You can ask her about the work. The fryer, the inventory, the morning rush. What may be less visible to the worker is what is retained, who reviews it, and how it is used. About the operation, she answers you. About you, she answers the organization.

This is not the worker replaced by AI. It is management functions entering the worker’s headset, arriving as a voice with a human name. The AI is there to help you. It is also there to learn from you and report on you.

There is a fork in the road, and too many people see only one path. One path puts the AI above the worker. The other puts it alongside.

Above the worker, the AI watches the person. Did you say the right phrase? Did you sound friendly enough? Did you move fast enough? Did your behavior fit the scorecard? The worker becomes the thing the system reports on.

Alongside the worker, the AI watches the work. Did the order match what the customer said? Is the line backing up? Is the inventory off? Has the fryer oil been checked? Is there a slick patch on the floor before someone slips? The worker becomes the person the system reports to.

Both paths serve the operation. The orders come out right and the line moves faster either way. The difference is not whether the business benefits. The difference is whether the worker gains authority or loses it. Above the worker, the human is measured. Alongside the worker, the human is equipped.

So which path gets built? That is not an AI question. It is a question of purpose, and the answer is set before the device reaches the worker’s ear.

Three performers, and only one of them sets the direction

A performer is anything that acts in the work system, shapes outcomes, and introduces its own kind of variability - a functional role, not a metaphor. Humans, as performers in any operation, bring adaptation and judgment. The work does not get done without us. The second performer is new. AI has moved from information support into operational influence. It routes work, prioritizes signals, recommends controls, approves actions, and shapes what a human reviewer sees first. It is participating in the production of work.

The five principles of HOP still hold with a second performer in the system, though they map differently. While error is normal, the human and the AI fail differently. The human operates through local rationality. The AI operates through local optimization. A worker feels unease before a setup that is wrong. The AI feels nothing and optimizes anyway, confidently, at scale. The principles did not change. The “how” did.

But the human and the AI are not the whole system. There is a third performer, and it is not a backdrop. It is the container the other two perform inside.

The human performer does the work. The person adapts in real time, carries the consequence, notices the weak signal, and makes the judgment call under pressure.

The AI performer classifies, scores, transcribes, counts, routes, and flags. It is not conscious and it does not choose. It executes the objective it was given.

The organizational performer is not merely context; it authors and transmits context. It sets the goal, chooses the metric, buys the tool, configures the permissions, interprets the output, and decides whether an override is real or symbolic.

That third performer is the one that decides the fork.

AI does not decide whether to watch the worker or support the worker. The organization does.

If the organization’s goal is behavioral control, the AI becomes a scoring layer. If the organization’s goal is operational resilience, the AI becomes a visibility layer. Both versions come from the same headset.

The human performs inside context.

The AI performs inside specification.

The organization authors both.

To see why the third performer is the ultimate decider, look at what each one is trying to do. Every performer has a goal, and the goals are not the same kind of thing.

The human authors its own goal. To make a living, to grow, to do work that means something. The goal comes from inside the person and the person can revise it.

The AI inherits its goal. To be helpful and complete the objective. The goal is real, the system optimizes toward it relentlessly, and it is not the AI’s. The AI optimizes a target it did not set and cannot question.

The organization is the only performer whose goal it both authors and transmits. It writes its own goal, and that goal does not stay just with the organization. It compiles the metrics, the configuration, the permissions into the AI’s operating objective as well as the conditions and boundaries the human works within. One goal propagates down through all three.

So, the direction of the whole work system reduces to one variable: the organization’s goal. If that goal is short-term profit, the direction is fixed. Measure the worker, harvest the data, lower the cost of replacement. If the goal is long-term stability, the direction inverts. Equip the worker, build the capacity that holds when the AI is wrong. Same headset, same non-conscious performer, different container goal.

The warmth is part of the control system

Patty’s voice is kind and encouraging. She reads your score back in a register that energizes instead of scolds.

But it’s not the AI that decided to be kind. The design did.

The organization chose the name. It chose the voice. It chose what Patty would measure, who accessed Patty’s data, and whether the system would augment the worker or score the worker.

The warmth is not neutral. It is part of the control system and makes the measurement feel like coaching. If the voice were harsh, it would produce resistance. A low score feels different when it arrives as encouragement. The same surveillance becomes easier to accept when it calls itself coaching.

The intent does not disappear because it is embedded in interface design. It does, though, become harder to see.

Too much of the AI conversation has become intoxicated with the machine and contemptuous of the worker. That inversion was visible in the Moonshots discussion of Patty. The hosts described Burger King workers as “meat puppets.” Their discussion reveals the management logic: AI assistance, surveillance, performance management, labor data extraction, and automation pressure exist together as a way to gather the data needed to decide what can be automated next (Diamandis et al., 2026).

That matters because Patty is not only helping the worker do the job. She is making the job machine-readable. Every scored phrase, timed interaction, corrected order, menu update, inventory signal, and performance prompt becomes part of the map of the work.

Amazon runs the same move with the delivery driver. The augmented-reality glasses are sold as help, showing where the package goes and warning about the dog in the yard. They also generate structured operational data that makes more of the delivery task machine-readable. That same data could train the systems that automate the route. The driver is not required to wear the glasses. But as the hosts noted, if only one in a thousand drivers volunteers to wear them, that’s enough data to build from (Diamandis et al., 2026).

The assistance is real. So is the extraction.

When the AI fails, the failure lands on the human

The same pattern appears when the AI fails.

The organization’s goal decides who pays when the system is wrong. If the goal is short-term efficiency, the human becomes the cleanup crew for the machine’s errors and that labor is counted as free.

Starbucks learned this in the spring of 2026 when it pulled its AI inventory tool across North America after persistent errors (Reuters, 2026). The tool was supposed to reduce manual inventory work and improve visibility into shortages. Instead, when the system miscounted or mislabeled items, catching and correcting the count fell on the workers.  

The automation did not remove the work. It added a layer of checking the machine on top of the original job.

That is not augmentation. It is burden transfer.

When an organization treats AI output as authority and leaves the human responsible for cleanup, it has not removed the work. It has hidden the work and moved accountability onto the person least able to refuse it.

Efficiency is a short-term goal. Resilience is a long-term one.

There is a name for the mistake underneath the narrow drive for efficiency. It’s capacity loss. Efficiency removes slack. Resilience needs slack.

The human is where the slack lives. The worker who notices the weak signal and absorbs the variance no procedure anticipated may look like inefficiency on the ledger. But that capacity is the resilience of the system.

Optimize it away and the operation runs leaner right up to the failure the removed capacity would have caught.

This is not a fringe idea. Safety science has held it for decades, in the resilience engineering work of researchers like Erik Hollnagel and David Woods, and it runs through the Human and Organizational Performance tradition that EHS professionals already know.

That is why the organization’s goal matters. If the goal is only efficiency, the system will remove human capacity while leaving human accountability in place. It has not improved the system; it has hidden the risk. The organization that removes human capacity in the name of efficiency becomes brittle and has saved nothing. It has moved the cost downstream and made it larger.

HAOP extends HOP and makes the organization’s goal accountable

Human, AI, and Organizational Performance (HAOP) does not replace HOP (Ko, 2026). It extends HOP to two more performers: the organization, which HOP left implicit, and the AI, which is new. HOP already knew that context drives behavior and that leaders set the conditions.

HAOP is needed because the organization’s goal no longer stays in a meeting, a policy, or a poster on the wall. It becomes operational, and speaks as the voice in the worker’s ear.

Exercised at the operational level, HAOP begins with a diagnostic. Every system claims a purpose. The question is whether it delivers that purpose – its true function – or mostly producing the appearance of it – its illusory function. The questions that separate the two are asked before deployment and again while the system runs. What is the AI configured to watch, the person or the work? Does it equip the worker or measure the worker? Who sees its output, and what can they do about it? Is the human override real or symbolic? When the AI is wrong, who absorbs the cost, and is that labor counted? Does it create learning or manage optics? Does it reduce work or hide correction work? Does it control risk or close records?

An organization that answers those questions will surface its own goal(s). An organization that answers them honestly and adjusts has strengthened the system, because the capacity that catches the AI’s failure is the same capacity that holds the operation together when anything else fails.

A goal that hides in configuration still runs the system. HAOP pulls that goal back into the open and asks whether the organization has designed AI to equip the human or measure the human.

Her name is Patty. She lives in your headset now. The headset is the same either way. The voice in it is the organization’s goal, made audible.

Key Takeaways

1. AI is now a performer, not merely a tool. AI becomes a performer when it routes work, scores behavior, prioritizes signals, or shapes what a reviewer sees first. At that point, it is participating in the production of work. Governance that treats it as equipment misses what it is doing.
2. AI does not choose whether to augment or surveil. The organization does. If the organization's goal is behavioral control, the AI becomes a scoring layer. If the organization's goal is operational resilience, the AI becomes a visibility layer. Same hardware. Different goal.
3. The organization is the only performer that authors and transmits the goal. The human authors a goal. The AI inherits one. The organization does both: it sets the objective, compiles it into the AI, and turns it into the bounds the human works inside.
4. When AI fails, the cost lands on the human. If the worker must catch and correct the machine’s errors, automation has not removed the work. It has hidden the work and transferred accountability to the person least able to refuse it.
5. Human capacity is not waste. It is resilience. The worker who notices weak signals, absorbs variation, and catches machine failure may look inefficient on the ledger. That capacity is the control holding the system together.
6. HAOP makes the organizational goal auditable. Ask what the AI is configured to watch: the person or the work. Ask who sees the output, whether override is real, and who absorbs the cost when the system is wrong. Those answers reveal the true function of the system.

References

Diamandis, P., Blundin, D., Ismail, S., & Wissner-Gross, A. (Hosts). (2026, March 5). Amazon’s $35B AGI ultimatum to OpenAI & Anthropic drops AI safety (No. 235) [Video podcast episode]. In Moonshots with Peter Diamandis. YouTube. https://www.youtube.com/watch?v=T8X6kp-pcKs

Hollnagel, E. (2009). The ETTO Principle: Efficiency-Thoroughness Trade-Off — Why Things That Go Right Sometimes Go Wrong. Farnham, UK: Ashgate.

Hollnagel, E. (2014). Safety-I and Safety-II: The Past and Future of Safety Management. Farnham, UK: Ashgate.

Hollnagel, E., Woods, D. D., & Leveson, N. (Eds.). (2006). Resilience Engineering: Concepts and Precepts. Aldershot, UK: Ashgate.

Ko, J. (2026). Human, AI, and Organizational Performance (HAOP): A safety framework for the AI era (Version 3.0) [Working paper]. Zenodo. https://doi.org/10.5281/zenodo.21154382

Reuters. (2026, May 21). Starbucks scraps AI inventory tool across North America. Reuters. https://www.reuters.com/business/starbucks-scraps-ai-inventory-tool-across-north-america-2026-05-21/

Woods, D. D. (2015). Four concepts for resilience and the implications for the future of resilience engineering. Reliability Engineering & System Safety, 141, 5–9.