Agentic AI in recruitment: 4 scientific guardrails before delegating pre-qualification

After generative AI comes agentic AI: autonomous agents that screen applications, run voice pre-interviews, follow up with candidates and push a shortlist to the recruiter. The promise is attractive: processing in hours what used to take weeks. But between saving time and delegating judgement, the line is thin. Under what conditions does an AI agent improve a hiring process instead of degrading it?

What agentic AI actually changes

An AI agent no longer just answers a question: it chains actions towards a goal. Applied to recruitment, this already covers sourcing, CV pre-screening, interview scheduling and, increasingly, conversational exchanges with candidates.

On logistical tasks, the gain is real and hardly contested. The difficult question starts when the agent produces an evaluation: a candidate ranking, a fit score, a recommendation to reject a profile. At that precise moment, the agent is no longer doing logistics, it is doing measurement. And measurement has rules.

Guardrail 1: demand a published measurement framework

When an AI agent ranks candidates, the first question is not "which model does it use?" but "what does it measure, and against what framework?". A ranking without an explicit framework is an automated opinion: fast, systematic, and unverifiable.

A serious measurement framework defines the competencies assessed, their operational definitions and the associated observable behaviours. This is exactly what Rising Up publishes with its framework of 18 behavioural competencies: three families, consultable definitions, described behaviours. The assessment mechanics remain proprietary, the scientific framework is public. If your agent vendor cannot produce the equivalent, you are not buying an assessment, you are buying a black box.

Guardrail 2: separate what is inferred from what is measured

A conversational agent can infer many things from an exchange: fluency, vocabulary, responsiveness. The risk is confusing these surface signals with underlying competencies. The smoothness of a conversation with a voice agent measures neither rigor, nor perseverance, nor the ability to decide under uncertainty.

A behavioural assessment worthy of the name places the person in concrete professional situations and analyses their responses, rather than inferring a profile from the style of the exchange. That is the principle of the Soft Skills Scan: contextualised scenarios, combining several assessment formats, backed by a verifiable framework. Conversational inference can complement a structured measure; it cannot replace it.

Guardrail 3: keep a human who decides, not a human who rubber-stamps

The European AI regulation classifies recruitment systems as high-risk and requires genuine human oversight. But beyond the legal obligation, there is a cognitive reason: a human who merely endorses an agent's outputs is no longer exercising oversight, they are exercising a reflex.

The difference plays out in the tool's design. A well-designed system presents its results as a decision aid: a score, an analysis, explicit limitations, and the hand left to the recruiter. That is Rising Up's stance: the Scan informs the manager's decision, it never takes it in their place. An agent that automatically rejects candidates without effective human review does exactly the opposite, and you carry the legal and ethical responsibility for it.

To dig deeper into this ridge line between assistance and substitution, our analysis of AI and recruiters details why tooled human judgement remains the most robust model.

Guardrail 4: audit the agent as you would audit a test

A serious assessment test is judged on established criteria: what does it measure, with what consistency, with what documented biases? An AI agent that evaluates candidates must undergo the same examination, with an added difficulty: its answers can vary from one interaction to the next and its behaviour can drift with updates.

Three minimal audit questions before any deployment:

• Traceability: can you reconstruct why the agent ranked this candidate at this level?
• Consistency: does the same profile, assessed twice, obtain a comparable result?
• Bias: does the vendor document how it detects and corrects biases in its data and models?

These requirements join those we detail in our article on the AI Act and recruitment tests: the European regulatory framework turns these good practices into obligations.

AI tools the judgement, it does not replace it

Rising Up's position is constant: artificial intelligence is a formidable measurement and analysis tool in the service of human judgement, not a substitute for that judgement. Our proprietary CoreSkills AI analyses responses in situation and produces scores and recommendations; decisions remain human.

This position is not conservatism, it is a scientific conclusion. Behavioural skills predict professional success, contribute to it causally, and can be developed (research by James Heckman, Nobel laureate in economics). Stakes of this weight deserve rigorous measurement AND a decision owned by someone who can answer for it. The market is moving in this direction: according to APEC, 41% of companies recruiting executives now formally assess soft skills, up 9 points in one year. The question is no longer whether to automate, but to automate what should be automated, and to measure what should be measured.

Recognised as Deeptech by the European Innovation Council (EIC) in 2023, Rising Up integrates AI where it excels: the standardised analysis of responses in situation. To see how this approach fits into a complete hiring process, discover our recruitment test grounded in cognitive science.