A Debate about Going From AI Enthusiasm to Accountable HR Adoption

[00:00:00] Welcome to the debate. You know, when structural engineers set out to build a suspension bridge, they don't wait until the morning rush hour to figure out where the load-bearing cables should go. Right, with all the cars driving over it. Exactly. They build the structural support first. I mean, they establish those safety thresholds early, long before the first commuter ever rolls across the span. Yeah, but, well, if you require those engineers to spend 10 years testing the tensile strength of every single bolt in a wind tunnel before they can lay a single plank of wood,

[00:00:28] you never actually get the cars across the river. You're saying it stalls out? Completely. You end up with a perfectly safe, completely theoretical bridge that serves absolutely no one. And that tension, that exact push and pull is where we find ourselves today. We are examining the structural cables of our modern workforce. Specifically, the organizational conditions required for the responsible adoption of artificial intelligence and human resource management. Right. And we are looking at this through the lens of Jonathan H. Westover's article,

[00:00:56] from AI enthusiasm to accountable HR. A really fascinating piece. It is, yeah. Westover unpacks this 2026 mixed method study by Poo Dom on AI adoption among HR practitioners in Thailand. And he basically uses it to build a framework for how companies should actually handle this technology. So the central question we are exploring today is fundamentally about sequencing and safety.

[00:01:17] Must organizations establish formal, stringent ethical governance structures prior to expanding AI deployment in HR to achieve what Westover calls justified adoption? Or conversely, does an over-reliance on perceived governance frameworks risk stifling practical AI utility by conflating survey-based intentions with verified organizational effectiveness?

[00:01:36] Right. And to be clear on where we stand, I represent the position that robust ethical governance and accountable decision-making are essential prerequisites that must be formalized before AI can be responsibly integrated into consequential HR workflows. And I represent the perspective that building heavy preemptive governance structures based primarily on practitioner perception data risks prioritizing bureaucratic reassurance over the iterative local testing necessary to realize AI's actual value.

[00:02:02] So let's lay the groundwork for why this prerequisite of governance is just so critical. Go for it. Human resource decisions are, you know, uniquely complex. If you're an HR manager listening to this, you know we are not talking about optimizing a supply chain route here. No, it's highly personal. Exactly. We are talking about livelihoods. Hiring, performance management, workforce planning. I mean, these actions directly impact human lives. They do. And Poo Dom's study found something incredibly striking in this regard.

[00:02:32] Among the hundreds of HR practitioners surveyed, ethical governance, which includes things like formal safeguards, transparent policies, human oversight, that was the largest positive predictor of their intention to adopt AI. Yeah, the data point there is a standardized coefficient of 0.412 for ethical governance. Right, 0.412. And to translate that statistical jargon into plain English, that coefficient is a massive signal. It's definitely significant.

[00:02:58] It means that out of all the reasons an HR professional might choose to use an AI tool, governance was the undisputed heavyweight champion. I mean, it far outweighed performance expectancy, basically how well the tool does its job, which was only at 0.245. Right. It outweighed social influence, which sat at 0.133. So what Westover argues, and what I fundamentally agree with, is that establishing accountable decision-making is not bureaucracy. It is the strict definition of justified adoption.

[00:03:26] But see, I come at it from a completely different angle. How so? Because we really need to look closely at the methodological foundation of that rigorous governance framework. Poo Dom's study is valuable, sure. It's a great snapshot. But it has distinct limitations. Okay, what limitations? It only measures adoption intention, right, and perceived effectiveness. It does not measure actual full-stage integration or, you know, verified value realization in a live corporate environment. Wait, let me make sure I'm understanding your critique here.

[00:03:56] Are you saying that because it's a survey of intentions, the demand for governance just isn't real? No. I'm saying we have to be incredibly precise about what the data actually proves. That 0.412 coefficient, it helps explain about 66% of the variance in a practitioner's intention to use a system. Right. But conflating a practitioner's perception of strong controls with the actual operational effectiveness of those controls is really dangerous.

[00:04:21] When we look at the heavy oversight models Westover references, like Duke Health's ABCDS oversight or Microsoft's Sensitive Uses program, we are looking at massive institutional bureaucracy. Well, they are large organizations. Exactly. If you take a survey where people say, yes, I like safe AI, and use that to mandate a Duke Health-level oversight board for every HR department, you are going to completely bottleneck the natural learning curve that teens need to understand human-AI collaboration.

[00:04:49] I think you're dismissing the qualitative side of the research, though. Pooidam's study wasn't just a multiple-choice survey in a vacuum. It had interviews, yeah? Right. It included deep interviews with 22 senior HR leaders alongside the 284 practitioners surveyed. This mixed-methods approach definitively shows that practitioners will not fully embrace AI without clear accountability. They say they won't, yes. And this brings up a concept from Westover's paper that I think is the absolute linchpin of this entire debate.

[00:05:17] It's the difference between trust and trustworthiness. Oh, it's an elegant distinction. I'll give you that. It is because it completely reframes the conversation. Trust is simply a user's willingness to rely on a system. That's a feeling. Sure. But trustworthiness concerns whether that reliance is actually warranted. Like, does the system deserve to be trusted? When ethical governance emerges as the dominant predictor of adoption, it tells us that HR professionals intuitively understand they need structural trustworthiness before they can grant their personal trust.

[00:05:47] So they want to know it's safe before jumping in. Exactly. If the governance isn't there, the adoption will be superficial. It's a great philosophical anchor. Um, but practically? I just don't buy the leap from that survey data to an enterprise-wide mandate for preemptive governance. Why not? Let me tell you why. You are taking explanatory statistics for survey responses and treating them as an ironclad law of physics. We have to consider what researchers call common method bias. Oh, from Podsikoff? Yes.

[00:06:15] Podsikoff and his colleagues warned us about this back in 2003. Think about how surveys work in the real world. You capture perceptions at one single point in time. Right. You're talking about the mood of the respondents skewing the data. Exactly. If someone takes a survey on a Friday afternoon, their company just had a great quarter, and they are generally optimistic about technology, they are highly likely to rate their intention to use AI favorably. Okay. Sure. And because they are in an agreeable mood, when the very next question asks, are transparent policies good? They rate that highly too.

[00:06:44] You are capturing a generalized positive sentiment. It's a vibe, not a structural reality. A vibe. Yes. To say this definitively proves practitioners won't embrace AI without rigid accountability frameworks is overstating the empirical evidence. We just cannot build resource-heavy preemptive mandates just because survey respondents expressed a concurrent favorable view of transparent policies. Well, you can call it a vibe, but it aligns perfectly with the fundamental nature of the work being automated.

[00:07:12] Let's look at Westover's first proposed priority, establish accountable decision-making. Okay. Let's look at it. You cannot simply slap the label assistive on an AI system and think that bypasses the need for scrutiny. Reich and Krakowski in their 2021 work on the automation-augmentation paradox argue convincingly that human and automated tasks are deeply interdependent. Yes, but... You can't cleanly separate them. But we separate human and machine tasks all the time. An HR rep writes the prompts, the AI drafts the text.

[00:07:42] It's not that simple when we are talking about evaluating human beings. Let me give you an analogy. All right. Right? Using an AI candidate ranking tool is not like using a fancy calculator. It is like driving a car with a GPS navigation system. If that GPS suddenly hallucinates a bridge that isn't there and the driver doesn't have their eyes on the road, you are going over a cliff. Well, that's a bit dramatic. But it's accurate. This is why Westover demands an AI HR register.

[00:08:09] It explicitly separates advice from action, documenting the purpose of the tool, the data inputs, and the accountable human owner before anyone gets behind the wheel. Hold on. I think your GPS analogy reveals exactly where this governance framework overreaches. How so? An AI that ranks thousands of candidates might be like a GPS. But an internal AI tool that, say, drafts a job description, that is closer to a highly sophisticated work processor than it is to navigating a car. But they're both AI. Right.

[00:08:37] But applying Duke Health-level clinical oversight to administrative HR tasks is an enormous overreach. And frankly, it ignores the contextual findings within Pu'udum's own study. Remind me which finding? The study explicitly found that facilitating conditions, meaning the technical infrastructure and support, showed a stronger association with adoption intention among respondents in larger organizations. Sure. Bigger companies have more support. Exactly. Because large organizations have the budget for AI HR registers, cross-functional review groups, specialized legal counsel.

[00:09:06] Demanding heavy oversight budgets up front paralyzes small employers. I don't think it paralyzes them. It forces a 50-person company to treat an AI tool that formats onboarding manuals with the same bureaucratic scrutiny as an algorithm making automated hiring recommendations. I'm glad you brought up the small business angle because I want to push back on that. Are you suggesting that small businesses get a free pass on safety just because they don't have Microsoft's budget? No, I'm saying... Because Westover specifically addresses this through the concept of proportional escalation.

[00:09:34] He isn't saying a local retail chain needs a 20-person ethics board. They still need formal governance under his model. He's saying they need a named owner for the software and an executable stop procedure if it starts generating faulty logic. That doesn't cost millions of dollars. It just requires intentionality. But intentionality still requires resources and time, which are in short supply for a small HR team. But, okay, let's assume they establish that owner. What happens next?

[00:10:00] What happens next is we confront the capability gap that Powodum's interviews revealed. Knowing how to do traditional HR work does not automatically prepare someone to evaluate AI-supported HR work. True. Which naturally brings us to the need to build the capability to question AI, not just operate it. Westover calls this critical AI literacy. Practitioners must be trained on the failure modes relevant to their jobs and explicitly given the authority to reject imperfect AI outputs. Right.

[00:10:30] The authority piece is big. Look at DBS Bank's framework, for example. They call it pure. Purposeful, unsurprising, respectful, and explainable. The pure framework, yeah. Yes. It means an HR professional doesn't just ask, did the machine give me an answer? They are trained to ask, is this output unsurprising based on the data? Is it explainable? They embed ethical reasoning into everyday capability. And you think that's required beforehand? Absolutely.

[00:10:56] You cannot achieve that level of critical literacy if you are just iterating on the fly without formal training. I agree that critical literacy is vital. I don't think anyone is arguing that we should just let algorithms run wild while HR practitioners take a nap. Glad we agree on that. But I challenge the underlying assumption here that you need perfect preemptive governance to create that literacy. Let's look at Bryn Yolson's 2025 study on generative AI at work. Okay. They studied over 5,000 customer support agents. And what did they find? A 15% average increase in productivity. In customer support, yes.

[00:11:26] But here is the critical mechanism of how that happened. Those gains were highly concentrated among the less experienced and lower skilled workers. Because the AI helped them catch up. Exactly. The AI acted as a leveling mechanism, capturing the implicit knowledge of top performers and giving it to the novices. If we follow your framework, requiring practitioners to have perfect specialist access, deep critical AI literacy, and a proven ability to interrogate model outputs before we allow deployment, we are gatekeeping the technology. Gatekeeping? Yes.

[00:11:56] We are keeping it out of the hands of the very novices who stand to benefit from its assistive capabilities the most. I'm not convinced by that line of reasoning at all, because it completely conflates customer support efficiency with HR decision quality. A 15% increase in Resolve IT support tickets is fantastic. It is. But a 15% increase in processing resumes, if those resumes are being filtered through a biased system that a novice recruiter doesn't know how to question, that is a massive organizational liability. Well, obviously bias is a liability.

[00:12:25] Tom Bay emphasized this back in 2019. HR problems have unique accountability challenges. Critical AI literacy isn't about requiring everyone to be a data scientist. Right. It's about ensuring that a novice recruiter knows how to spot when a system is producing a confident answer without sufficient supporting evidence. If you give a novice a powerful tool without that preemptive training, they don't become an expert. They just become a rubber stamp for a machine. But how do they actually learn to identify that missing evidence? Do they learn it in a seminar? No.

[00:12:54] They learn by using the tool in their actual workflow, seeing its outputs, and comparing it to their local context. But that's risky. Westover talks about requiring local evidence before expanding deployment. I am 100% in agreement with that. But you generate local evidence through supervised, hands-on use, not through preemptive decision maps and classroom role-playing about hypothetical failure modes. Role-playing prevents real-world harm. A novice recruiter learns the limitations of an AI ranking system by testing it on real,

[00:13:23] local data and seeing where it misses the mark. If you burden that learning process with a mandate that they must document their ethical reasoning before the pilot even begins, you stifle the exact local experimentation Westover claims to want. I would argue that practicing disagreement before deployment is exactly how you prevent systemic harm. But let's broaden the lens here, because HR systems don't just affect the HR users. Right. They affect the candidates. They affect the employees. Kellogg and colleagues showed in 2020 how algorithms fundamentally reshape organizational

[00:13:52] control. They direct, evaluate, and discipline workers. Yeah, algorithmic management. Exactly. This isn't just about computational efficiency. It's about authority. And accountability has to extend beyond the users to the stakeholders. Under Westover's framework, people need to know when an AI is materially influencing a consequential process about their career. Absolutely. They need a practical means to correct irrelevant information, and they need to be able to obtain human review. Ethically, yes, stakeholder voice is necessary.

[00:14:21] Practically, Westover's mechanism for this is incredibly fraught. Why? He suggests organizations should create recurring opportunities for candidates. It's to review pilots. It's to review pilots. And mandates a documented governance record for every single complaint. It has to be tracked to an owner, investigated, and the outcome communicated. That sounds like basic due process to me. In theory, it is. But in a high-volume recruitment environment where you have 10,000 applicants for 50 roles,

[00:14:47] requiring a documented investigation cycle every time an external applicant doesn't understand an AI-assisted instruction risks turning HR experimentation into an endless, litigious communication exercise. It's just communication, though. It stops being about efficient process improvement and becomes an administrative nightmare. Designing feedback loops is good practice. Mandating that every concern enters a formal decision-making tree and receives a reasoned, documented response before a pilot can be considered successful, that is a recipe for complete organizational paralysis.

[00:15:17] I have to push back hard on the idea that this causes paralysis. I see it as organizational responsibility. But the cost? Let me finish. Westover explicitly notes that a pilot shouldn't be labeled unsuccessful merely because it leads to a decision not to deploy the tool. Okay, fair. If a well-designed evaluation prevents an unsuitable, hallucinating application from becoming embedded in your hiring process, that is a massive success. The true failure is learning about an algorithm's limitations only after employees have been harmed,

[00:15:44] or after your dependence on the tool makes withdrawal impossible. Yeah, vendor lock-in is real. Right. So the cost of a documented investigation cycle is microscopic compared to the cost of scaling a faulty evaluation tool across an entire enterprise. And I would counter that the cost of inaction is also profound. We are talking about human resources, a field that is already imperfect. Of course. If we demand that every AI application in HR meets the evidentiary and governance standards of clinical health care before it can scale, we will be stuck using legacy systems that

[00:16:14] we already know are rife with human error. Well, we know humans are biased. We know human hiring is subjective. We know human performance reviews are inconsistent. If we over-engineer preemptive governance, we hold AI to a standard of perfection that we have never demanded of ourselves, thereby impeding the practical testing where true incremental value is discovered. That's a fair challenge. As we draw this discussion toward a close, though, I think my position remains grounded in the sheer stakes of the work. The livelihood aspect. Yes.

[00:16:42] Responsible AI and HR demands a fundamental shift in how leaders approach technology. We must move away from simply asking, what can this system do? And start asking, under what conditions should we allow to influence decisions about people? Right. By prioritizing accountable decision-making, critical AI literacy, and local evidence, we build a structural trustworthiness that Poo Wudum's data shows practitioners are implicitly demanding. Governance isn't a bureaucratic hurdle. It is the bridge's support cable.

[00:17:11] It is the only viable path to justified adoption. And from my perspective, while oversight is undeniably necessary, organizations have to clearly distinguish between a practitioner's survey-based intentions and verified operational benefits. The survey versus reality gap. Exactly. A favorable attitude toward transparency policies in a survey does not automatically validate the implementation of rigid, heavy governance frameworks in practice. Overengineering these preemptive controls risks building a fortress of compliance that

[00:17:39] locks out the practical, assistive value these tools can provide, especially for the less experienced practitioners who could use the boost the most. I will say, despite our different approaches to how heavy that governance should be, there is significant common ground here. For sure. We both firmly agree on the necessity of critical AI literacy for HR professionals. Knowing how to click buttons in a software interface is not the same as exercising professional judgment around its outputs. Absolutely. And we also agree that a generic vendor demonstration of an AI tool's capabilities should never,

[00:18:09] ever substitute for local, context-specific testing. Never. What looks flawless in a controlled sales demo may fail entirely when applied to the unique workflows, languages, and populations of a specific company. Generating local evidence is non-negotiable. Ultimately, integrating algorithmic management into a domain as deeply human as human resources is incredibly complex. Moving from general enthusiasm to truly accountable HR requires ongoing rigorous exploration. It really does.

[00:18:38] As Poo Dam rightly noted at the end of the study, future research must move beyond just measuring user intentions and begin examining objective organizational outcomes. Exactly. Because until we measure what actually happens when these tools are integrated into the daily messy flow of work, we are just theorizing about traffic patterns without looking at the road. So we return to the river, looking at our suspension bridge. Whether you believe every load-bearing cable must be exhaustively stress-tested before the

[00:19:06] first plank is laid, or whether you believe we must start carefully moving traffic across to truly understand the physics of the span, one thing is certain. What's that? The bridge between human judgment and artificial intelligence is being built beneath our feet right now. The question remains, how much weight are we ready to let it carry?