Gartner has published two forecasts that should be read together. By the end of 2026, forty percent of enterprise applications will include task-specific AI agents. And by the end of 2027, over forty percent of agentic AI projects will be cancelled, due to escalating costs, unclear business value or inadequate risk controls.
Both predictions will prove correct. The interesting question is what separates the projects that survive from the projects that die. Nowhere will that question be answered faster than in recruitment, because recruitment is where agentic AI is being adopted first and where the cost of getting it wrong is highest.
Gartner describes the destination as small teams of specialised agents supervised by a recruiter. Call it the agentic recruiter. Whether that model works will depend far less on the agents than on the quality of the supervision.
Why recruitment is the proving ground
Recruiting is now the single most common enterprise AI use case in HR. Adoption of AI across HR tasks roughly doubled in a year, and the measurable wins are real. Organisations using AI across the hiring workflow report time-to-hire reductions of a third or more. Screening that took days happens in hours. Job descriptions that took a week of back-and-forth are drafted in minutes.
The risks are just as measurable. Nearly one in five organisations using automated screening admit their tools have screened out qualified applicants. Keyword-based matching misses an estimated thirty to forty percent of qualified candidates simply because they describe their skills differently. A hiring decision is also one of the most regulated decisions a company makes. The EU AI Act classifies recruitment AI as high risk. GDPR Article 22 and UAE PDPL Article 18 both restrict solely automated hiring decisions. Recent litigation in the United States has made clear that AI vendors and the employers who deploy them can share liability for discriminatory outcomes.
So recruitment concentrates everything that makes agentic AI valuable and everything that makes it dangerous, in one workflow, at volume.
The real failure mode is not the model
Most analyses of failed AI projects focus on model quality. In hiring, that is rarely the cause. The models are good enough to score a CV against a structured rubric. What kills these systems is trust, and it fails in two opposite directions.
When recruiters do not trust the system, they override everything. The scores get ignored, the shortlists get rebuilt by hand, and within a quarter the expensive AI platform has quietly become a job-posting tool. Nobody announces the failure. The value simply evaporates.
When recruiters trust the system too much, they rubber-stamp everything. That is worse. European regulators have stated plainly that token human involvement does not count: oversight of an automated decision must be meaningful, carried out by someone with the authority and competence to change the decision. A reviewer who spends two seconds per application is not a human in the loop. They are a legal liability wearing one’s badge.
The override button as an instrument
This is why I have come to believe the most important feature in any AI recruitment system is not the matching algorithm. It is the override button, and more precisely, the telemetry behind it.
A healthy human-AI hiring workflow shows an override rate somewhere between ten and twenty percent. That number says the humans are genuinely engaging with the machine’s recommendations and genuinely disagreeing some of the time, which is exactly what expert judgment looks like. An override rate near zero signals rubber-stamping and should trigger an investigation, not a celebration. An override rate near one hundred percent signals the system has already failed, whatever the dashboard says.
Treated this way, the override stops being an escape hatch and becomes an instrument.
Every override, captured with a reason, is a labelled data point about where the system and the expert diverge. That stream is the most valuable tuning signal the system will ever receive. It is also the audit trail that answers the question every regulator, and increasingly every candidate, is entitled to ask: who decided, and on what evidence?
Evidence is what earns the trust
A score without rationale demands faith. A score with cited evidence invites judgment. That distinction decides whether recruiters engage or disengage.
When a system says a candidate scored well and shows the specific lines of the CV, the specific interview answers and the specific rubric criteria behind that assessment, the recruiter can agree or disagree in seconds. The human is doing what humans are good at, which is judgment, and the machine is doing what machines are good at, which is consistent, evidence-linked evaluation at volume. Structured, rubric-based assessment roughly doubles predictive validity compared with unstructured interviews, and the research suggests standardised formats can cut bias effects by half. The machine did not replace the expert. It gave the expert better raw material.
There is a second reason evidence matters more every month. Candidates are using AI too. CVs are increasingly tailored by AI to match job descriptions, which weakens every keyword-based filter at once. As the surface signals inflate, the only reliable filter left is structured assessment against evidence: what did this person actually do, decide and deliver. Both sides of the hiring table are moving up the stack.
What good looks like
From everything I have seen building and running enterprise systems, the AI recruitment deployments that will still be operating in three years share a small set of design decisions. The AI drafts, ranks, explains and recommends. A human decides at every consequential gate, and the system makes that easy rather than theatrical. Nothing is silently auto-rejected. Every score carries its evidence. Override rates are monitored as a first-class health metric. And accuracy is measured continuously against expert judgment, not assumed from a launch-day benchmark.
None of this slows the process down. The time savings come from automating the grind: drafting, posting, parsing, scheduling, summarising. The judgment stays where it always belonged. The agentic recruiter is not a diminished recruiter. The role shifts from doing the work to owning the decisions, which is a promotion, not a threat.
Recruitment is about to become the reference case for agentic AI in the enterprise, in both directions. The organisations that treat human oversight as friction to be engineered away will supply the cancellation statistics. The ones that treat trust as the core of the architecture will quietly compound an advantage in the scarcest market of all, which is talent. The difference between the two is not the sophistication of their models. It is whether they balanced the speed of their agents with the judgment of the humans supervising them. Getting that balance right is the real work of the agentic recruiter, and of everyone building systems for one.
Fayyaz Alam is the Founder and CEO of Zirora, and a technology executive with three decades of experience building and running enterprise systems, including leading software engineering for one of the world’s largest aviation groups.
Get in touch