Something has shifted in client conversations over the last two quarters.

AI agents are no longer being discussed only as pilots. They are already being used across sourcing, screening, scheduling, assessments, and candidate engagement.

The speed gains are real. But speed is still where much of the conversation begins and ends.

I think that is the wrong question.

The more important issue is whether the hiring process becomes more reliable as it becomes faster.

I have spent close to two decades across consulting and talent businesses, and the last ten of those years inside recruitment technology: at Findem, at foundit where I ran consumer businesses across the Monster India and APAC markets, and now building the Agentic AI business at SelectPrism, where we are contracted to run close to 80,000 AI interviews this year and have completed more than 15,000 already.

In that time I have watched several waves of technology arrive with the promise of fixing hiring. Each wave did reduce manual effort somewhere in the cycle. None of them fixed the harder problems underneath: weak candidate signals, fragmented systems, inconsistent assessments, poor coordination, and too much dependence on individual judgement.

Aptitude Research put a number on this last year: 82 percent of companies report significant gaps in their applicant tracking systems, even as a third keep increasing spend on that infrastructure. That matches what I have seen from the inside. Two decades of spending have made hiring more efficient. The reliability problem is still open.

AI can remove some of that friction. It can also amplify it.

The Signals We Built Hiring On Are Losing Their Meaning

Recently, I spent an hour with talent acquisition practitioners discussing what is changing inside enterprise hiring teams.

One issue came up repeatedly: the information used to evaluate candidates is becoming harder to interpret.

Resumes are the clearest example. Across the applications we process, 60 to 80 percent can appear to be a strong match for the job description.

When we look closely, there are two kinds of AI work candidates do on a resume. One improves the language and the structure, and I have no problem with that. The other rewrites the resume to mirror the job description itself, so it scores well on whatever matching system sits on the other side. The first makes a candidate easier to read. The second makes every candidate look the same.

Neither means the person lacks the required capability. It does mean the resume, by itself, carries less signal than it once did.

Candidate behaviour is shifting in parallel, and I would argue this is the more consequential change. The same AI that helps recruiters operate at scale helps candidates apply at scale. A motivated candidate can now identify relevant roles, tailor an application for each one, rehearse likely questions, and run several hiring processes at the same time.

We see individual candidates applying to 100 or 120 positions in a single day, often through agents that apply the moment a role opens. And because candidates can hold multiple offers in parallel, offer to join ratios in some segments have slipped to 60 or 70 percent. When volume rises that sharply, the application stops telling you much about intent. Genuine interest and casual interest look identical on paper.

Underneath all of this, the skills being hired for are moving too. IBM has estimated the half-life of technical skills at roughly two and a half years, and in areas close to AI I would say parts of the stack turn over even faster. Prompt engineering was a premium skill two years ago. Today it is table stakes.

Which is why hiring teams increasingly find themselves evaluating candidates for roles the market has not finished defining. Forward-deployed engineers, RAG architects, evaluation engineers, AI governance specialists: 18 months ago some of these titles did not exist, and today they are among the hardest positions to fill. Yet few organizations have screening rubrics or panel training for them. The old playbook was written for roles that sat still.

Put these shifts together and the real challenge comes into focus. It is no longer simply finding more candidates. It is establishing which evidence is credible and which capabilities are genuinely relevant.

What Enterprise Talent Leaders Are Wrestling With

The poll responses reflected this shift.

Participants expected Talent Acquisition to be among the HR functions most affected by AI agents over the next two to three years. Screening and assessments were seen as the clearest starting point for greater autonomy.

There is a practical reason for that. These stages consume considerable recruiter time, and they involve large volumes of information that AI can process effectively.

They are also the stages where mistakes carry the greatest consequence.

Scheduling an interview is primarily an operational decision. Screening out a candidate or interpreting an assessment requires judgement. The system is influencing who progresses and who does not.

Participants also pointed to a readiness gap around AI-native roles. Software engineering was identified as the role changing fastest because of AI, while many organizations said they were still at an early stage in defining how to hire for newly emerging roles.

The discussion focused on a few practical concerns.

How should recruiters compare candidates when resumes have been extensively refined using AI? What makes an AI-led interview rigorous enough to support a hiring decision? How should hiring teams detect unauthorized assistance, impersonation, or synthetic identities? How should agents connect with existing HR systems without adding another layer of complexity?

These questions have a common thread. Organizations need to understand the basis of an AI recommendation before they are prepared to act on it.

Trust Is the New Constraint on Autonomy

AI is changing both sides of the hiring process.

Employers can source and screen candidates at greater scale. Candidates can apply and prepare at greater scale. Hiring teams can automate assessments, while candidates can use AI to improve or generate responses.

Fraud is becoming more sophisticated for the same reason.

Four in ten candidates use some form of AI assistance in interviews where it is not permitted. We are also seeing deepfakes enter the picture, where someone else takes the interview on a candidate's behalf, alongside hidden support, manipulated evidence, and synthetic identities.

Some companies are responding by bringing more interviews back into the office.

That may help in selected situations, but it does not address the broader issue. Enterprise hiring will continue to operate across locations, time zones, and digital channels.

The more durable response is to make the evidence behind each hiring decision easier to verify.

Look at where the market actually stands. Almost every organization I speak to is using AI somewhere in talent acquisition. Very few are using it across the full hiring cycle, fewer still can point to a clear return, and almost nobody has taken agents to scale. All of this is landing on recruiting teams that are shrinking even as their workloads rise.

And in the pilots that do exist, the agents creating the most value sit in interviews, assessments, and candidate experience. Those are exactly the stages where evidence matters most. The value of agentic hiring concentrates precisely where verifiability has to be strongest, which is why the organizations that solve verification early will set the standard for everyone else.

I see the trust curve moving in our own numbers. When we started running AI interviews, roughly one in six candidates completed them. Today it is closer to one in two, and 70 percent of candidates choose to interview outside work hours because the process finally fits their lives.

Candidates extend trust quickly when the system is upfront about what is recorded, why, and how their responses will be judged. Recruiters are no different.

For recruiters, that means being able to see why a candidate was recommended, which data points drove the recommendation, where the system has low confidence, and how much of the resume reflects the candidate versus a language model. Our own matching engine checks for exactly this and adjusts its ranking based on whether the AI edits were cosmetic or fundamental. Claimed proficiency and demonstrated ability have to be separable.

The same applies to candidate integrity. Identity checks, behavioural signals, assessment design, and interview controls need to work together rather than as separate interventions.

Trust cannot depend on a recruiter accepting the output because the model appears sophisticated. It has to come from a process that can be examined.

Designing Autonomy That Deserves Trust

Greater autonomy in hiring only works when the boundaries are clear, and I find the simplest way to think about this is to look at the decision rather than the technology.

An agent that schedules an interview is making an operational call. If it gets the slot wrong, the cost is an email. An agent that rejects a candidate is making a judgement call, and the cost of getting that wrong lands on a person's career. The two should never carry the same level of autonomy.

Within our team we describe hiring maturity in five stages: fully manual, ATS and workflow rules, fragmented point solutions, an orchestrated model where agents run the workflow end to end with humans deciding, and finally full autonomy where people step in only for exceptions.

We deliberately stop at orchestration. In a domain this sensitive, I advise clients to avoid the fifth stage. Agents can run the show and operate at many times human scale, but the decision has to stay with a person, and the reasoning behind every recommendation has to be visible to that person.

So when I talk to talent leaders about designing for this, the conversation keeps returning to three questions. Can your recruiters and hiring managers actually see the reasoning behind a recommendation, or are they being asked to take it on faith?

Does candidate data move cleanly across the applicant tracking system, assessment tools, interview platforms, and internal talent systems, or does every handoff create a gap someone has to bridge manually? And when a decision is made or overridden, does it leave a record that would stand up to scrutiny six months later?

Human accountability sits underneath all three. Hiring decisions affect careers and livelihoods. AI can improve the quality and the speed of those decisions, and final accountability should still remain with people.

That principle is easy to state in a keynote. The harder work happens in the operating rules: deciding which calls agents can make on their own, which require approval, what evidence must be shown alongside a recommendation, and how a decision can be challenged. Those choices, more than any model capability, will determine whether AI becomes trusted infrastructure for hiring or another layer of technology that recruiters quietly work around.

The Industry I Want Us to Build

I have spent two decades watching this industry chase speed. Faster sourcing, faster screening, faster scheduling. Every generation of technology promised time back, and every generation delivered some of it.

Speed was never the point.

Close to 40 percent of a recruiter's week still goes into coordination and follow-ups. AI will return much of that time. One of our clients, a 25,000-person IT services firm, brought its time to first interview down from two weeks to 24 hours once screening and interview links were automated.

That is real value. The bigger question is what recruiters do with the hours that come back: engaging candidates, selling the role, evaluating potential and motivation, advising hiring managers, and helping candidates make decisions they will live with. That work was always the job. The coordination was in the way.

That is the industry narrative I want us to write. A faster process is useful. A process that produces better evidence and more consistent decisions changes what talent acquisition means inside an enterprise. It moves recruiting from an operational function that fills seats to a judgement function that shapes the workforce.

The next phase of AI in hiring will be decided by how confidently organizations can stand behind the outcomes their systems produce.

Trust will decide how far autonomy can go.

About Dr Rishi Thussu

Dr Rishi Thussu leads Agentic AI at SelectPrism and brings nearly two decades of experience across talent technology, consulting, SaaS and AI-led hiring. His previous roles include leadership positions at Findem and foundit, as well as consulting at McKinsey. An alumnus of IIM Ahmedabad, he focuses on helping enterprises apply AI to hiring in ways that improve speed, decision quality and human accountability.

From Recruitment Automation to Autonomous Talent Acquisition

Dr Rishi Thussu (BU Lead, SelectPrism) and Saketh Sanka (GTM Director, SelectPrism) discuss how AI agents are reshaping talent acquisition, where greater autonomy can create value, and why trust, explainability and human oversight remain critical as hiring becomes more automated.

Watch the full webinar below.