Let the machine sort, but you sign it yourself

Hackerrank has released an AI interviewer called Chakra that conducts job interviews for developer positions. It asks follow-up questions and judges not only whether a working solution comes out at the end, but how someone got there. Almost at the same time it became known that the Bundeswehr wants to screen applicants automatically in order to spot extremists earlier.

Two very different cases, one shared question. Which part of an application may run through a machine before a human decides. The answer can be played through on a perfectly ordinary open position.

Your open position and the nine hours

Say you are looking for an inside sales clerk. The ad runs for four weeks, and 90 applications are sitting in the inbox. Reckon six minutes for a first look at each one, and that comes to nine hours before the first interview has even taken place. On top of that come the acknowledgements of receipt, chasing missing references, coordinating dates with two colleagues from sales, and 80 rejections at the end.

The larger part of this is busywork. None of these tasks calls for judgement of character, they call for time. And they are the reason applicants in mid-sized companies drop out after three weeks, not because the company is unattractive, but because nobody got back to them.

What the machine may pre-sort

Anything that can be checked by recalculating it is worth automating:

• Acknowledgement of receipt within an hour instead of on Friday afternoon, with a realistic date for when an answer will come.
• Completeness checked and missing documents requested before anyone opens the folder.
• Hard criteria that were stated in the ad and can be verified, such as a completed commercial apprenticeship or a driving licence, if the job really needs one.
• A summary per application, two paragraphs, with the source reference. If it says someone worked with your ERP system for five years, it has to say next to it which page of the CV that comes from.
• Proposed interview dates drawn from the calendars of everyone involved, as a draft ready to send.

The source reference is not cosmetic. If a system puts out an assessment and nobody can trace where it came from, the assessment is worthless. A summary without a source does not replace reading, it only shifts the risk.

Where a human signs

The machine sorts, it does not delete. An order of ranking is a suggestion, not a result. Nobody is out because a score falls below a threshold, and the pile of those sorted out gets looked through by a human rather than archived. The invitation and the rejection are decisions, and they are signed by someone who stands behind them.

We build our own systems to the same pattern. In the platform for Kunst gegen Bares that we operate, the AI writes drafts per channel, and nothing is published until a human has approved it. With a social post, the damage from an automation that ran all the way through is an embarrassing post. With an application, it is a person who gets no answer to the question why.

The legal position draws the same line. Article 22 GDPR gives people the right not to be subject to a decision based solely on automated processing that has a significant effect on them. A rejection that no human has looked at is exactly the case you do not want to bring upon yourself.

Every rejection needs a sentence you can read out loud

This is the point where pre-sorting gets expensive. Under § 22 AGG, it is enough for a rejected applicant to present indications of discrimination. After that, you have to prove there was none. They have two months to assert the claim under § 15 (4) AGG, and during that time you have to be able to explain why this particular person got the rejection.

"Score 62" is not an explanation. "No evidence of the training required in the ad, see attachment" is one. So record one sentence per rejection, ideally naming the criterion from the ad it refers to. That is not extra work, that is the actual yield of the pre-sorting.

That this is not a theoretical worry was reported by Reuters back in 2018, when Amazon switched off an internal selection tool again because it rated women applicants worse. The mechanism behind it is not an isolated case, it is the nature of the thing. A system that learns from your past hires carries your past hires forward. If the same type of person has been sitting in your inside sales team for ten years, that is exactly the type you get suggested again.

Three ways to introduce something like this

There are not just two. Renting means you take an applicant management service and pay per user or per position. Quick to set up, and the applicant data sits with the provider. Building and running it yourself means you need someone in house who maintains it, including in July when half the team is on holiday. In between lies the third way, having it operated for you. Third-party or purpose-built software runs on a server that belongs to you, maintained by someone who is paid to do it. The data stays with you, the on-call duty does not. This is the category we work in, which is why we say so rather than leave it out.

Whichever way you go, ask the provider three questions. How long is applicant data kept and who deletes it. Who can view it. And what happens to it if you terminate the contract.

The yardstick you can apply

AI takes the nine hours of sorting off your hands, not the signature. Whether your process holds up is something you test with a single question. If one of the 80 rejected applicants calls you, can you tell them in one sentence why. If yes, the machine in front of it may pre-sort as much as it likes. If no, your problem is not the automation, it is that the criterion was missing in the first place.