An AI-enabled recruiter is a human recruiter using AI to handle the high-volume parts of the job such as screening interviews, initial scoring, scheduling, and shortlist generation. The recruiter reviews the AI's output, applies context the model cannot see, and makes the final call.

The defining characteristics:

• Human oversight is structural, not optional. A recruiter must act on AI outputs before a candidate progresses.
• Every score is explainable. AI recommendations tie to specific evidence the recruiter can interrogate and override.
• AI handles consistency while the human handles context. Deterministic models apply the same criteria to every candidate; the recruiter applies the situational judgment a model cannot.
• Accountability stays with the employer. Because a human reviewed and acted on each recommendation, there is a clear audit trail of human decision-making.

This is the model Hubert is built around. Recruiters receive scored, auditable shortlists directly in their ATS and never autonomous decisions. Every score ties to a specific candidate response, but the final call is always the recruiter's.

An AI recruiter, in the fullest sense of the term, is an autonomous agent that performs hiring tasks end-to-end with no meaningful human checkpoints. It may source candidates, contact them, score interviews, and advance or reject applicants. Any human involvement tends to be supervisory; reviewing what the system decided rather than participating in the decision.

Where the risk concentrates:

• Accountability becomes diffuse. When an autonomous system rejects a candidate, who made that decision? The vendor? The employer? The model? Regulators are starting to answer that question, and the answers are uncomfortable for early adopters.
• Errors compound at scale. A human reviewer catches model errors before they affect large numbers of candidates. Remove the reviewer, and a single miscalibration can process thousands of applications before anyone notices.
• Candidate trust erodes. Candidates who discover an automated system handled their application with no human ever involved consistently report lower trust and lower satisfaction. In a tight talent market, employer-brand damage is real and measurable.
• EU AI Act exposure is material. Systems that make or substantially influence employment decisions are high-risk under the EU AI Act. Autonomous rejection at scale, with no human review, is a compliance risk most enterprise legal teams are not yet equipped to absorb.

None of this means autonomy in recruiting has no use. Outreach, scheduling, and FAQ-handling are lower-stakes tasks where autonomy is reasonable. The risk concentrates at the assessment and shortlisting stage; precisely where the decision most affects the candidate's future.

The easiest way to think about the difference is across six dimensions: who decides, how reviews happen, how scores are explained, the regulatory profile, the candidate experience, and the right use cases.

AI-enabled recruiter (augmented model)

• Who makes the final call: the human recruiter.
• Human review of each candidate: required before the candidate progresses.
• Score explainability: every score ties to a specific candidate response and criterion.
• EU AI Act risk profile: manageable with structural human oversight is built in.
• Candidate experience: a recruiter is accountable for the outcome and the audit trail.
• Best fit for: assessment, shortlisting, and decision-stage tasks.

AI recruiter (autonomous model)

• Who makes the final call: the system.
• Human review of each candidate: often retrospective, batched, or absent.
• Score explainability: often opaque, or generated after the fact by the model.
• EU AI Act risk profile: high and harder to defend at audit, particularly at scale.
• Candidate experience: no human is ever involved in the individual decision.
• Best fit for: lower-stakes tasks like outreach, scheduling, and FAQ-handling.

Both models can be fast. Only one of them is fast in a way enterprise legal, compliance, and TA teams can actually defend.

In practice, most enterprise deployments fall somewhere between the two extremes. It helps to think of recruitment AI as a spectrum with four reference points:

1. Assisted. AI drafts content, surfaces information, or suggests next steps. A human initiates every action. Example: an AI-drafted job description a recruiter edits and approves.
2. Augmented. AI handles defined tasks (screening interviews, initial scoring, shortlist generation) and presents structured outputs to a human who acts on them. The workflow is designed around human review. This is where Hubert operates.
3. Supervised autonomous. AI makes decisions across most of the process; humans review outputs in batches or at exception points. Risk grows as batch size grows and review becomes perfunctory.
4. Fully autonomous. AI acts end-to-end. Human involvement is governance-level, not workflow-level. Currently rare in enterprise hiring; more common in high-volume contingent labor.

Most TA leaders thinking carefully about AI adoption land in the augmented zone. Not because they lack ambition, but because they understand that speed and defensibility are not in tension at that level. The augmented model is fast enough to deliver the up to 80% reduction in screening time Hubert customers see, while keeping the human accountability that legal, compliance, and talent teams actually need.

Much of the anxiety about AI in hiring is rooted in a legitimate concern: if you cannot explain why the AI scored a candidate the way it did, you cannot defend that score to the candidate, to a regulator, or to your own hiring manager.

That concern is valid for probabilistic, LLM-based scoring; the same input can produce different outputs and the reasoning is opaque by design. It is not valid for deterministic AI models, which are architecturally different.

Hubert's assessment layer uses deterministic models: same input, same output, full explainability. Every score is tied to a specific candidate response, assessed against pre-defined competency criteria. There is no black box. There is no probabilistic drift. A recruiter reviewing a shortlisted candidate can see exactly why that candidate scored what they scored; and so can an auditor, a legal team, or the candidate themselves.

This is why the augmented model only works if the AI doing the augmenting is explainable. A shortlist a recruiter cannot interrogate does not augment their judgment; it replaces it with something they are forced to trust without basis. That is the autonomous model dressed in augmentation language, and TA leaders should push vendors hard on the difference.

The Candidate Pledge that underpins Hubert makes the commitment explicit: explainability is a human right. Every candidate deserves to know, at least in principle, how their performance was assessed. That standard holds regardless of hiring volume.

Before adopting any AI recruiting tool, get clear answers to these five questions:

1. Where is the human checkpoint? At what stage does a human review AI outputs before a candidate progresses? If the answer is "we review in aggregate" or "we review exceptions," the system is operating autonomously at the individual candidate level.
2. Can we explain every score? Not in general terms; specifically. If a candidate asks why they scored 62% on a communication competency, can you show them the response that generated the score and the criteria it was assessed against?
3. Is the model deterministic or probabilistic? This is a technical question vendors should answer directly. Probabilistic models are not inherently bad, but they require additional explainability infrastructure to be defensible at audit.
4. How does this interact with our ATS and existing compliance obligations? Integration quality matters as much as model quality. Shortlists that arrive outside the ATS create parallel records and compliance gaps.

What happens to candidates who are not shortlisted? Do they receive feedback? Is there a meaningful appeals path? A system that processes thousands of candidates and returns silence to the majority is a candidate experience and brand risk, regardless of how good the shortlists are for recruiters.

The augmented model, done well, is not a compromise position. It is the model behind the numbers TA teams actually quote on results decks:

• ManpowerGroup: 67% reduction in screening time for recruiters.
• Hemfrid: 15,000+ applications managed annually with 75% less manual screening effort, and 90% accuracy in predicting successful hires.
• Malmö Stad: a recruitment cycle from six weeks to two days.
• Ambea: one central recruiter supporting 3,000 hires and 100,000+ applications, across 500 care units and 1,200 hiring managers.

None of these required handing decisions to autonomous systems. They came from giving skilled recruiters the right tools and the right information at the right moment, in a legally defensible manner.

The choice between AI recruiter and AI-enabled recruiter will not resolve in favor of one model. Different hiring contexts will continue to sit at different points on the spectrum. What matters right now is being deliberate about where your program sits; and why.

If you are evaluating where your hiring program sits on the automation spectrum, or building a business case for structured AI interviewing, we would welcome the conversation.

Book a demo with the Hubert team to see how the augmented model works in practice, and what scored, auditable shortlists look like in your ATS.