500 applications and one person: how to actually screen them

Five hundred applications is the point where a hiring process stops being a task and becomes a backlog. Read at a genuine three minutes each, that is 25 hours — three working days in which nothing else gets done. Here is what actually has to happen, what our own measurements say about the machine part, and what stays slow no matter what you buy.

First, an honest declaration about our numbers

We have never screened 500 CVs in one batch. Our largest real batch is 22. What we have is 24 batch runs with a measurable start and finish, ten of them of ten CVs or more, and a rate that has been consistent across all of them. Everything below that is arithmetic from those runs, and labelled as such.

At 800 an hour, 500 CVs is about 37 minutes — arithmetic, not a run we have done.

Why one CV takes 20 seconds but 500 take 37 minutes

Because they do not queue politely. A single application arriving through the form is one job doing one thing: read the PDF, parse it, score it against the posting, store it. Nearly all of those 19.7 seconds is spent waiting for a model to answer, not computing anything.

A batch runs ten of those at once. The waiting overlaps, so the per-CV wall time collapses from twenty seconds to about four and a half. That is the entire trick, and it is worth understanding because it tells you what does not speed up: anything that has to happen one at a time, in order, with a human in it.

What 500 applications actually costs by hand

From our breakdown of hiring costs: at 20 CVs an hour — three minutes each, which is generous if you are reading properly — 500 applications is 25 hours of reading. Priced at an illustrative Tk 341 an hour, that is about Tk 8,525 of someone's attention, which is roughly three times what the job posting cost.

And the 25 hours is the optimistic version, because nobody reads 500 CVs at a steady three minutes each. Attention falls off after the first hour, which is where the real damage happens: in three of our five real postings the highest-scoring CV arrived in the second half of the pile.

The order to do it in

1. Fix the posting before you fix the process. A large share of 500 applications is often people who could never have been right, invited by a vague description. That is cheaper to fix at the source than to filter afterwards.
2. Score everything before reading anything. Whatever your ranking method, apply it to all 500 before opinions form. A rubric on paper works; it is just slower.
3. Read in ranked order. Then fatigue costs you the weakest CVs instead of a random slice.
4. Set the cut on the rank, not on the clock. "Everyone above this line" survives a long pile. "The first fifteen who look good" does not.
5. Reply to everyone, once. Cheap, and the only part of this a candidate ever sees.

What does not get faster

• Deciding. Ranking 500 CVs takes half an hour. Deciding which 20 deserve a conversation still takes a person, and should.
• Interviewing. Twenty interviews is twenty interviews.
• Replying. Sending is instant; deciding what to say is not, and nobody should be closed out without a person choosing to.
• Being wrong. Screening faster does not improve your requirements. If the posting asked for the wrong things, you will now be wrong 500 times more efficiently.

What breaks at volume, and what we cap on purpose

Two different things, and vendors tend to blur them.

What actually breaks: not much on our side, and we measured it rather than assuming. A burst test with the model calls removed handled 20 concurrent requests at about 3,350 CVs an hour using a third of one CPU. The limits that bite first are the model provider's rate limits, not the server.

What we cap on purpose: how many CVs are scored per month, per workspace, according to the plan. This matters because of what happens when you reach it. Applicants are never refused at a limit. They keep their place and wait until there is room, and they are never told a limit exists. An applicant should never discover that they applied on the wrong day of your billing month.

That is a question worth asking any vendor — along with five others — because "what happens when I hit the limit" has three possible answers and only one of them is acceptable.

If you take one thing

The reading is no longer the expensive part. That means the honest question about a 500-application role is not "how do I get through them faster" — it is "now that the reading is cheap, how many people do I actually want to talk to, and what evidence do I want before I do?" That is a better problem to have than the one you started with.