The oversight illusion: when accountability outruns agency

^{“Understanding ourselves, navigating the world around us, making sense of others, and recognising the impact we have on them. Because who are we, really, if not shaped by one another?“}

^{Happy Thursday, everyone. I’m Shaida Darian, a Business and Organisational Psychologist. I’m fascinated by three questions: Why does this happen? What can we do about it? And when should we act? Join me down the rabbit hole. With any luck, we’ll find Wonderland.}

Subscribe | Past essays | Advertise with us

🤝 This edition is kindly brought to you by Leapsome

Most HR leaders can see where AI is taking their function. Few have had someone beside them who has made the move before.

This is HR's moment to lead the company's AI transition, and Leapsome believes an ambitious people team can build an agentic HR team within six months. To help, customers who sign by December 31 get Leapsome's Agentic HR Accelerator Program (€24K value) and the Leapsome Agentic HRIS Foundation package for free, with any Leapsome Talent Suite module.

For six months, a dedicated Leapsome point of contact works with your team, alongside a cohort of HR leaders making the same move and a live expert series led by Jessie Zwaan. HRIS migration can be as fast as two weeks, handled end-to-end by Leapsome's migration experts, so you're live before your first program session.

Leapsome is the Agentic HR Platform. For lean people teams with outsized business impact.

Start your six months

_{We partner with a select group of brands that we use or admire, allowing this newsletter to remain free and independent. Reach out for your campaign here.}

A few weeks ago, I sat in an executive AI steering committee meeting at a multinational enterprise. The Chief Risk Officer was presenting a new oversight framework for AI agents authorised to negotiate vendor contracts and execute purchase orders. A dashboard showed that 99.2% of agent-generated transactions had been “reviewed and approved” by operations managers.

The average approval took less than four seconds.

During a break, I asked one manager how she evaluated the agent’s actions and changes to contract terms in that time. She lowered her voice and smiled wearily. There were fourteen pages of execution logs and two hundred live orders to manage each day, she explained. In practice, they checked whether the compliance indicator was green and clicked approve to keep work moving.

The dashboard recorded approval. What that approval represented was less clear.

I’m Shaida Darian, a business and organisational psychologist, and I examine how workplace systems shape the decisions people make.

That exchange stayed with me because the manager had been assigned a responsibility without an obvious opportunity to exercise it. Checking a recommendation requires sufficient understanding, relevant evidence and time to assess it. Challenging one also requires authority to interrupt the process.

AI may accelerate execution, but the judgement expected of the person signing off still takes work. In that meeting, the approval rate told us how consistently managers clicked the button. It could not establish how thoroughly they had evaluated what came before it.

Which brings me to the question: what does it mean to hold someone accountable for a decision they no longer have a realistic opportunity to evaluate?

The manager’s four-second approval illustrates a distinction that matters as organisations adopt AI: responsibility can be assigned more easily than the capacity to exercise it.

Job Demands–Resources theory provides a useful starting point. It examines how demands requiring sustained effort interact with resources such as autonomy, support and feedback. Applied to AI supervision, the concern is that expanding automated output may increase review demands without a corresponding investment in the people expected to assess it. This is an application of the theory, rather than a finding that all AI deployments create overload.

Attention is one constraint. Parasuraman and Manzey’s review explains how people can overlook automated failures or accept incorrect recommendations, particularly when attention is divided between competing tasks. These tendencies are influenced by workload, experience and system design. A requirement to “check the output” does not establish that meaningful checking is feasible.

Understanding is another. Endsley and Kiris’s experimental research examined the out-of-the-loop performance problem: people supervising automation can lose awareness of the task and struggle when required to resume control. The study concerned a simulated task, not contemporary AI managers, but it identifies a relevant risk. Someone reviewing a completed result may lack the context they would have acquired while producing it.

This does not make expertise loss inevitable. McKinsey’s recent organisational analysis explicitly raises the need to rebuild learning pathways as AI absorbs foundational work. The practical issue is how people develop the judgement that their supervisory roles require.

When responsibility exceeds practical control, Madeleine Clare Elish’s concept of the moral crumple zone becomes relevant. It describes how a human operator can absorb blame for a system’s failure despite having limited influence over its behaviour. This is a warning about how responsibility is distributed, not a universal rule about legal liability.

The organisational risk is therefore broader than an inattentive reviewer. An approval process may record individual responsibility while concealing gaps in workload, understanding and intervention authority. Those conditions determine whether human judgement can meaningfully affect the outcome.

Effective oversight requires a role people can realistically perform. Leaders need to define what must be reviewed, provide the resources to assess it and ensure that intervention can change the outcome. Three decisions matter.

Responsibility also extends beyond the reviewer. Leaders who determine staffing, incentives, system permissions and acceptable risk shape whether oversight is workable. Their decisions belong in the accountability structure alongside the final sign-off.

When you deploy autonomous AI agents without expanding human oversight capacity, the risk does not disappear. It simply travels somewhere else.

Increasingly, it lands on the manager.

Perhaps we need to stop asking managers how they can become better at verifying AI outputs in their spare time, and start asking why so much unmonitored risk is reaching them in the first place.

Until next week,

If something here speaks to you, I’d love to hear it.

Until next week,
Shaida

P.S. If you manage a team supervising automated or AI-assisted workflows, where does the oversight pressure land most heavily for you? Reply and tell me.

If you’re new here, I’m over the moon you’ve joined us! To help me craft content that’s actually useful (and not just noise in your inbox), I’d love it if you took 1 minute to answer this quick survey below. Your insights help shape everything I write.

💡 Dive into our other newsletters - where psychology meets the founders, creator economy, e-commerce marketing, and AI founders.