← Blog HR

Hiring with AI without discriminating: responsible screening in SMBs

Automated screening cuts up to 75% of time. It’s also the HR area with the highest legal risk in 2026. How to gain speed without stepping into a grey zone.

by Miguel Quina Ribeiro· 28 Mar 2026· 2 min read

In an SMB, every bad hire hurts far more than in a large multinational — there’s no redundancy to absorb the mistake. The good news is AI-based screening cuts up to 75% of time in the early stages. The bad news is 23% of organisations using AI in recruiting report compliance challenges, versus only 15% of those that don’t.

The EU AI Act classifies candidate screening as high risk. That changes the conversation. It doesn’t mean you can’t use it — it means you have to use it well.

What can be safely automated

flowchart LR A[Application received] --> B[CV data
extraction] B --> C[Ranking against
objective
requirements] C --> D[Scheduling of
next stage] C --> E{Borderline
match?} E -->|Yes| F[Mandatory
human review] E -->|No| D D --> G([Human decides
who advances])

Three principles that keep the operation clean:

Automate process, not decision. Extraction, matching, scheduling — yes. Final rejection based only on an AI score — no.
Objective, verifiable requirements. Language, certification, years of experience with a system. Never personality traits or "cultural fit" assessed by AI.
Human review in every borderline case. The score orders the queue — it doesn’t close the door.

Where discrimination sneaks in

Systems learn from the past. If you historically hired more men into engineering, the model will prefer CVs from men — even without gender being provided. This bias has cost several well-known companies hundreds of thousands in fines and more in reputation.

How an SMB defends itself:

Quarterly output audit (comparing demographics of candidates advancing vs. those applying).
Don’t give the model name, address or age — work only with role-relevant data.
Document decisions: why candidate A advanced and why B didn’t.

The vendor contract needs three clauses

If you buy an AI recruiting tool, demand in writing:

Explainability — the tool must say why it ranked what it ranked.
No protected-attribute bias — the vendor certifies gender, ethnicity or age are not used as features (directly or by proxy).
Right to audit — you can request output samples for external analysis.

Without those three clauses, it’s your CEO going to court, not the vendor.