I have spent my career building AI and ML responsibly, so I have a fairly strong incentive to believe in what the technology can do.
As a chief AI officer of an AI recruitment software company, I know well that AI can remove enormous amounts of friction from recruiting, help recruiters work through volumes of information that would otherwise be unmanageable and improve the quality and consistency of decisions people make.
There is also a genuine need for it. When we surveyed our customers, improving recruiter efficiency through AI ranked as their number one priority, while research with Aptitude Research found that 50% of companies cite efficiency as their primary reason for adopting AI. Hiring teams have a real volume problem and a real time problem, and AI has an important role in solving both.
The question I care more about is where that role ends. I think we should automate as much as possible leading up to the hiring decision, while being much more deliberate about automating the decision itself, because hiring systems operate on incomplete information, imperfect models and decisions that can materially affect someone’s life.
Hiring systems start with incomplete representations
Most AI hiring systems begin with two things: a job description and a resume. Neither paints the whole picture. A job description is a representation of the work that needs to be done, while a resume is a representation of someone’s experience and capability. Both are necessarily incomplete.
Job descriptions are often inherited, adapted from previous roles or written before the hiring team has fully understood what will distinguish an excellent candidate from an adequate one, with some of the most important requirements only becoming clear once the team begins meeting candidates. Resumes have the same problem from the other direction. They compress years of work into a few pages, reflect a candidate’s ability to describe their experience as much as the experience itself and increasingly contain text written or substantially edited with generative AI. And frankly, if you could fully describe a person on a single sheet of paper, the world would be a very boring place.
That does not make either document useless, but it does mean we should treat them as imperfect measures of the underlying things we are trying to understand. The existence of an interview process is evidence of the same problem. If the written information were sufficient to make a reliable hiring decision, we wouldn't speak to a single candidate and would simply jump to an offer. A model can become extremely effective at processing, comparing and identifying patterns in the information we give it, but it cannot recover information that was never present in the first place. It can only infer from the information that is present, and those inferences introduce their own uncertainty.
Bias is a system property, not a single model defect
Removing the human from an individual decision does not make the resulting system neutral. Bias can enter through historical data, labels, sampling, proxy variables, model objectives, evaluation criteria, product design and the decisions people make on top of a model’s outputs, which is why I am wary of treating it as something that can be fixed once during model development.
Good engineering, better data, careful evaluation and continuous monitoring can all reduce risk, while product design can constrain how much authority a model is given. But responsible AI requires those things to work together. That matters particularly in hiring because the failure mode is not simply that a model makes a poor prediction. A poorly designed system can make the same type of mistake repeatedly, consistently and at very large scale, so one of the risks of automation is that it can operationalize human error efficiently.
Responsible AI, bias monitoring and human oversight therefore need to be part of the system design from the beginning rather than simply tacked on at the end.
Efficiency without trust can become a false economy
Recruiting teams naturally measure factors such as time to hire, recruiter capacity and cost per hire, and they should. But those measures become dangerous when they are treated as the whole objective rather than part of it. Candidate trust is harder to measure, but it still has a business impact.
Research cited in our work suggests that a significant proportion of candidates are uncomfortable with AI making hiring decisions about them. I would not infer from that, without evidence, that some candidates are simply going to refuse to participate or that those who remain are lower quality. The narrower business risk is enough. If some strong candidates self-select out because they do not trust how decisions are being made, then an apparent efficiency improvement can come at the cost of candidate-pool quality.
Recruiter behavior points in the same direction. Our research found that when AI conflicts with recruiter judgment in practice, recruiter judgment wins 58% of the time. I do not think that should automatically be read as evidence that the AI has failed. George Box’s observation that “all models are wrong, but some are useful” applies rather well here, because a useful hiring model does not need to replace recruiter judgment if it can materially improve it.
Ranking and screening are different product decisions
Ranking and screening can look superficially similar because both use information about candidates and roles to help recruiters work through an applicant pool. But they allocate decision authority differently.
Automated screening establishes a decision boundary, with candidates on one side continuing and those on the other not. This places considerable weight on the assumptions that the available data is sufficiently complete, the criteria are sufficiently correct and the model is sufficiently reliable around that boundary. In hiring, or any other context where systems influence significant decisions about people, I think that is a difficult standard to meet.
Ranking asks a different question by helping a recruiter decide where to look first rather than deciding who should be excluded. That does not mean ranking eliminates exclusion, because recruiter attention is finite and someone ranked 8,000th is unlikely to be considered if the recruiter reviews only 100 candidates. The difference is where and when the effective boundary is established. Ranking allows it to emerge as the recruiter learns more about the candidate pool and often more about the role itself. In contrast, automated screening fixes that boundary in advance and asks the system to enforce it.
Internally, we sometimes describe the principle as “the AI proposes, the human decides.” I like that framing because it doesn't constrain AI to trivial tasks. AI can retrieve information, compare candidates against relevant criteria, rank, summarize, surface evidence and make recommendations, substantially reducing the amount of work required to reach a good decision without owning the consequential decision itself.
Explainability is necessary, but not enough alone
Hiring is already receiving significant scrutiny under AI regulation and employment law. New York City has introduced rules governing automated employment decision tools; Illinois has expanded employment protections related to AI; and the EU AI Act treats many employment-related AI systems as high risk, while other jurisdictions continue to develop their own approaches. The precise regulatory requirements will change, but organizations need to understand what their systems are doing, why they are producing particular outputs, how those outputs are being used and who remains accountable for the resulting decisions.
Explainability is an important part of that, but it should not be confused with correctness or fairness. According to our research, 82% of companies say transparency and explainability in AI are important, while only 10% say they consistently receive them. A biased system does not become acceptable because it can explain how it produced a biased result. The value of explainability is that it gives people a better opportunity to inspect, challenge and understand a system’s output, which in turn makes accountability and effective oversight possible.
For a recruiter, that means understanding why candidates are being prioritized rather than receiving an unexplained score. For a hiring manager, it means having confidence that a shortlist can be defended rather than simply trusting the algorithm that produced it. For a candidate, it means that a consequential decision should be capable of meaningful explanation. That makes it not just a responsible AI requirement but a product requirement.
Where I would draw the line
None of this is an argument for using less AI or automation in recruiting. In many parts of the process, I think we should be using considerably more. AI can help source candidates, identify relevant experience, rank large applicant pools, summarize information, prepare recruiters for conversations, support scheduling and communication, surface evidence, answer questions and reduce administrative work. I expect the proportion of the recruiting workflow that can be automated to continue increasing as the technology improves.
I therefore care much less about how many stages of the process contain AI than I do about where, within that process, we allow the technology to acquire decision-making authority. For consequential decisions such as whether someone should continue in a hiring process, I think a person should remain accountable and have enough information to exercise meaningful judgment, rather than merely approving an algorithm’s recommendation.
That also means organizations need to understand where automated decisions already exist in their technology stack, because “AI-assisted” can describe a wide range of behaviors. The difference between recommending an action and taking it is significant. Measuring whether AI makes recruiting faster is necessary, but not sufficient. We should also ask whether it produces better outcomes, whether recruiters can understand and challenge it, whether candidate trust is being protected, whether different groups experience the system fairly and whether someone is clearly accountable when the system gets something wrong.
AI has a real and important role in talent acquisition, and I have spent my career building it responsibly. I expect that role to grow considerably as the technology improves. But the objective should be to give people better information, remove work that does not require human judgment and make the judgment that remains more consistent, more informed and easier to defend. That is a much more useful measure of the value of AI in hiring than simply asking how much of the process we can automate.

Dan Joplin is chief AI officer at ICIMS, an enterprise talent acquisition platform, where he leads the AI and Data organization across Dublin and India. He sets the company’s AI strategy and works closely with engineering and product leadership to embed AI across the hiring journey through ICIMS Coalesce AI, the unified intelligence layer across the ICIMS platform.
With more than 15 years of experience in data science, machine learning and AI, Joplin combines technical depth with a pragmatic approach to deployment, focusing on systems that solve real problems for recruiters, candidates and employers without adding unnecessary complexity. He is a strong advocate for responsible AI in talent acquisition, with an emphasis on transparency, customer control and compliance in how AI is applied to hiring processes.
Before joining ICIMS, Joplin served as director of data science at Mastercard and Director of AI at Sword Health, and held senior leadership roles at SparkBeyond and Visokio. Based in Dublin, he regularly speaks on AI in hiring, candidate experience, trust and regulation.