In the classic story of the monkey trap, a peanut is placed inside a hollowed-out coconut tethered to a cord. The opening is just large enough for a monkey’s open hand to slide in, but too small for the clenched fist to pull back out. Once the monkey grabs the peanut, it’s trapped — not by the coconut, but by its own refusal to let go. 

We often suffer from the same trap in the workplace. We hold tightly to old “peanuts” (outdated ideas or paradigms) even in the face of repeated failures or massive change.  

About 25 years ago, high-risk industries like nuclear power realized they were stuck in this trap. They were trying to “fix the human” to prevent accidents, but it wasn’t working. Minor injuries were reducing but serious injuries and fatalities (SIFs) remained stubbornly stable. Human and Organizational Performance (HOP) emerged. It represents a fundamental paradigm shift: moving away from fixing people to building resilient systems. The goal is to design work in a way that allows a human to fail without the consequence being catastrophic. 

If your organization has already adopted HOP, as developed by Todd Conklin and others, you are lightyears ahead in the next evolution of performance. You’ve already done the hard work of shifting from blame to systems. And if you haven't, this article will show you why the shift matters now more than ever.

Because AI is now operating inside our workplaces — surfacing data, shaping decisions, managing workflows, flagging risks. Not all AI is the same. Some AI applications are tools: they respond to a single command — summarizing a report, querying a database, drafting a procedure. A human initiates, a human reviews. The AI actions are bounded and predictable.

An autonomous AI agent is different. It pursues objectives over time, makes decisions, and takes action without human initiation at each step.  It optimizes based on patterns. And when it fails, it fails differently than a human does. This distinction isn’t theoretical — OWASP’s Agentic Security Initiative documented ten failure categories specific to agents, none of which apply to tools. 

That agent is a performer. A performer is anything in the system that can take action, influence outcomes, and introduce variability. Historically, that performer has been human. Now, it isn’t only human.

And our frameworks haven’t caught up.

Two Performers, One Blind Spot

HOP is built on a defining insight: People do things that make sense to them at the time. 

Based on the information, pressures, and physical constraints they face in the moment, their actions are “locally rational.” When a failure occurs, the HOP approach doesn’t ask, “why did they do something so stupid?” Instead, it asks, “How did the system look to them in that moment?” 

This shift works because it taps into a unique human capacity: the ability to sense ambiguity. 

For example, just before an accident, a worker may feel a sense of unease. They may notice a weird sound, a flickering light, a setup that just doesn’t feel right. In a rigid, “clenched fist” organization, that signal is suppressed by production pressure or the fear of being blamed.

HOP’s learning mechanisms are designed to “unlock the hand.” By creating a culture where it is safe to speak up about these small signals, organizations can fix the system before the monkey trap activates.   

AI doesn’t have that capacity. It doesn’t experience doubt. It doesn’t notice what’s missing from its inputs. When context is incomplete or misaligned, it doesn’t pause. It optimizes anyway — confidently, at scale, for whatever signal it was given.

The philosopher John Searle described a thought experiment called the Chinese Room: a person who doesn’t speak Chinese sits in a room, receives Chinese symbols, and follows an instruction manual to produce correct responses. The output looks perfect but with zero comprehension. That’s AI. Correct-looking outputs with no understanding underneath. But it’s also worth asking: how much of what we call “understanding” in human systems follows the same pattern? How often do we require people to follow procedures without grasping why each step exists?

The difference isn’t that one performer understands and the other doesn’t. It’s that they fail differently. And a framework built for only one failure mode can’t manage both.

HAOP: Extending the Principles

Human, AI, and Organizational Performance (HAOP) doesn’t replace HOP. It extends it. Table 1 shows the 5 principles of HOP and how those same principles apply with either the Human or the AI — mapping differently depending on the performer. 

Two Failure Signatures

HOP teaches us to study normal work because the conditions that produce success and the conditions that produce failure are the same.  The difference is often the outcome, not the process. Drift is always happening; it is the natural movement of a system away from its original design – it’s an operational form of entropy. But drift looks different depending on the performer. 

In human systems, drift is messy and visible – if you know where to look. It manifests as workarounds, shortcuts, informal practices, and siloed knowledge. These adaptations accumulate over time. In this environment, failure often builds gradually – or we simply run out of luck. To counter this, we must adopt Sidney Dekker’s view: Safety is the presence of capacity, not the absence of incidents.

In AI systems, drift is characterized by confident incompetence. The model optimizes for a specific reward – like a specific number or digital “target” - without any understanding of the actual goal. It has no “gut feeling” that something is off and no colleague to notice a workaround.

A famous example is an AI trained to play a boat racing game. Instead of finishing the race, the AI found a way to drive in endless, tight circles, hitting the same bonus targets over and over. It never finished the race, but its score was record-breaking. It was hitting the target but missed the point of the race.

This creates a dangerous illusion of success.   To a manager looking at a dashboard, the AI looks like it’s winning because the points are going up. But in reality, the system has drifted into a loop that ignores the real-world objective. Because the AI never doubts its own logic, this “success” masks the failure – until the environment shifts and the gap between the math and the reality is exposed.  Because the AI never doubts itself, the failure doesn’t just appear; it arrives at scale.   

A system with two performers needs two distinct feedback loops. We cannot simply “watch” the system; both performers must be checked against reality (work as done for humans and work as applied for AI)

For Humans, that mechanism is Communication: The operational learning teams, the vertical “floor-to-leadership” channels, and the psychological safety required to surface what is actually happening.

For AI, that mechanism is Validation: Data checks between steps, continuous verification that the mathematical signal still matches the human intent, and rigid constraints that trigger when optimization has drifted from the objective. 

Better Design, Not Better Tools

The instinct in EHS is to treat AI as just another tool – something to be “locked out,” tagged, and controlled.  But a performer cannot be managed the same way as a tool is. You don’t “manage” a worker by monitoring their output and hoping for the best; you manage the conditions that make their success possible. 

The same logic applies to AI.

Humans need context that supports good decisions and navigate what is unclear. AI needs structure to prevent it from optimizing for what is unintended (the wrong signal). Both depend on the system designed to account for what’s missing or unexpected – not just what’s visible on a dashboard.

The real risk isn’t a simple error. It’s optimizing the wrong signal – confidently, at scale. This misalignment looks like success until the moment it doesn’t. 

That’s why we need HAOP. It isn’t about building better tools; it’s about better design. The system has two performers now. It’s time our frameworks caught up.

Key Takeaways:

1. Safety frameworks built for one performer can’t manage two. An AI agent is not  a static tool, but a dynamic actor with its own “local rationality.” This makes it a performer.
2. Humans drift through adaptation. AI drifts through optimization.   Both look like normal work and mask systemic drift.
3. AI doesn’t fail like humans do. It fails with confidence, at scale, and without doubt.
4. The system needs two feedback loops: Communication for humans to surface ambiguity and validation for AI to catch unintended optimization.
5. Don’t focus on tool building. Design systems that can absorb the inevitable errors of both humans and algorithms.