17 September 20266 minutes
An agent-run process does not hire for you: it gives you the hours back and takes the doubt away
What actually changes when the repetitive work of an opening is done by an agent, how much time comes down, and why the part that matters most is still yours.
The cost nobody invoices
An open role has a cost that appears on no invoice: the seat that is not filled. While it lasts, someone on the team does that work on top of their own, deadlines slip, and the decision gets made in more of a hurry than anyone would like.
The other cost is worse because it arrives later: the wrong hire. It is paid in months of salary, in the time of the team that onboarded them, and in reopening the same search with less appetite.
What follows is about those two costs and nothing else. Not about replacing anyone.
Where the hours go
The work of an opening is not spread evenly. The part that decides, talking to someone and realising whether they fit, is short. What eats the hours is what comes before: opening forty CVs to find three worth reading properly.
And it is work you can describe precisely: checking what the posting already asked for. Whether the role is on site in a city, whether it needs a specific language, whether it asks for a number of years. That is not judgement, it is verification.
What actually predicts performance
Corrected validity for each method, from the Sackett et al. meta-analysis (2022). Higher predicts better.
- Structured interview0.42
- Job knowledge test0.40
- Work sample (the exercise)0.33
- General cognitive ability0.31
The exercise built from the posting is a work sample. The call about what the person wrote is a structured interview. Both sit at the top of the table.
And one figure to stop repeating: cognitive ability went from 0.51 to 0.31. The 0.51 from Schmidt and Hunter (1998) still circulates in decks and is no longer the number.
What an agent does, and what it does not
An agent does the checking and writes down what it checked. That second half is the one that matters: a number on its own is not enough. A score with nothing behind it cannot be defended to the client or to the candidate.
So every requirement comes back with one of three answers: verified and met, verified and not met, or there was nothing to check it against. The third is the one almost nobody shows, and it is what stops you discarding someone over a fact that was never there.
What the agent does not do is choose. The conversation with a person is still where it gets decided, and there the agent brings something else: that you arrive already knowing what that person solved and how.
Follow-up, the invisible half
The other half of the time goes on answering. What stage am I at, when do I hear back, what do I have to do now. They are reasonable questions, and answering them one by one over email is what makes a process feel slow even when it is not.
An agent that answers the candidate there and then changes two things at once: it brings the process time down, and it brings down how many people drop out halfway because they got tired of waiting. The second is the least measured and the most expensive.
How wrong hires get rarer
Not with more filters. With more evidence before the interview.
A short exercise from the actual role, built from the posting itself, says more than half an hour of general questions. And when the person then talks about what they solved, it shows straight away whether they understood it or copied it.
That is the whole mechanism: arriving at the conversation with something concrete on the table. It is not a new trick, it is what any good interviewer does when they have time. What changes is that now there is time.
What is worth measuring
Three numbers, and none of them is how many CVs were read: how many days pass between posting and closing, how many candidates drop out along the way, and how many of this year's hires are still there at six months.
If an agent-run process does not move those three, it did not help, however tidy the dashboard looks.