn’t happen in a month.

Because what changes isn’t the tool.
It’s the decision-making logic.

Final Thought

This is not a story of “we implemented it — and everything became perfect.”

It’s a story of an ongoing transformation.

If you’ve already defined competencies, updated templates, delivered training — and it still feels messy — that’s not failure.

That’s the stage where a new professional habit is forming.

And that stage is always harder than it looks at the start.

Most companies either revert to the old model
or move all the way through the transformation.

My client chose the second path.

I’d value your thoughts and professional perspective in the comments.

Leave a comment

Lately, I’ve been noticing something strange: after working with AI, I feel more exhausted than after working with people.

And that honestly concerns me.

Especially because I’m neurodivergent — and this feels counterintuitive. I’ve always been more drained by calls, messages, and emotional interactions than by focused solo work.

But here — it’s the opposite.

That’s why I started digging deeper. Because in theory, AI should make work easier.

And this is not about AI being плохий or ineffective. Quite the opposite — I use it daily, and in many tasks it genuinely speeds things up.

But along with that, something else appeared: a constant sense of mental tension that’s hard to explain at first.

What AI is actually doing to work

We usually talk about AI in terms of productivity: faster, more, cheaper. Much less often — in terms of how it changes the nature of work itself.

Because AI doesn’t remove work. It restructures it.

Where work used to have a clearer cycle — do → finish → move on — that cycle is now blurred.

Working with AI becomes a continuous loop of micro-decisions: write a prompt, evaluate the output, fact-check, уточнити, rewrite, check again.

And these loops can happen dozens — sometimes hundreds — of times a day.

Formally, it’s faster. In reality, it feels more intense.

Why this is exhausting (at least for me)

I’m not a neuroscientist, so this isn’t a scientific explanation. It’s me trying to understand what’s happening in my own head.

After learning more about how my brain works, I started paying closer attention to cognitive load — because in my work and lifestyle, thinking is the main resource.

And I started noticing a pattern.

With AI, the mechanical part of work decreases — but the number of decisions explodes.

Every output is not a result. It’s another decision point:

• Can I trust this?

• Is this accurate enough?

• Should I rewrite it?

Each decision is small. But together, they create a constant background noise of tension.

There’s also something else.

We’re losing natural “end points.”

Before, you would finish a task and feel that it’s done.

With AI, that feeling almost disappears. There’s always a way to “improve it a bit more,” “refine it again,” “run one more iteration.”

At some point, work stops having a clear definition of DONE.

What actually worries me

There’s another aspect I’m seeing more and more in teams — and honestly, this worries me more than the fatigue itself.

AI easily amplifies confidence. But it doesn’t always amplify the quality of decisions.

This is especially visible in managers (including myself).

Instead of deeply thinking through a problem, it becomes very easy to get an answer that sounds structured, logical, and convincing.

It looks right. It feels right.

But without the habit of critical evaluation, this quickly turns into an illusion of expertise.

I’ve already seen cases where decisions sounded smart — but didn’t hold up under even basic context checks.

And this is far more concerning to me than whether AI is “effective.”

We adopted the tool — but not the rules

So for me, the question isn’t whether AI is good or bad.

The question is how we work with it.

Right now, it feels like mass adoption of a powerful tool without redefining how work should actually be done.

We haven’t agreed:

• where verification is required vs. where “good enough” is fine

• which decisions can be delegated — and which shouldn’t

• where AI supports thinking — and where it starts replacing it

And we definitely haven’t accounted for the fact that this isn’t just automation.

It’s a new type of cognitive load that we haven’t adapted to yet.

What’s next (a question, not an answer)

I don’t have a clean solution yet.

But it feels like the next step isn’t “use more AI.”

It’s learning how to work with it without losing the quality of thinking — and without burning out from constant micro-decisions.

With rules.
With boundaries.
With clarity on where it actually helps — and where it just adds noise.

Because otherwise, instead of simplification, we’re just creating a different — and more complex — form of fatigue.

Using AI to write resumes.
Optimizing profiles for algorithms.
Testing different wording.
Applying to hundreds of roles.

But there’s a question that gets asked much less often.

Are we actually filtering out dishonest candidates —
or simply those who don’t know how to play the algorithmic game?

Because modern hiring increasingly looks less like evaluating people and more like an interaction between two optimization systems.

Candidates optimize visibility.
Companies optimize filters.

And in the middle — people simply trying to find a job.

Hiring has become an Algorithmic Market

The modern candidate experience increasingly looks like this:

LinkedIn shows jobs through recommendation systems.
AI tools filter resumes using keywords.
Algorithms rank candidates.
Job boards use matching systems.

The result is a new reality.

Being a strong candidate is no longer enough.

You also need to be visible to the system.

And visibility has become its own skill.

Candidates optimize:

• keywords in resumes

• experience descriptions

• profile structure

• headlines

• skill ordering

• sometimes even adding “hidden” text in resumes

This isn’t necessarily deception.

It’s adaptation to the infrastructure of the market.

(Though the white-text-in-your-resume trick is a classic example of the internet inventing hacks that either don’t work or just look… incredibly dumb.)

Algorithms are the new First Interview

For a long time, the first step and barrier in hiring was a recruiter.

Today, more and more often, it’s an algorithm.

LinkedIn uses ranking systems to:

• search for candidates

• recommend jobs

• sort profiles in recruiter searches

ATS systems may also apply simple filters — for example by keywords, location, or basic role criteria.

And it’s worth saying this clearly because there’s a lot of mythology around it:

Filters are not AI screening.

Social media sometimes makes it sound like ATS platforms are some kind of hyper-intelligent systems automatically deciding the fate of candidates.

In reality, most ATS platforms barely handle basic search inside their own databases.

Yes, automated filters exist.
But the reality is far more mundane than the popular fear of “AI rejecting everyone.”

Still, the effect is real.

A large portion of candidates never reach a human at all.

They are filtered out at the level of algorithmic visibility — not decision.

Visibility algorithms: how LinkedIn decides who You are

LinkedIn today is no longer just a database of resumes.

It’s a recommendation platform.

Newer algorithms increasingly rely on behavioral signals to understand where a person belongs professionally.

For example:

• who you follow

• which posts you like

• what you comment on

• which topics you engage with

• which companies and pages you interact with

These signals help the system determine which professional ecosystem you belong to.

Based on this, LinkedIn decides:

• which jobs to show you

• where your profile appears in recommendations

• how you rank in recruiter searches

In other words:

LinkedIn is becoming an algorithm of visibility, not just a resume database.

This creates a new reality for candidates.

Your professional visibility is no longer determined only by your experience.

It’s also determined by your behavior inside the system.

If someone doesn’t engage with professional content, doesn’t follow industry topics, and doesn’t show activity within their field, the platform may simply fail to associate them with that professional niche.

And then their profile appears less often:

• in recruiter searches

• in job recommendations

• in professional feeds.

The market starts optimizing for Algorithms

Once this becomes clear, behavior changes.

Candidates begin optimizing not only for resume filters, but for visibility algorithms.

They:

• follow industry topics

• like professional posts

• comment on discussions

• shape their profiles so the system reads them as part of a specific industry

This is no longer just networking.

It’s adaptation to an algorithmic labor market.

And gradually the market begins rewarding not only competence — but algorithmic presence.

When everyone plays the System

Once an algorithm becomes a gatekeeper, people adapt.

Candidates experiment with:

• multiple versions of their resume

• different ways of describing roles

• AI-generated experience summaries

• SEO-style optimization of profiles

• optimization for LinkedIn search algorithms

This isn’t a new phenomenon.

SEO has existed for decades.
Content is optimized for Google.
Marketing is optimized for social media algorithms.

Now the same thing is happening with people.

Hiring has become another algorithmic environment.

A new Inequality: Algorithmic Literacy

This creates a new type of inequality.

Not only:

• education

• experience

• professional networks

But also: algorithmic literacy.

The ability to:

• understand how systems work

• present experience in machine-readable ways

• stay visible in search

• use AI to structure information.

Two candidates with identical experience may have completely different outcomes simply because one understands how the system works — and the other doesn’t.

Non-Linear candidates become Invisible

This affects candidates with non-traditional careers the most.

For example:

• people returning from career breaks

• career switchers

• interdisciplinary professionals

• candidates with non-linear work histories

• people with experience in smaller companies.

Algorithms struggle with complex stories.

They prefer clear patterns.

Senior engineer → senior engineer.
Product manager → product manager.

The more complex the trajectory, the harder it is for the system to interpret it.

The result is a paradox.

A system designed to optimize talent discovery may actually reward predictability instead of potential.

What this means for Candidates

Being good at your job is no longer enough.

Today you also need to:

• understand how the market works

• remain visible inside systems

• structure your experience clearly

• use AI as a tool.

This is simply the new reality of the labor market.

The real Question

When companies say: “Candidates are gaming the system.”

It may be worth asking a different question.

What if the system itself rewards optimization?

Because when the entire market begins operating according to algorithmic rules, we risk losing more than dishonest candidates.

We risk losing people who simply don’t know how to be algorithmically visible.

And a small meta observation.

I hope this post performs slightly better with the algorithms than my previous one — the one that had the word “women” in the title and ended up getting the lowest reach on my LinkedIn.

Apparently algorithms have their own biases too.