AI in Recruitment 2026: What Actually Works (and the Legal Risks)

AI adoption in HR went from 26% to 43% in a single year. The question is no longer whether to engage with it, but how to separate what genuinely produces results from what adds legal exposure without improving quality.

What changed in twelve months

Moving from 26% to 43% in a year is not ordinary growth; it is the shift from pilot to standard. What was an innovation project in 2023 has become, for nearly one HR function in two, a production tool.

The second change is qualitative. The conversation is now about agentic AI: systems that chain steps together — source, screen, follow up with a candidate — without a human trigger at each transition. That moves the question from “time saved” to “control”.

What actually works: three measurable uses

1. Broader sourcing

This is the most mature and least risky use. The AI does not decide; it widens the search perimeter — reformulating queries, surfacing adjacent profiles, spotting equivalent skills hidden behind different job titles. The reported gains apply to a stage where a human still makes the call.

2. Skills-based matching

Matching on skills rather than on degree or job title reports predictive accuracy around 78%. That is meaningful — and it is also a ceiling worth reading honestly: a prediction that is right 78% of the time is wrong one time in five. It should prioritise a shortlist, not compose one on its own.

3. Recruitment administration

Interview write-ups, follow-ups, scheduling: this is where the return is clearest and the exposure lowest, because no rejection decision is at stake.

What does not work yet

Three uses are sold well beyond what they deliver. Automated video-interview analysis (expression, tone, pace): the scientific literature does not support the claimed predictive validity, and discrimination exposure is high. Personality scoring from free-text analysis. And individual attrition prediction, which mostly produces managerial side effects when it leaks.

Legal exposure in Morocco: law 09-08 and algorithmic bias

Two issues overlap. First, data protection: processing applications falls under law 09-08 and the oversight of the CNDP. A tool that stores CVs outside Morocco, or reuses candidate data to train its models, engages your liability as data controller — not the vendor’s.

Second, bias. A model trained on your past hiring reproduces your past patterns, including those you would not defend explicitly. If your history favours certain schools, certain neighbourhoods, or one gender for a given role, the model learns that filter and applies it at scale, with the appearance of neutrality.

The countermeasure is methodological, not technological: define the assessment criteria before looking at candidates, and make them objective through instruments that measure capability rather than pedigree — which is precisely what an assessment center is for.

How to frame a pilot around the right KPI

1. Pick a stage, not a process. A sourcing pilot can be evaluated; a pilot “on recruitment” cannot.
2. Set the KPI on quality, not speed. Confirmation rate at end of probation and at twelve months — not CVs processed.
3. Keep a control group. Without comparison you are measuring a hiring season, not a tool.
4. Document rejection criteria. If you cannot explain why a candidate was screened out, you cannot defend it.
5. Train recruiters on the tool. A recruiter who cannot contradict the machine is not using it — they are subject to it.

That last point is the most underestimated: the value of a screening tool depends entirely on the recruiter’s ability to argue with its output.