Recruiting with KI - a practical guide

Current KI models do three things reliably in recruiting: score applications against a requirements profile, draft responses to standard mails, and parse resumes into structured fields. Not spectacular, but real time-savers.

What they don't do reliably: judge cultural fit, spot 'red flags' in unstructured info, or make a hire decision. Using KI for that builds systematic discrimination risk into the process.

Article 22 GDPR bans automated individual decisions with legal effect without human review. KI may pre-sort, it may not decide. Concretely: KI gives a score, a human reads it, a human decides.

The EU AI Act (in force since August 2024) names recruiting KI explicitly as high-risk in Annex III No. 4; the high-risk duties apply from 2 August 2026. That means: documentation, discrimination testing, candidate-facing transparency, audit logs. Tools that take this seriously make it visible by default - not as an add-on. For the GDPR side of those duties, see the GDPR checklist for recruiting 2026.

Since February 2025 a hard ban runs in parallel (Art. 5 AI Act): emotion recognition in video interviews via facial analysis, social scoring from unrelated sources, and biometric categorisation are forbidden in hiring - whatever the tool. Fines are two-tier: up to €35M or 7% of global annual turnover for prohibited practices, up to €15M or 3% for high-risk breaches. Not a residual risk - a reason to look hard when buying a tool.

Honestly counted: KI pre-sorting doesn't pay from the first role but from volume. In studies on German mid-market firms the ROI turns positive at roughly 25-30 hires per year, with payback around nine months. Below that the gain is quality (a second, even-energy reading voice) more than time.

The numbers once volume is there: time-to-hire drops 30-40% in practice, cost-per-hire 20-35%, and early attrition up to 25%, because fewer good people are missed in the energy dip of the list. Audi publicly reports -30% time-to-hire from KI-assisted hiring. Crucially: these effects come from faster reaction to good applications, not less care. If you're still weighing whether a dedicated ATS is the right step at all, the overview on applicant-management software helps.

We recommend starting with five concrete steps. No 'KI transformation project', no quarterly roadmap, just five tasks you can do in one afternoon. For deeper guidance on writing and calibrating the screening prompt from the start, see Setting up KI screening correctly.

Three things: no auto-rejection without human review. No cultural assessment from photo, name, or language. No hidden scoring not disclosed to the applicant.

These limits aren't politeness, they're legally binding. Skipping them risks lawsuits, anti-discrimination claims, and long-term brand damage.

The transparency duty sounds heavy but is one paragraph. Place it visibly (not in the small print) in the job ad and the receipt email. A proven snippet:

"To pre-sort applications we use a KI-assisted tool that produces an initial estimate of professional fit. The selection and hiring decision is made solely by our HR team. You have the right to object to the KI-assisted pre-assessment; in that case we review your application without the tool."

That covers AI Act transparency, GDPR information, and the right to object in one paragraph. KI BMS shows such a notice in the careers-page default - you only fill in the exact retention window.