Short answer: AI helps in recruiting mainly with job ads, scheduling, replies and onboarding material. Evaluative steps such as screening applications count as high-risk under the AI Act and need careful handling. The biggest time saving sits in the organisation around hiring, not in the judgement itself.
Skills shortages hit small companies harder than corporations, because nobody there works on hiring full time. The pattern is familiar: a role goes out, the ad reads like ten others, applications sit unopened for two weeks, and the best candidate has already accepted elsewhere. That is exactly the gap where AI is useful, and it is useful in unglamorous places. Where it has no business without a proper assessment is covered here too.
Where the time actually goes
Before choosing a tool, spend a week measuring. In most companies the split looks like this: writing and publishing the ad, answering questions, coordinating interviews, writing rejections, assembling paperwork for day one. Actually assessing people is the smallest slice, and it is the slice you should not hand over.
Bitkom surveyed 604 German companies with 20 or more employees: among those using AI, 14 percent apply it in HR, far behind customer contact at 88 percent (Bitkom press release, 15 September 2025). In the follow-up survey of 14 September 2026, customer contact stands at 72 percent; it gives no figure for HR. The caution has good reasons, but it costs applications.
What the AI Act requires in HR
Annex III of the AI Act lists employment explicitly as a high-risk area. It names, among others, AI systems intended for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates (Annex III EU AI Act).
In practice that means an assistant helping you phrase an ad is a different thing from a system that picks the best ten out of 80 applications. The first is a writing tool, the second an evaluative application with elevated requirements. On top of that, Article 4 requires measures supporting the AI literacy of the people involved (Article 4 EU AI Act). Clarify both with your legal adviser before deployment [check legally]. The wider frame is in EU AI Act: what SMEs actually have to do.
Step by step: what pays off and what does not
| Step | Effort today | With AI | Classification |
|---|---|---|---|
| Draft the job ad | 3 to 5 hours | 45 minutes including editing | Uncritical, output is edited |
| Adapt the ad per channel | 2 hours | 20 minutes | Uncritical |
| Acknowledgements and questions | 10 minutes per application | 2 minutes, template plus approval | Uncritical with approval |
| Interview scheduling | 15 minutes per interview | 3 minutes | Uncritical |
| Screening applications | 5 minutes per application | Technically possible | High-risk under Annex III, assessment required |
| Interviewing and assessing | 60 to 90 minutes | Note-taking support only | Decision stays with people |
| Onboarding material | 4 hours per role | 1 hour | Uncritical |
The three uses with the best ratio
1. An ad that does not sound like every other ad
Do not ask the assistant to write a job ad. Feed it material instead: the actual shape of a working day, three sentences from the person doing the job today, your real conditions. Generate three variants and pick the most concrete. The gain is not time but quality, because specific ads attract fewer but better-matched applications.
2. Response times under 24 hours
An automatically generated acknowledgement that reads personally, states the next step and gives a timeframe costs almost nothing once set up and stops people drifting away. Add two interview slots. How to chain such steps without AI is described in Five processes small teams should automate.
3. Onboarding that does not start from scratch
From existing material you can generate role-specific induction plans, first-week checklists and access lists. It is the least glamorous and most dependable benefit, because no people are being assessed. A wider set of use cases is in AI automation for SMEs.
A worked example: one vacancy, two routes
A trades business is hiring a service technician. The posting attracts 40 applications.
- Today: ad 4 hours, adaptation for two boards 2 hours, correspondence 40 times 10 minutes equals 6.7 hours, scheduling 12 interviews 3 hours, onboarding material 4 hours. Around 19.7 hours in total. At 55 euros fully loaded, 1,084 euros.
- With AI support: ad 0.75 hours, adaptation 0.3 hours, correspondence 40 times 2 minutes equals 1.3 hours, scheduling 0.6 hours, onboarding 1 hour. Around 4 hours, or 220 euros.
- Saving: 15.7 hours and 864 euros per vacancy.
- Across six vacancies a year that is 94 hours and 5,184 euros, against tool costs of roughly 40 euros a month and a one-off setup of 1,500 euros. Net, about 3,200 euros in the first year.
The real value is not the 3,200 euros. It is that the first reply arrives after one day instead of twelve. That is where acceptances are decided.
The ad itself is the biggest lever
In almost every company complaining about too few applications, the cause is the ad rather than the channel. Four things are nearly always missing.
The real working day. Instead of a varied role in a dynamic environment, three sentences describing what a Tuesday actually looks like: when it starts, who you work with, what you go home with.
A salary range. Ads without one get skipped more often, because nobody invests effort in an unknown number. A range tied to experience is enough.
The honest drawback. Every role has one. Shift work, travel, software from the previous decade. Naming it loses the wrong applications early instead of expensively in the third interview.
A concrete next step. Not a portal full of mandatory fields, but a short route: a call, a message or a form with three fields. Everything else belongs in the conversation.
For all four points a language model is a good tool if you give it the raw material. It is a poor tool if it is meant to invent the substance. That distinction decides whether your ad stands out or sounds like one of a thousand.
Four rules to settle in advance
- No automatic rejections. Every rejection is triggered by a person, even where the text is prepared.
- No application documents in personal accounts. Applications contain personal data, sometimes special categories. Approved services with a data processing agreement only [check legally].
- Transparency. State in the process where AI is used in support [check legally].
- Deletion periods. Define when documents disappear from tools and mailboxes, and stick to it.
These points belong in your internal AI policy so they are not renegotiated with every vacancy. For a starting point use the AI check, and run the economics through the ROI calculator. For implementation, reach us through the contact form.
Frequently asked questions
May I let AI pre-screen applications?
Technically yes; legally this is a high-risk application under Annex III of the AI Act with corresponding requirements. For small companies the effort usually outweighs the benefit. Keep the step with people [check legally].
Can applicants tell an ad was written with AI?
They can tell when it is generic. That is why collecting material matters more than the tool: the more specific your input, the less the result sounds prefabricated.
What data may go into an AI tool?
For ads and onboarding material, anything you would publish anyway. For application documents, only in services with the right contract and training switched off, and only as much as necessary [check legally].
Is a dedicated applicant tracking system worth it?
From roughly ten vacancies a year, usually yes, mainly for deadlines and traceability. Below that, a well-run mailbox with templates and a scheduling tool is normally enough.
How do I measure whether it paid off?
Through three figures: time to first response, share of suitable applications and time to fill. Record the baseline before you change anything, otherwise nothing can be evidenced later.
