At a Wharton event last week, I met someone who works in commercial real estate. Twenty to thirty deals land on his desk every day. Each needs a valuation and a decision about whether it deserves a closer look.
That sounds like a good job for AI. It also sounds like the job every other buyer will give it.
Years ago, I opened a role on LinkedIn and had hundreds of applications within 30 minutes. Plenty weren’t close to qualified. I read them by hand. And if someone had offered me a decent filter, I definitely would have taken it.
These days, the person applying has AI, too. They can tailor the résumé, write the cover letter and move on to the next opening while the employer’s software decides whether they deserve a conversation.
AI made applying and rejecting cheap. Now candidates and employers are automating their way past each other.
Ashby’s analysis of more than 100 million applications found that applications per hire have tripled since 2021, to more than 300 on average. Not all of that is AI. All of it lands on someone’s desk.
I understand both sides. Writing cover letters after dinner is a lousy way to spend your evening. Reading them all week has its own limits. Each side has a good reason to automate, and each side’s automation gives the other more to deal with.
A tailored application used to suggest someone had spent time learning about the company. Now it might mean they pasted the job description into a chat box. The person could still be excellent. On Freelancer.com, the premium for tailored applications fell sharply after AI writing tools arrived.
Polish used to cost effort. Now it tells you much less about intent.
Then there’s the software’s taste.
In a research preprint on AI hiring, researchers generated new résumé summaries while keeping the candidates’ underlying qualifications fixed. Models tended to favor their own summaries over the human-written versions. In simulated hiring pipelines, candidates whose summaries were written by the screening model were 23% to 60% more likely to be shortlisted than those with human-written summaries.
Models turn out to be like the rest of us. They like their own writing best.
For the applicant, adapting is reasonable. If the machine likes a certain voice, use it. For the employer, the applications get harder to distinguish. Time to turn up the filter.
The machinery can keep improving while the information each side gets about the other gets worse.
I used to joke that by the time I left Apple, I couldn’t have passed the interview.
That was before AI. Hiring by consensus was already good at making an unusual candidate difficult to approve. One person sees a standout skill. Someone else sees a gap. The gap is easier to put on a scorecard.
A filter can carry that habit across far more people than an interview panel ever could.
This already shows up in the data. Researchers studied 4 million applications screened using game-based assessments from one vendor. Of people who applied to 10 jobs, 4% were recommended for rejection by all 10, “a rate higher than expected by chance.”
We tend to think of applying to another company as getting another chance. That depends on whether you get another judgment.
Jon Kleinberg and Manish Raghavan modeled this problem in 2021. They showed conditions under which employers adopting the same algorithm can produce worse collective hiring decisions, even when that algorithm is more accurate for each employer in isolation.
A tool can pass its individual test and still make the market worse.
Companies can get different decisions from the same language model by giving it different criteria and evidence. My concern is how easily “find me a good candidate” becomes a default judgment nobody has thought very hard about.
I’d have wanted that filter in my 30-minute flood. I doubt I would have stopped to ask who else was using it.
Which brings me back to those 20 to 30 deals a day.
Every buyer wants a property worth more than its price. Investors call that edge alpha, and everyone in a matching market wants some. Employers want the candidate everyone else overlooked. Candidates want the job nobody else spotted.
Alpha takes a view about the property that other buyers haven’t acted on. If AI helps everyone arrive at the same valuation, it may save them all time. It also helps them find the same bidding war.
There’s nothing wrong with agreeing. Plenty of unusual deals deserve to be passed over.
But if I’m looking for an edge, I need some reason to value the opportunity differently. A model can help investigate that reason. I still have to have one.
Warren Buffett tells a version of this in his 1995 shareholder letter. On a Saturday in January 1951, he took the train to Washington to learn about GEICO. The building was closed. He pounded on the door until a custodian appeared and pointed him to the one man working upstairs: Lorimer Davidson, who later became CEO. Buffett’s only credential was being Ben Graham’s student. Davidson gave him about four hours anyway.
Buffett bought GEICO stock that year. In 1995, Berkshire agreed to buy the rest of the company.
That’s the alpha. That conversation gave him a reason to value GEICO differently. Today everyone can pull the same filings and ask the same model about them. The train ride is the part that doesn’t scale.
I’d like AI to help more people find a conversation worth the trip.
My first instinct for hiring was to cap applications. Make people choose where to spend them, the way Google limits applicants to three jobs in any 30 days. Cap the number of résumés an employer has to read, too.
Thinking it through, a first-come cap would mostly reward whoever had the fastest bot.
A limited signal of interest seems more promising. Keep applications open, but give candidates a few chances to say, credibly, “I chose you.” Give recruiters a similar limit on outreach. The limit would have to hold across the platform. Otherwise everyone just buys more signals.
I’d try that as an experiment. It could help me decide where to spend attention. It would still leave me with the work of finding out whether we fit.
Referrals carry something similar: someone is willing to put their reputation behind an introduction. On Ashby’s platform, referrals are about 1% of applications and about 18% of hires. They also favor people who already know someone. I’d be wary of responding to an automated hiring market by making personal access even more necessary.
I’d also spend some of the time AI saved me looking at applications it wanted to reject. What evidence did it discount? Is there a record of rejections and reasons? Did a missing title outweigh work that actually mattered to this role? Could I check that work?
A tidy shortlist can hide a lot. You only meet the people it lets through.
It’ll probably get worse before it gets better. What interests me is how much better it could get.
An AI could read more of someone’s work than I ever had time to. It could help connect experience in one industry to a problem in another, then show me enough evidence to decide whether the connection holds.
That’s a much more interesting assignment than making every résumé look like the job description.
It also asks more of me. I have to explain the work we need done, decide which gaps we can tolerate and be willing to look at someone who makes the scorecard awkward. Handing that judgment to software is tempting precisely because it feels like getting the work off my desk.
I keep thinking about those hundreds of applications. I remember the volume. I have no idea what I missed.
Who’s the best person you found that a filter would have missed?
Turn today's calls into tomorrow's proposal
Wispr Flow Notetaker brings your meetings into Claude or ChatGPT through a built-in connector. Ask for a proposal draft or a deal recap, and your AI works from an accurate record of what was said.
