Interprets conditions
creates, adapts, and intervenes- Situated judgment
- Detection and coordination
- Verification and intervention
- Recovery under variation
Human, AI, and Organizational Performance
HAOP The safety framework for the AI era
AI governance is getting all the attention.
But who's designing the work systems the technology just joined?
HAOP is the framework for that missing design work.
Framework under active development · pilots and research collaborations invited
HumanAIOrganizationPerformance
HOP3: Three Performers, One Work System
Human performance is a field of study going back decades. Recognizing the organization as a performer in a work system is more recent. Human and Organizational Performance (HOP) put organizational performance in the name. HAOP considers AI the third performer, acting and introducing variability. Each performer contributes to the work system. Understanding hazards begins with considering each performer's contributions and interactions.
Acts on representation
optimizes, amplifies, and accelerates- Selection and classification
- Generation and transformation
- Ranking, routing, and suppression
- Execution within permissions and constraints
Authors conditions
structures, governs, and provides- Function allocation
- Staffing, evidence, and throughput
- Permissions, incentives, and interfaces
- Escalation, authority, and acceptable tradeoffs
Eight operating principles of HAOP
- Performance is distributed.Operational outcomes are produced by the interaction of human performers, AI performers, and organizational performers.
- Each performer has a distinct failure signature.Failure signatures are diagnostic: they tell an investigator what kind of breakdown occurred before the accountability trace begins.
- Systems execute signals, not intentions.An organization may intend safety, but the system executes what is measured, rewarded, automated, and enforced.
- Grounding is a required condition.Because AI acts on a representation of the worksite rather than the worksite itself, the work system must establish and maintain verified contact between the representations it permits the AI to rely on and relevant operational reality before an AI-shaped representation or proposed action crosses a consequential transition.
- Human oversight is work, not a label.Oversight requires time, competence, authority, attention, access to underlying data, and the protected ability to pause, verify, escalate, or intervene.
- Constraints must be designed before optimization.AI performers produce and select outputs according to the signals, criteria, objectives, and constraints they are given.
- Accountability follows control.Control includes not only the immediate ability to act, but also the authority, resources, information, and capacity to change the conditions under which action occurs.
- Learning must include all three performers.Investigations must examine what the human adapted to, what the AI optimized, and what the organization signaled, funded, ignored, rewarded, or normalized.
The five-hazard register
Understand the hazard before it becomes the event.
The interaction between the three performers creates the five hazard classes.
01
Physical interaction and machine agency
Embodied AI carries energy, mass, speed, force, and software-defined or adaptive action into physical work.
Five Hazard Classes in AI-Enabled Systems
Embodied AI carries energy, mass, speed, force, and software-defined or adaptive action into physical work.
Surveillance, scoring, pacing, scheduling, and machine-generated attribution can reduce autonomy, intensify work, suppress reporting, and make power continuous.
AI can reshape attention, situational awareness, competence, judgment, alarm burden, and verification demand. Verification Overrun appears here, but the condition is authored through work-system design.
Representations are repeatedly summarized, classified, merged, selected, stored, retrieved, and transformed until provenance, context, uncertainty, low-frequency conditions, or situated knowledge disappears.
Consequential control fragments across vendors, procurement, configuration, leadership, and local operation while verification and accountability concentrate at the visible point of use.
The work-system cockpit
The framework in one view.
One panel, a worked example of how HAOP reads a workflow. Hover or tap the panel to learn each concept.
Control path
Humancreates, adapts, verifies, judges, intervenes
AIoptimizes, amplifies, selects, transforms, routes, recommends, executes, accelerates
Organizationstructures, governs, allocates functions, provides and preserves
Representationwhat the workflow presents as true
Verification gateRepresentation tested before passageevidence, anchors, criteria, authority
Consequencewhat becomes harder to correct
Five operating modes
Track the AI's behavior, not the product label.
Operating modes classify what an AI system does in the work. A product purchased to answer questions looks like a Responder. Permitted to plan, browse, and invoke tools before responding, it operates as an Orchestrator. The full mode notes are one click below, along with grounding and anchors, which keep the work in contact with operational reality.
1. Responder2. Prioritizer3. Executor4. Orchestrator5. Ensemble
From human-in-the-loop to human-in-the-design
The person in the loop is not proof of control.
The workflow sets the demand for verification. The organization supplies the capacity. A person can be present and still lack the knowledge, evidence, time, or authority the check requires.
The HAOP method
Establish Purpose before technology.
Know the why behind the planned deployment. Then examine the unseen hazards adapting the same Industrial Hygiene method to AI deployment. ARECC is translated across the three performer relationships and verification architecture.
A
Anticipate
Establish True Function and unacceptable outcomes before selecting a product.
R
Recognize
Reconstruct work-as-done, decompose functions, and locate consequential transitions.
E
Evaluate
Test Verification Capacity, grounding, Anchor Access, reversibility, and alternatives.
C
Control
Allocate functions and engineer gates, evidence, constraints, pause, and handback.
C
Confirm
Deploy, learn, monitor, and revalidate as the work and operating conditions change.
Version 1.0 deposits the application. The method continues to develop; Revision 6.0.1 Section VIII is the current statement.
Tools and Research
Practitioner tools and workbook are in development. Collaboration on studies and pilots welcome.
White paper
The controlling source for the three performers, five hazards, operating modes, Accountability by Control, grounding, verification, and the applied method.