ARTICLE Human Resilience

How Should Organizations Measure Human Risk in AI-Enabled Work?

SHARE
By Team CM · Oct 28, 2026, 8:06:08 PM
How Should Organizations Measure Human Risk in AI-Enabled Work?

AI measurement is developing much faster on the technology side than on the human side.

Organizations can track which models are being used, how often employees interact with AI, where agents are deployed, how much time automation saves and whether technical controls are functioning. They may know exactly how many people completed AI training and how confident employees say they feel using the tools.

Those measures are useful. They do not necessarily tell the organization whether people are becoming safer, more capable or more resilient as AI changes the work.

That requires a different measurement question.

At Cybermaniacs, we think organizations should measure human risk in AI-enabled work across the conditions that determine whether people can use, supervise, challenge and remain accountable for AI safely. That includes competency, behavior, psychology, culture and organizational conditions, interpreted in the context of the role, the AI system and the consequence of the work.

The objective is not to create one more employee score.

It is to produce evidence that helps the organization understand where human risk is emerging, why it is emerging and what should change because of it.

Why isn't AI training completion enough?

Training completion tells you that an intervention reached a population. It does not establish that the population can perform the behavior the organization now requires.

Someone can complete an AI literacy course and still accept a plausible but incorrect answer. They may understand the concept of automation bias while becoming steadily more dependent on a system they use every day, or know that high-risk actions should be escalated while remaining uncertain about what “high risk” means in their actual workflow.

This is not a new measurement problem.

Cybersecurity awareness programs spent years relying heavily on completion rates because they were easy to collect and easy to report. Those numbers told organizations whether training had been delivered, but considerably less about whether people could recognize, resist or respond to actual threats.

AI creates an opportunity to start with a richer model.

Training data belongs in the measurement environment. It simply should not be asked to prove outcomes it was never designed to measure.

Start with the human responsibility inside the AI-enabled work

The measurement model should follow what the person is actually expected to do.

An employee using an AI assistant for routine productivity needs a different set of capabilities from someone reviewing AI-supported financial analysis. A manager supervising autonomous agents has another role again.

Before deciding what to measure, organizations should ask where human judgment still matters.

Does the person need to verify an output, make a consequential decision, recognize unusual conditions, supervise autonomous activity, escalate exceptions or remain accountable for the outcome?

Those responsibilities define the human-risk question.

If the organization expects a person to catch an AI error, measurement needs to tell us whether they can recognize that error. If they are responsible for supervising an agent, measurement needs to address whether they understand the agent's mandate, can recognize when it has moved outside normal conditions and know what authority they possess when that happens.

This role-first approach prevents AI measurement from becoming a generic survey about how employees feel about the technology.

What should organizations measure about AI competency?

Competency provides part of the foundation.

People need enough understanding of AI capabilities, limitations, organizational policy, relevant security risks and the systems they actually use. Higher-risk roles may also require stronger domain knowledge or more sophisticated understanding of how AI influences the work.

But competency should be measured at the level at which it matters.

An employee does not need to explain transformer architecture to recognize that a confident AI-generated legal claim needs independent verification. A supervisor does need enough understanding of an agent's capabilities and boundaries to know when a request or behavior deserves attention.

That means the useful question is not simply “Does this person understand AI?”

It is “Do they understand enough about AI and the work to perform the human responsibility this workflow assigns to them?”

The required depth will change by role, which is why role and consequence belong inside the measurement model rather than being treated as segmentation added afterward.

AI Judgment needs to be measured separately from AI literacy

Knowledge does not guarantee good decisions.

An employee can know that generative AI sometimes produces incorrect information and still accept the wrong answer because it is well written, fits expectations and arrives during a busy afternoon.

Cybermaniacs uses AI Judgment to describe the human ability to determine when, whether and how AI-generated information, recommendations or actions should influence a real-world decision.

That capability is especially important because AI use increasingly moves beyond asking employees to remember rules.

People have to interpret circumstances. They need to decide when an output is credible enough to use, when independent verification is warranted, when AI should not be used at all and when an apparently routine situation has become consequential enough to require escalation.

Our Guide to How Do You Measure Whether Employees Can Exercise Good AI Judgment? explores this measurement problem in more detail.

The important implication for the broader measurement model is that organizations should not infer judgment from knowledge alone.

What behavior should organizations measure around AI?

Behavior tells the organization what people actually do around the technology.

That can include how employees verify AI-generated information, what types of work they delegate, when they rely on AI recommendations, whether they escalate unusual situations and how they respond when AI and human judgment disagree.

The purpose is not to monitor every employee interaction with AI.

Useful measurement focuses on behavior that matters to risk.

If an organization depends on people to independently verify consequential AI output, evidence about verification behavior matters. If humans are expected to intervene when an agent crosses a boundary, escalation and intervention behavior become relevant.

The context is critical because the same behavior can have different meanings.

High acceptance of AI recommendations may indicate that the system is performing extremely well. It may also indicate that people have stopped applying meaningful independent judgment.

Behavior becomes risk intelligence only when the organization can interpret it alongside the quality of the system, the nature of the work and the capability of the people involved.

Reliance should be measured as calibration, not trust

Organizations frequently ask employees whether they trust AI.

That is interesting but incomplete.

Trust becomes useful to risk management when it is connected to the behavior that follows from it. A person may report moderate trust while verifying almost nothing, while another may report high trust in a well-performing system and still apply careful scrutiny to consequential cases.

Cybermaniacs therefore looks at reliance through the idea of calibrated reliance.

Our Guide to How Do You Know When Employees Are Over-Relying on AI? examines whether human reliance remains appropriate to the system's performance, the uncertainty involved and the consequence of the work.

Measurement should therefore look for relationships.

Does verification decline as familiarity grows? Do people increase scrutiny when consequence rises? Does confidence increase faster than capability? Are certain roles becoming unable or unwilling to complete work without AI support?

Those patterns tell the organization much more than a generic trust score.

Psychology helps explain why the behavior occurs

Behavior tells us what happened. Psychology often helps explain why.

Self-efficacy matters because employees who lack confidence may either avoid useful AI or defer excessively to it. Risk perception influences whether someone views an unusual AI output as worth investigating, while perceived control affects whether a person believes they can influence what the system does.

Attitudes matter too.

Some employees may approach AI with excessive skepticism, while others assume that increasingly sophisticated systems deserve greater confidence. Either extreme can produce poor decisions depending on the work.

The measurement model should therefore examine psychological factors where they help interpret behavior.

A reviewer who rarely challenges AI because they genuinely observe high-quality performance presents a different risk condition from someone who rarely challenges because they assume the technology knows more than they do.

The action looks similar. The underlying intervention is not.

Culture shapes whether people exercise the controls they have

Human risk measurement also needs to account for the environment in which people make decisions.

An organization can train employees to challenge AI while leadership simultaneously signals that rapid adoption is the priority. It can grant formal escalation authority while teams learn that raising concerns creates delay or makes them look resistant to change.

Those conditions affect whether controls are exercised.

Cybermaniacs' culture model examines enabling factors such as leadership, psychological safety, group cohesion, individual commitment, prioritization and sense of control, alongside behaviors such as communication, accountability, teamwork and incident reporting.

For AI-enabled work, the relevant question is how those cultural conditions affect safe use, challenge and intervention.

A technically sound human control can become weak inside an environment where people do not feel comfortable exercising it.

That is why culture belongs in the evidence model where it materially affects AI risk rather than being treated as a separate employee-engagement issue.

Organizational conditions need to be measured too

Not every human-risk problem is caused by the person.

Sometimes the organization has created a role that cannot realistically be performed.

A supervisor may be responsible for too many agents. A reviewer may lack access to the evidence required to evaluate a recommendation, or an employee may be expected to intervene while the system executes too quickly for intervention to matter.

These are organizational risk conditions.

Workload, role clarity, decision rights, authority, access to information and workflow design can all determine whether the human contribution succeeds.

This becomes particularly important in agentic environments.

Cybermaniacs uses Supervision Sufficiency to examine whether the human supervisory capacity surrounding an AI-assisted or agentic workflow is adequate for the autonomy, complexity and consequence involved.

A measurement model that examines people but ignores the environment around them can easily misdiagnose a poorly designed control as an employee capability problem.

How should organizations measure human oversight?

If humans are being used as controls, the organization needs evidence about whether those controls work.

The fact that a reviewer exists, an approval step appears in the workflow or a supervisor has been named does not establish control effectiveness.

Cybermaniacs uses Human Control Effectiveness to describe whether a human control relied upon in an AI-enabled process can actually perform the risk-reducing function expected from it.

Measurement should therefore consider the full control environment.

Does the person possess the relevant capability? Do they receive enough information? Are they applying independent judgment? Can they intervene before the consequence occurs, and do they have authority to influence what happens next?

Behavioral evidence is especially useful here.

Override patterns, escalation, response time, exception handling and scenario performance can all help, but no single metric proves the control works. The evidence needs to be interpreted together.

What should be measured when employees supervise AI agents?

Agentic work changes the human measurement problem because people may no longer participate in every action.

The supervisor may set objectives, define boundaries, monitor exceptions and remain accountable while the agent performs much of the execution independently.

That means measurement should reflect the supervisory role.

Does the person understand what authority has been delegated? Can they recognize meaningful deviation? Do they know when intervention is required and what authority they possess? Is their workload compatible with meaningful supervision?

Cybermaniacs uses Human-Agent Accountability for the wider relationship among ownership, delegated authority, supervision, intervention and evidence.

Measurement should help determine whether that relationship is viable in practice.

A person may be named accountable for an agent while being unable to explain what the agent can do, unable to observe meaningful activity or unable to stop it when circumstances change.

That is an accountability problem, but it is also a measurable human-risk condition.

Intervention and override create measurable control conditions

Agentic systems increasingly depend on selective human intervention rather than constant approval.

Our Intervention Points work examines where human control should be able to enter, while intervention triggers examine what conditions should activate that involvement.

From a measurement perspective, the interesting question is what happens when those conditions occur.

Did the person recognize that intervention was warranted? Did the escalation reach someone with appropriate expertise? Was the decision made before the action became irreversible?

Cybermaniacs uses Intervention Authority for the question of who can act and what powers they have, while Human Override covers the stronger case in which human authority supersedes the AI system's behavior.

Those concepts create measurable conditions around authority, recognition and action without requiring the organization to turn every intervention into an employee-performance metric.

The objective is understanding whether the control system behaves as intended.

Skill atrophy should be measured before it becomes a recovery problem

Some human-risk conditions develop slowly enough to escape ordinary monitoring.

Skill atrophy is one.

If AI increasingly performs the analysis, writing, diagnosis, coding or decision preparation that people once performed themselves, the organization may gradually reduce the practice through which expertise was maintained.

That is not automatically a problem.

Skills that no longer matter can disappear without creating meaningful enterprise risk. The issue arises when the organization still relies on people to supervise, verify or recover the AI-enabled work.

Our Guide to How Do You Prevent AI Skill Atrophy in the Workforce? focuses on identifying which capabilities need deliberate preservation.

Measurement should therefore ask not simply whether a skill has declined, but whether a capability the operating model still depends upon is becoming weaker.

This is one of the reasons Human-AI Workforce Resilience is a useful higher-level concept. It asks whether the human capabilities and organizational conditions needed around AI remain viable as the technology takes on more execution.

Should organizations create a Human-AI risk score?

They can, but the score should summarize the model rather than become the model.

Executives need concise signals. A score can help identify where risk is moving, which populations deserve attention and whether interventions appear to be working.

The danger is compression.

Two teams can receive the same score for entirely different reasons. One may lack AI competency, another may demonstrate excessive reliance, while a third has capable employees operating inside a badly designed supervision model.

Those conditions require different action.

This is why Cybermaniacs measures across multiple dimensions rather than treating Human-AI risk as one attribute that can be captured by a survey question.

We do not publish the detailed scoring algorithms, weighting, diagnostic relationships or assessment engine behind that work because those mechanics form part of our proprietary methodology.

The principle is public: a useful score should preserve enough diagnostic information to tell the organization why the risk exists and what type of intervention is warranted.

How often should human risk be measured?

Different human-risk conditions move at different speeds.

Competency may change gradually as people learn, while behavior can change quickly after the introduction of a new system. Reliance may shift as familiarity grows, and supervisory capacity can deteriorate almost immediately if an agent receives broader permissions or workload increases.

That makes one annual measurement cycle insufficient for many AI environments.

A stronger approach combines deeper periodic assessment with operational and event-driven evidence.

A baseline can establish capability, psychological and cultural conditions. Targeted assessments can follow major deployments or role changes, while operational measures such as escalation, intervention and incident behavior can reveal how the system is functioning between formal assessment points.

The exact cadence should follow how quickly the work and risk are changing.

The objective is not continuous measurement for its own sake. It is to avoid relying on an old picture of the workforce after AI has materially changed the work around it.

Measurement should produce Human-AI Risk Intelligence

Collecting more human data is not the goal.

At Cybermaniacs, we use Human-AI Risk Intelligence to describe the evidence organizations need to understand how human capability, behavior, psychology, culture, authority and working conditions interact with AI risk.

The distinction between measurement and intelligence matters.

A confidence score is measurement. A pattern showing that confidence is increasing while verification behavior declines in a high-consequence workflow starts to become intelligence.

Training completion is measurement. Evidence that a population completed training but continues to misidentify conditions requiring escalation tells the organization something operationally useful.

Intervention counts are measurement. Understanding that interventions are falling because supervisors are overloaded rather than because the agent has improved changes the management decision.

Human-AI Risk Intelligence emerges when those signals can be interpreted together and connected to action.

How does measurement fit into Human Risk Assessment & Assurance?

Assessment provides a more structured way to establish the human-risk condition around AI-enabled work.

Cybermaniacs' Human Risk Assessment & Assurance work examines competency, behavior, psychology, culture and organizational conditions rather than inferring human risk from training completion or sentiment alone.

For AI-enabled work, those models can be applied to questions such as AI Judgment, reliance, verification, supervision, authority and Human Control Effectiveness.

The detailed methodology, diagnostic logic, rubrics and scoring remain proprietary because they are part of how Cybermaniacs translates evidence into customer recommendations.

What organizations should expect from a strong assessment is transparency about the dimensions being examined and a result that helps explain why risk exists.

The output should support decisions about learning, governance, workflow, supervision, culture or technical controls rather than simply creating another dashboard.

What should CISOs measure first?

Start with the human responsibilities the organization most depends upon.

If employees are expected to verify important AI output, measure whether they can and whether they do. If managers supervise agents, examine whether they understand delegated authority and can recognize when intervention is required. If the governance model relies heavily on human approval, determine whether those approvals involve real judgment or routine confirmation.

Then look for the conditions explaining those outcomes.

Does the issue arise from capability, reliance, confidence, culture, workload, unclear authority or poor workflow design?

That sequence matters because measurement becomes much more useful when it begins with the risk decision rather than with whatever data happens to be easiest to collect.

The result should tell leaders something they can act on.

Measure the human system, not simply the human

The most important shift in measuring Human-AI risk is recognizing that people do not operate independently of the systems and environments around them.

A person can possess excellent judgment and still operate inside a workflow that gives them no time to exercise it. Another can understand when escalation is required while working in a culture that discourages challenge, and a capable supervisor can become ineffective when the volume of autonomous work exceeds available attention.

Those are human-risk conditions even though they cannot be explained by employee competency alone.

This is why Cybermaniacs approaches AI-enabled human risk as a system across competency, behavior, psychology, culture and organizational conditions.

The purpose of measurement is not to determine whether humans are the weak link in AI.

It is to understand whether the human and organizational system surrounding AI is strong enough for the work the technology has been allowed to do.

That evidence is what allows Human Risk Management to support Secure AI Enablement: enabling AI to scale while preserving the judgment, accountability and control conditions the organization still depends upon.