The AI doom loop in recruiting

Because both sides armed themselves at the same time, with the same weapon. Candidates apply using generative tools, recruiters screen and reach out using generative tools, and the channel through which the two talk to each other is collapsing while everyone involved can demonstrate, with numbers, that they made the right individual choice.

Greenhouse’s chief executive has a phrase for it: hiring is stuck in an “AI doom loop”. Harvard Business Review was blunter still in January 2026, with a headline that travelled a long way on LinkedIn, AI has made hiring worse. These are not sceptics talking; these are people whose business is selling recruitment tools.

What the numbers say, and they are brutal

LinkedIn connection-request reply rates fell from 3.5% in May 2025 to 2.2% in April 2026. In under a year the channel lost more than a third of its effectiveness, and nothing in the platform’s algorithm explains that fall: volume produced it.

On messaging, a study covering more than five million sends produces a result few vendors enjoy quoting. On email, hand-written outreach gets 12.6% replies against 5% for generated outreach, more than double. On LinkedIn the ratio inverts, around 16.9% for the assisted version, which deserves saying plainly rather than leaving out.

And on the candidate side, the scale of the flood is captured better by an anecdote than a statistic: an American newsroom posted one job and was buried under AI-generated applications within twelve hours.

A tragedy of the commons, in the strict sense

What is happening has a name in economics, and naming it helps explain why nobody stops.

Automated personalisation works. The first recruiter to send two hundred tailored messages where competitors send twenty generic ones gets excellent results, and is right to do it. The problem appears once everybody has done it: the recipient now receives twenty equally well-turned, equally flattering, equally false messages, and stops answering all of them.

The mechanism is vicious because every vendor can simultaneously be telling the truth. Yes, their tool improves their customer’s results relative to what that customer would get without it. And yes, the sum of those individual improvements lowers the channel average for everyone, their customers included. Both claims hold, which is precisely what makes the loop so hard to break: there is no moment at which stopping is rational for a single participant.

At least the next step is predictable. A saturated channel eventually closes in one of three familiar ways: the platform puts a price on what was free, it hardens detection, or the recipients move elsewhere. All three are already underway, which is one more reason not to build your outreach strategy on a channel whose rules and account you do not own.

The disagreement that does not exist

There is one thing everybody agrees on, which is rare in this debate.

71% of Americans oppose AI making the final hiring decision. On the other side, 85% of recruiters say they want to keep that decision. Both camps want the same thing, which ought to settle it, and yet the market keeps selling end-to-end automation as though the demand were there.

Our reading is that this supposed debate hides the real one. The question was never whether the machine should decide: almost nobody wants that. It is what you hand it immediately before the decision, and that is where positions genuinely differ, without many people bothering to spell them out.

How to tell the loop has caught you

Three indicators are enough, and they come from data you already hold.

The first is your own reply rate over a rolling twelve months, channel by channel. Its absolute value does not matter much, since it depends on your profiles and your brand, but its slope does: a steady decline while you send more is the exact signature of the phenomenon described above. Many firms measure only volume sent, which means watching the one variable that always goes up.

The second is the ratio of applications received to interviews held. When it degrades without any change in the quality of your postings, you are not receiving more candidates: you are receiving the same candidates spread across more applications. The difference is enormous and invisible on a dashboard.

The third is the time your team spends screening, divided by completed hires. It is the most painful one to look at, and the only one that justifies a budget in front of a finance director.

One last sign, less measurable and very reliable: the day your own recruiters start saying they can spot a generated cover letter in three seconds, the loop has run its course. It means the channel no longer carries information, and a channel that carries no information is eventually abandoned by the people who were listening.

What still works, and why

The way out is not adding AI to the saturated channel. It is working the ones where volume confers no advantage, because they rest on a prior relationship.

Three of them are available to any staffing firm from tomorrow morning.

People you have already met. A candidate interviewed two years ago remembers you, and a message opening with an exact reference to that conversation has no equivalent in a campaign, precisely because it cannot be mass-produced by somebody who was not there. That assumes the trace exists, which brings you back to the real state of your talent pool.

Referrals from your own consultants. They remain the best channel in the industry, they are under-exploited because they require regular follow-up nobody has time for, and that is exactly the kind of repetitive, decision-free task an agent handles well.

Applications you never processed. It is the most embarrassing and cheapest resource you have: these people already wrote to you, and the only work is getting back in touch: a subject we covered elsewhere from the angle of the silence candidates experience.

There is a version of this where the loop simply closes and hiring goes back to being a relationship business conducted mostly through people who already know each other, which is roughly how it worked before job boards existed. We would not bet on that outcome, but it is worth noticing that the firms least disrupted by the last two years are the ones that never depended on cold volume in the first place.

What we took from this when building Balt

We hesitated, while building the product, over bulk sending. The demand is real, it comes up explicitly in almost every sales conversation, and refusing to serve it is an expensive decision.

We decided otherwise, for a reason that is not moral. A tool that helps its customers saturate a channel destroys the value of that channel for its own customers, usually within less time than the contract runs. Put another way, the feature sells well in year one and turns on the buyer in year three: we would rather not build that kind of debt.

What we do instead fits in a sentence: the agent works memory rather than volume. It remembers who you met, what was said, who has been waiting eleven days for an answer; it drafts the message and you send it. That is less impressive than a send counter, and it is the only advantage that does not devalue the moment your competitor buys the same tool.