Illustrative redesign · A real 2017 hiring round, rebuilt for 2026
Then and Now: Keeping People in Charge of Hiring, with AI Only for the Paperwork
How I'd rebuild a hiring round I once ran by hand, with AI doing the reading and people making every decision
Please read this first
The 2017 hiring round is real, and no AI was used on it: every application was read by a person. The AI workflow is a design study: it has not been built or used in a live hiring round, and it is not client work. No applicant's name or details appear here. This page explains the rules for learning purposes and is not legal advice.
In 2017 I ran the hiring for my mother's tutoring business, advertising two teaching roles and sorting every application by hand. Here I take that real hiring round and redesign it as an AI-assisted workflow that would comply with the EU AI Act and GDPR if the business operated in Germany today.
The 2017 round shows where the time and the risks really were. The redesign shows how to fix them within the rules. The key design decision is to keep the AI out of the hiring decision, which can keep a small business out of the AI Act's high-risk category.
What the AI never does
In this design, people make every decision about a candidate. The AI never:
- Scores, ranks or shortlists candidates
- Recommends who to interview, hire or reject
- Rejects anyone
- Analyses interviews, video, voice or facial expressions
- Sees a candidate's name, photo or age
- Contacts a candidate on its own
At a glance
| Details | |
|---|---|
| The business | A small, family-run tutoring business |
| Roles hired | A Commerce teacher and an IB/IGCSE teacher |
| Then (2017, real) | I advertised two teaching roles on Indeed and read, compared and sorted every application by hand |
| Now (2026, illustrative) | The same hiring round, redesigned as an AI workflow for a small business in Germany today |
| My role | Then: I ran the hiring for the family business. Now: AI governance engineer, designing the workflow and its compliance |
| Rules applied | EU AI Act, GDPR, German anti-discrimination law (AGG) |
| Outcome | A hiring workflow where AI saves time on reading and admin, and people make every decision |
| Status | Illustrative redesign, not yet used in a live hiring round |
Then: hiring by hand in 2017
A small business owner hiring for several roles at once spends most of their time reading applications, and very little choosing between good candidates.
We advertised two teaching roles on Indeed. Applications arrived as email notifications with CVs attached in different formats (PDF and Word). Each one had to be opened, read and compared by hand against what the role needed: subject knowledge, curriculum experience such as IB or IGCSE, and location.
The pain points:
- Time: reading every CV in full, even clearly unsuitable ones.
- Inconsistency: without written criteria, each application was judged slightly differently.
- Hidden bias: names, photos and personal details were visible during screening.
- Data left behind: applications, including CVs from people who weren't hired, stayed in the inbox for years.
Today, the obvious fix would be to let an AI tool read and rank the CVs. In the EU, that is exactly where the legal difficulty starts.
Now: why AI in hiring is a hard case
Hiring is one of the few everyday business uses that the EU AI Act treats as high-risk, because decisions about who gets a job deeply affect people's lives.
| Rule | What it says | Applies from |
|---|---|---|
| AI Act, Annex III point 4 | AI used to recruit or select people, for example to place targeted job ads, analyse and filter applications, or evaluate candidates, is high-risk | 2 Dec 2027 |
| AI Act, Art. 6(3) | An Annex III system is not high-risk if it only does a narrow procedural or preparatory task and doesn't materially influence the decision, but never if it profiles people | 2 Dec 2027 |
| AI Act, Art. 5 | Emotion recognition in the workplace is banned | Since Feb 2025 |
| GDPR, Art. 22 | People have the right not to be subject to a decision based solely on automated processing that significantly affects them | Since 2018 |
| GDPR, Art. 35 | A data protection impact assessment is needed for high-risk processing, such as new technology used on applicants' data | Since 2018 |
| AGG | Applicants must not be disadvantaged because of gender, ethnic origin, religion, disability, age or sexual identity; rejected applicants can claim within two months | In force |
Why this matters now: after the Digital Omnibus on AI (Regulation (EU) 2026/1744), the high-risk rules for hiring apply from 2 December 2027. Any business planning to use AI in recruitment needs to decide over the coming year where the AI sits in its process, because that decision sets which rules apply.
What this means for a small business: an AI tool that scores, ranks or rejects candidates makes it a deployer of a high-risk AI system. If the owner builds the tool themselves, they also become its provider, with the heaviest set of obligations in the Act. For a tutoring business hiring a few teachers a year, that burden is out of proportion.
The key design decision: keep the AI out of the decision
The workflow lets AI do the reading and admin, and leaves every judgement about a person to a person. That keeps it within the AI Act's Art. 6(3) exemption for preparatory tasks, while still saving most of the time.
| AI may | AI may not |
|---|---|
| Draft the job ad text for the owner to edit | Decide who sees the job ad (targeted ad placement is high-risk) |
| Remove names, photos, age and other personal markers from CVs | Score, rank or shortlist candidates |
| Extract facts into fixed fields: subjects, curricula, years of teaching, qualifications, availability | Recommend who to interview or reject |
| Draft interview invitations and polite rejection emails for a person to send | Analyse video, voice or facial expressions in interviews |
| Log what it did, for review | Send any decision to a candidate on its own |
Why the line sits here. Extracting facts that a candidate wrote is a preparatory task. Evaluating whether those facts make someone a good teacher is profiling, and profiling removes the exemption. The extraction prompt therefore reports facts only, with no ratings, adjectives or comparisons, and the owner always sees the original CV alongside the summary.
Documenting it. A provider who relies on the Art. 6(3) exemption must document the assessment (Art. 6(4)) and register the system in the EU database (Art. 49(2)). The Omnibus simplified what that registration must contain but kept the duty. The assessment should be checked by a lawyer, because where "fact extraction" ends and "evaluation" begins is not yet settled.
The redesigned workflow
The workflow has nine stages. AI helps at four of them, always with reading or writing, and people make every decision about a candidate.
1 Job ad written (AI assists)
The owner writes the criteria first; AI drafts the ad text, the owner edits it · Mistral
2 Applications arrive
Form with a privacy notice that explains the AI step; talent pool only with consent · n8n
3 Personal details removed (AI assists)
Names, photos, age and addresses removed before anyone screens the CV · Presidio
4 Facts extracted (AI assists)
Subjects, curricula, years of teaching, qualifications; no scores or rankings · Flowise, Mistral EU
5 The owner screens every application
Against the written criteria, with the anonymised summary and the original CV side by side
6 Interviews
Run by people; no AI video or voice analysis (emotion recognition at work is banned, Art. 5)
7 Decision and reply (AI assists)
A person decides; AI may draft the email, a person checks and sends it
8 Data deleted after 6 months
Rejected applicants' data removed once the claim period is over · scheduled n8n job
9 Logged and reviewed monthly
Every AI step recorded; checked for errors and unfair patterns · Langfuse
The AI never sees a candidate's name or photo, and never produces a score. The owner still reads every application, but from a short, consistent summary instead of a stack of CVs in different formats.
Then and now, side by side
Each pain point from the 2017 round maps to a specific fix in the redesign.
| Then: the 2017 pain point | Now: how the redesign handles it | Stage |
|---|---|---|
| Time: reading every CV in full | AI turns each CV into a short summary with the same fields; the owner still reads every application, just faster | 3, 4 |
| Inconsistency: no written criteria | Criteria are written down before the job ad goes out, and every summary uses the same fields | 1, 4 |
| Hidden bias: names, photos and personal details visible | Names, photos, age and other personal markers are removed before anyone screens | 3 |
| Data left behind: CVs kept in the inbox for years | Rejected applicants' data is deleted 6 months after the rejection; a talent pool only with consent | 8 |
Compliance map
Every requirement maps to a specific stage of the workflow, so compliance is part of the design rather than a document written afterwards.
| Requirement | How the workflow meets it | Stage |
|---|---|---|
| AI Act Annex III / Art. 6(3): stay out of high-risk | AI only drafts, anonymises and extracts facts; no scoring, ranking or recommending; assessment documented and registered (Art. 6(4), 49(2)) | 1, 3, 4 |
| AI Act Art. 5: no emotion recognition at work | Interviews are run by people, with no video or voice analysis | 6 |
| AI Act Art. 4: AI literacy | The owner gets short training on what the AI does and doesn't do, recorded in a simple training log | Set-up |
| GDPR Art. 6: lawful basis | Steps taken at the applicant's request before a possible contract (Art. 6(1)(b)) | 2 |
| GDPR Art. 13: transparency | The application form explains how data is used, including the AI extraction step, and for how long it's kept | 2 |
| GDPR Art. 22: no solely automated decisions | A person reviews every application and makes every decision; the AI cannot reject anyone | 5, 7 |
| GDPR Art. 28 and Chapter V: processors | Mistral API through its EU endpoint, data processing agreement in the terms, training on inputs switched off | 3, 4 |
| GDPR Art. 35: impact assessment | A data protection impact assessment is carried out before launch, as AI is used on applicants' data | Set-up |
| AGG: no discrimination | Screening uses anonymised summaries; criteria are written down before the job ad goes out | 1, 3, 5 |
| GDPR Art. 5(1)(e) and AGG § 15(4): retention | Rejected applicants' data is deleted 6 months after the rejection (the 2-month claim period plus 3 months to sue under § 61b ArbGG, plus a buffer); a talent pool only with separate, active consent | 8 |
| Oversight and accountability | Every AI step is logged; the owner reviews the log monthly for errors or unfair patterns | 9 |
The workflow uses open-source, EU-friendly tools: n8n to run it, Microsoft Presidio to remove personal details, Flowise with the Mistral API for the extraction, and Langfuse for the log.
What changes if the AI ranks candidates
One extra feature, letting the AI score or shortlist applicants, would turn this into a high-risk AI system, and the obligations multiply. From 2 December 2027:
| As the provider (whoever builds it) | As the deployer (the business using it) |
|---|---|
| Risk management system (Art. 9) | Use it according to its instructions (Art. 26(1)) |
| Data governance and bias checks on the data (Art. 10) | Human oversight by trained, competent people (Art. 26(2)) |
| Technical documentation (Art. 11) | Make sure input data is relevant and representative (Art. 26(4)) |
| Automatic logging (Art. 12) | Keep the logs for at least six months (Art. 26(6)) |
| Instructions and transparency for deployers (Art. 13) | Tell candidates a high-risk AI system is used in decisions about them (Art. 26(11)) |
| Built-in human oversight (Art. 14) | Give an explanation of the decision on request (Art. 86) |
| Accuracy, robustness and cybersecurity (Art. 15) | |
| Quality management system (Art. 17), conformity assessment and EU database registration |
For a small tutoring business, that is a compliance project in its own right. The design decision here avoids it while keeping most of the time saved. For what the high-risk route looks like when it can't be avoided, see the CV-screening scenario in my illustrative scenarios.
What this shows
- Hands-on experience shapes the design. Having screened every application myself, I knew where the time actually went, which is why the AI takes on reading and admin, not judgement.
- Compliance is an architecture decision. Where the AI sits in the workflow decides which rules apply. Choosing that early is cheaper than fixing it later.
- The hard cases are where the value is. Hiring is high-risk, so getting it right shows how to handle any sensitive workflow.
- Small businesses can use AI safely, with free, open-source tools and a clear line between what the AI does and what people decide.
Limitations. This is a design study based on a real hiring process in India in 2017, re-imagined in Germany today. It has not yet been built or used in a live hiring round. The natural next step is a pilot on the business's next hiring round, which would turn it into a measured, real-world case study. Whether fact extraction counts as a preparatory task under Art. 6(3) is a judgement that should be checked by a lawyer before use.
For real client work, see the in-house AI support chatbot case study.
Planning AI in your hiring?
The high-risk rules for recruitment apply from December 2027. If you want to work out where AI can sit in your process, and which rules that brings, I'd be glad to talk.
Vinita Silaparasetty, AI Governance Engineer