Good candidates are gone after two weeks of silence.
A well-advertised position brings applications through three channels at once: the job portal, a central recruiting mailbox, and still paper by post. Every document looks different. CVs as a table, as running text, or as a scan. Cover letters sometimes detailed, sometimes three lines. References in any order.
That is why screening takes time. HR reads every application individually, notes in a spreadsheet what fits, and sends a selection to the hiring department. There it sits. The colleagues who can judge the professional side are on the shop floor or with customers and rarely answer the same day.
The pattern is always the same: one to two weeks pass between arrival and the first real response. In that time good candidates have had two other conversations and withdraw, without anyone ever having judged their documents professionally. The bottleneck is not the decision, it is everything before it.
Sharpen the requirement profile first, then structure.
The start was not a software demo but one real open position and the question of what HR and the hiring department actually use today to judge who fits. Usually it turns out that the requirement profile in the job ad and the one in the hiring manager’s head are two different things. One condition from the team became the specification: the AI employee must never reject anyone.
- Make the requirement profile explicit: Write down must-have and nice-to-have criteria together with the hiring department, professional and verifiable. Protected characteristics are not criteria. Whatever cannot be evidenced in an application belongs in the interview, not in the screening.
- Test against real past applications: The AI employee runs over closed processes whose outcome the team already knows. What gets checked is not whether the AI selects correctly, but whether it structures correctly and evidences correctly.
- Fix roles and permissions: Who sees which application, who decides on the shortlist, who releases invitations and rejections. The two human stations live in the system, not in a convention.
- Start small, then roll out: One job family with high application volume first, typically skilled-trade or commercial positions. Leadership roles and special cases come later or not at all.
80 percent is done by the AI employee. The human decides at two points.
The AI employee pulls all three channels into one inbox: applications from the job portal, emails with attachments from the recruiting mailbox, and scanned paper applications. From every application it builds the same structure: education, career stations with time spans, skills, and certificates. For every item it records which document it came from and where.
It then compares this structure against the professional must-have and nice-to-have criteria of the position and marks each criterion as evidenced, not evidenced, or unclear. There is no score, no ranking, and no recommendation label. Every application that arrived lands in the comparison overview for the position, in order of arrival. For unclear points the AI employee drafts a follow-up question. HR and the hiring department decide on the shortlist and release invitation or rejection. Only after that does the AI employee coordinate appointments and monitor the retention deadlines.
Application for an industrial mechanic position, received by email: CV as PDF, short cover letter, two employment references as scans.
One field is flagged for decision: the shortlist. The AI employee assigns no score, produces no ranking, and rejects nobody. Four of five evidenced criteria are information, not a verdict.
That was before, this is today.
This is how candidate screening ran before, and this is how it runs today with the AI employee. Conservatively calculated, that saves about 80 percent of the screening time per application. The decision itself does not get faster, it only gets better prepared.
| Before | Today | |
|---|---|---|
| Time per application | 10 to 15 minutes of screening | About 2 minutes of review |
| First response | 1 to 2 weeks, often longer | Within one day |
| Channels | Portal, email, and post screened separately | One inbox for all three channels |
| Comparability | Every CV in its own format | One uniform structure, one overview per position |
| Traceability | Assessment rarely documented | Every criterion backed by a source reference |
| Selection decision | Human, under time pressure and without evidence | Human, on a complete and evidenced basis |
* Savings conservatively calculated: 10 to 15 minutes of screening per application before, about 2 minutes of review of the structuring today, calculated with the lower value. At 100 applications that is about 13 hours less screening work. The vacancy figure assumes the usual range of 4,000 to 10,000 € cost per month for an open position and a position filled two weeks earlier. The selection decision stays with the human in both columns.
How do you build an AI employee like this, technically?
The knowledge from this scenario to take away, whether you build with us or on your own:
Reading: vision models instead of CV parsers
Classic CV parsers fail on table layouts, on scans, and on anything that does not follow the expected structure. Modern vision models like GPT from OpenAI or Claude from Anthropic read a CV the way a human reads it: as a table, as running text, or as a photo of a paper document. They map it into a fixed schema of education, stations with time spans, skills, and certificates.
Structuring: every criterion with a source reference
For every extracted item the system records which document it came from and where, for example welding experience, CV, station 2013 to 2019. Anything that cannot be evidenced is marked as unclear rather than guessed. An invented career station is not a cosmetic error in recruiting.
Criteria: professional only, from the requirement profile only
Only criteria derived from the requirement profile are checked: education, work experience, certificates, language and technical skills, driving licences. Protected characteristics under section 1 AGG are not criteria and must not enter as proxies either, so no photo, no year of birth, no name. The criteria list is visible in the system and not hidden in model behaviour. This is also where the line to a high-risk classification under the EU AI Act runs: pre-structuring and evidencing yes, scoring and ranking no.
Integration: ATS, mailbox, and retention rules
The AI employee connects to the existing applicant tracking system and to the recruiting mailbox in Outlook. There is no new platform. GDPR retention periods are not an AI topic but a fixed rule set: intake date plus closing date give the deadline, six months after the process ends is common. The AI employee reminds and lists, deletion happens on release.
UX: the interface decides adoption
The hiring department gets a comparison overview where all applications for a position sit side by side, complete and without ranking. Follow-up questions to applicants wait as editable drafts, nothing goes out unseen. And every decision is a deliberate click, never a silent pre-sorting. Exactly these three things turned legal skepticism into approval.
What does it cost in comparison?
Superkind charges per use case. The price grows with application volume, not with headcount. Here is the honest comparison:
| Manual screening | ATS keyword filter | Superkind AI employee | |
|---|---|---|---|
| Cost | 10 to 15 minutes of HR time per application | 3,000 to 15,000 € per year for licence and upkeep | Price per use case, a fraction of a full-time position |
| What is included | Reading, comparing, and chasing, all by hand | Keyword hits in the CV, portal only | Portal, email, and post read, structured, and backed by source references |
| Scales with | More HR staff | Number of users | Application volume, without new positions |
| Exceptions | Left until someone has time | Drop silently out of the filter | Flagged and sent to a human, nobody is silently discarded |
| Who decides | Human, under time pressure | The filter pre-decides | Human, on 100 percent of the decisions |
| Rollout | Onboarding new colleagues | Configuring filter rules | 2 to 3 weeks to the first productive version |
The honest comparison is the cost of the open position. An unfilled role typically costs 4,000 to 10,000 € per month in lost output, overtime, and postponed orders. And a keyword filter silently discards good candidates that nobody ever sees again.
What we learned about AI in recruiting.
The market reflex is to think of AI in recruiting as a selection machine: 200 applications go in at the top, the best five fall out at the bottom. We consider that the wrong build, legally and above all professionally. What distinguishes good recruiters is not fast rejection but spotting candidates who do not obviously fit on paper. A system that rejects takes exactly those cases out of view.
The bottleneck in mid-sized companies is a different one anyway. There is no shortage of judgement, there is a shortage of preparation time: gathering documents, reading them, making them comparable, chasing people on the phone. Exactly that work can be handed over completely, without a single decision leaving the house. Teams quickly notice that an evidenced overview changes their discussion with the hiring department. They talk about criteria instead of gut feeling.
What it is not suited for: If a position brings five applications, you read them faster yourself. For leadership roles filled through executive search, the decisive part is never in the CV. And anyone who wants an automatic ranking or a filter that rejects is in the wrong place with us. We deliberately do not build that, neither as a feature nor as an option.
