Why More AI Screening Is Making Your Hiring Worse

Two numbers, side by side. Applications per job opening have roughly doubled since 2022. Application volume jumped over 45% year on year, with LinkedIn processing something like 11,000 applications a minute globally. And yet employers report more difficulty filling roles than at almost any point in the last decade. More candidates. Harder hiring. If that does not strike you as strange, it should. It breaks the assumption underneath basically every recruitment tool sold in the last three years.

What Actually Broke

78% of job seekers now use AI in their applications or say they would. 64% of recruiters report seeing more look-alike applications as a result. Meanwhile, 83% of companies plan to use AI to review CVs, and on some estimates AI now touches the screening of the large majority of incoming CVs. So: a machine writes the application. A machine reads the application. Both machines were trained to optimise for the same keyword patterns in the same job specification. This is where the usual AI is transforming recruitment post ends, with a note about human oversight. I want to make a sharper claim. Adding AI screening to an AI-written applicant pool does not filter noise. It manufactures it. Filtering only works when the signal you are filtering on is expensive to fake. A CV used to cost something, time, thought, and some evidence that the person had actually read the advert. That cost is now near zero. When the input cost collapses, the filter is not separating good from bad. It is separating well-prompted from badly-prompted. Your shortlist is no longer a ranking of candidates. It is a ranking of prompt quality. Those are not the same thing, and one of them has nothing to do with whether the person can actually do the job.

The Cost Moved, and Most Hiring Processes Have Not Noticed

For twenty years, the expensive part of recruitment was sourcing: finding people. Almost every tool, budget line and KPI in talent acquisition is built around that problem. Sourcing is now nearly free. Post an advert in South Africa, where we have 8.1 million unemployed people and 32.7% unemployment as at Q1 2026, and you will get volume. Guaranteed. You will get 400 applications for a bookkeeping role in Midrand, and you will feel briefly successful about it. The expensive part is now verification. Is this person real? Are the claims true? Can they do the work? Will they stay? Most companies are still spending their money on the free problem. You can see the consequence in the funnel. Every additional screening round you add to absorb the volume adds days. Every added day increases the chance your genuinely strong candidate, the one already talking to three other employers, possibly two of them offshore, accepts elsewhere. High-volume, slow-verification hiring does not just cost more. It systematically selects against the best people in your pipeline, because they are the ones with alternatives and the least patience.

What the Evidence Says Actually Predicts a Good Hire

The most interesting statistic in recruitment right now is old-fashioned, and it is about referrals. Referrals are roughly 7% of applications but account for 30–50% of hires. Referred candidates are around 4x more likely to be hired, get hired about 55% faster, and show meaningfully higher retention at the two-year mark than job board hires. Something like 88% of companies rate referrals as their best source for quality of hire. That gap has existed for years. Nobody could scale it, so it got treated as a nice-to-have. Look at why referrals work, though, and it is not magic and it is not loyalty. It is that a referral is a verified claim from someone with reputational skin in the game. That is expensive to fake. It is a real signal. Which means the question for 2026 is not how do I screen faster? It is: where do I get signal that is costly to fake?

A Verification-First Checklist

Here is the practical version, for whoever owns hiring at your company:

• Stop measuring applications received. It is now a vanity metric, and it may be inversely correlated with pipeline quality. Measure verified-candidate-to-offer instead.
• Replace one CV screening round with one work sample. A 30-minute realistic task tells you more than 200 CVs. It is also the one thing a generic AI application cannot produce for you.
• Ask about a specific decision, not a general responsibility. Walk me through the last time a month-end did not balance is very hard to bluff. Tell me about your strengths is not.
• Track days-to-offer as a leak, not an admin stat. Put a number on it. Most South African employers I speak to are surprised, and not pleasantly.
• Buy pre-vetted, not pre-sorted. An algorithmically ranked list is not a vetted list. Someone must have actually spoken to the human.

AI Is Not the Problem, Verification Is

None of this is anti-AI. I use it daily, and it is excellent at scheduling, drafting, summarising, note-taking and translation. It is just not a truth detector. And it cannot be, because the thing generating the applications is the same thing reading them. The employers who will hire well over the next two years will not be the ones with the most automation. They will be the ones who worked out which part of hiring was never a volume problem in the first place.