ARTICLE Metrics

What Is Human-AI Risk Intelligence and What Should CISOs Measure?

SHARE
By Team CM · Oct 28, 2026, 8:18:48 PM
What Is Human-AI Risk Intelligence and What Should CISOs Measure?

Most enterprises know more about their AI systems than they do about the people working with them.

They know which applications have been approved, where models are deployed, which teams are experimenting with agents and how quickly adoption is growing. More mature programs may also have model evaluations, risk registers, usage telemetry, incident reporting and increasingly sophisticated technical monitoring.

Then the AI reaches a person.

An employee decides whether to trust a recommendation. A supervisor decides whether an agent can continue. A manager reviews an exception. Someone notices that an answer looks plausible but wrong, while someone else accepts the same kind of answer because the system has been right often enough before.

That human layer is much harder to see.

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 important word is intelligence. Training completion rates, survey responses, usage statistics and confidence scores are all data. They become useful risk intelligence when the organization can connect them to what people are actually being asked to do, how they behave in those situations and what needs to change when risk begins to emerge.

For CISOs, CAIOs and risk leaders, the challenge is therefore not collecting more AI metrics. Most organizations are heading toward an abundance of those already. The harder question is which human signals explain whether the AI-enabled operating model is actually working.

Why do CISOs need a human intelligence layer for AI?

AI risk increasingly appears through the interaction between technology and work.

A model may behave exactly as designed while an employee relies on it inappropriately. An agent can stay inside its technical permissions while operating beyond what the business owner believed had been delegated. A human review step may appear exactly where governance said it should and still provide very little meaningful scrutiny because the reviewer lacks time, expertise or confidence to challenge the system.

Technical monitoring will identify some of those problems, but it will not explain all of them.

NIST's work on post-deployment AI monitoring has increasingly emphasized this socio-technical problem, including human-AI feedback loops, the difficulty of scaling human-led monitoring and the challenge of combining automated signals with human validation. That matters because AI behavior does not exist separately from the people using, supervising and responding to it.

Human-AI Risk Intelligence looks specifically at that interaction.

Are people becoming better at supervising AI, or simply more accustomed to it? Is confidence increasing because capability has improved, or because familiarity has made employees less cautious? Are reviewers exercising judgment, or has review gradually become a formal step everyone expects to approve?

These are not secondary workforce questions. They affect whether controls work, whether emerging risks are detected and whether governance survives contact with everyday operations.

How is Human-AI Risk Intelligence different from workforce analytics?

Workforce analytics can tell an organization a great deal about skills, roles, engagement, performance and organizational change. AI adoption programs can add another layer of information about usage, training and readiness.

Human-AI Risk Intelligence has a narrower purpose.

It focuses on the human conditions that affect risk and control in AI-enabled work. Capability matters because people may need to evaluate AI-generated work. Confidence matters because it influences reliance. Culture matters when people technically have permission to challenge AI but believe doing so will slow the team down, irritate leadership or make them look resistant to change.

The same workforce signal can therefore mean very different things once risk enters the picture.

High confidence might indicate that employees have become capable, thoughtful AI users. It could also mean that they have become extremely comfortable with a system whose limitations they no longer examine carefully. A useful intelligence layer needs enough context to tell those conditions apart.

That distinction is central to the Cybermaniacs approach. Human-AI Risk Intelligence is not a new label for employee analytics. It is the evidence layer used to understand whether the human side of AI-enabled work is strengthening or weakening enterprise control.

What should organizations understand about AI reliance?

Reliance is one of the first behaviors worth measuring because it rarely changes dramatically overnight.

When a system performs well, people reasonably begin to trust it. They verify less often, stop recreating work the AI has already completed and gradually incorporate the system into how they understand the task itself.

None of that is automatically problematic.

The risk appears when reliance no longer fits the consequence of the work, the quality of the system or the person's ability to recognize an error. Someone may use an AI assistant dozens of times during routine administrative work with almost no meaningful downside when it makes a mistake. The same level of trust matters considerably more when AI output influences a financial decision, employment action, customer commitment or production change.

Our Guide to How Do You Know When Employees Are Over-Relying on AI? examines this through the idea of calibrated reliance. The question is not simply whether employees trust AI, but whether the level of trust remains appropriate to the work.

Human-AI Risk Intelligence should therefore help organizations see how trust changes behavior. Do people verify differently when consequences increase? Can they recognize when independent evidence is warranted? Does repeated system success make them less likely to question the unusual case?

Those questions reveal considerably more than asking employees whether they “trust AI” on a five-point scale and admiring the chart.

How do you know whether people still have the judgment to supervise AI?

This question becomes harder as organizations ask people to evaluate work they perform less often themselves.

An analyst who once created the first draft may now review one produced by AI. A developer may supervise code generated by an agent. A manager may make a decision based on analysis assembled through a process they never personally performed.

That can be highly productive, but it also changes how expertise develops and how people recognize weak work.

This is why Cybermaniacs separates AI literacy from AI Judgment. We use 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.

Measuring that capability requires more than testing whether someone knows that AI can hallucinate. The evidence needs to get closer to the decisions people actually make: when they verify, when they challenge, when they accept the system's reasoning and when they recognize that the situation has moved outside the conditions where AI should be trusted.

Our Guide to How Do You Measure Whether Employees Can Exercise Good AI Judgment? goes deeper into that measurement problem, while How Do You Prevent AI Skill Atrophy in the Workforce? looks at the longer-term question of whether AI is removing some of the practice through which that judgment was historically developed.

The two issues are closely connected. Organizations may depend increasingly on human judgment while simultaneously changing the work that used to build it.

What should organizations measure about human oversight?

A human review step is one of the easiest AI controls to document and one of the easiest to overestimate.

The organization knows a person is present. Governance knows which role is responsible. Training records show that the reviewer completed the required material. Those facts tell us that the control exists, but they do not establish that it works.

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.

That creates a more useful measurement question.

What is the human expected to recognize or decide? Do they possess the capability to do it? Does the workflow give them enough context and time? Can they intervene when something is wrong? What evidence shows that those interventions occur when conditions genuinely warrant them?

Our Guide to What Is Human Oversight in AI? establishes the concept, while How Do You Build Effective Human Oversight for AI? addresses the design problem. Human-AI Risk Intelligence provides the evidence needed to determine whether that design remains effective once people begin working inside it.

This distinction matters because the organization can have perfectly respectable governance documentation and a human control that has quietly stopped reducing much risk.

What does agentic AI change about human risk measurement?

Quite a lot, because the human may no longer participate in every individual decision.

Once agents move from assistance toward execution, the person's job can shift toward setting objectives, establishing constraints, supervising exceptions and remaining accountable for outcomes produced through a chain of autonomous activity. Prompt counts and AI usage statistics become considerably less informative when the person is no longer directly interacting with the system at every step.

Human risk measurement needs to follow that change in role.

Does the accountable person understand what authority the agent has been given? Do they know which conditions should trigger escalation? Can they interpret what the agent is doing when the situation becomes unusual? Do they have enough time and attention to supervise the amount of autonomous activity assigned to them?

Our Guide to What Is Human-Agent Accountability? looks at the relationship among ownership, delegated authority, supervision and intervention. The measurement question underneath it is whether those relationships work in practice rather than merely appearing in the governance model.

That can include evidence about escalation behavior, intervention patterns, supervisory workload, decision authority and whether the people carrying accountability understand what they are accountable for.

Agentic AI does not eliminate human risk.

It changes where the human contribution sits.

Why should CISOs care about confidence and psychological safety?

People can recognize risk and still do nothing about it.

An employee may correctly identify that an AI recommendation looks wrong, understand the policy and know where escalation should go. The decision then enters the social environment of the organization.

Will challenging the system slow everyone down? Is leadership pushing hard for adoption? Has the employee seen colleagues rewarded for speed while questions are treated as resistance? Does the reviewer feel genuinely authorized to disagree with a system that has already been approved by people considerably more senior than they are?

These are human risk conditions even though none of them appear in model telemetry.

Cybermaniacs therefore includes psychological safety, leadership behavior, perceived control and group norms within the human risk environment surrounding AI. The point is not to turn every AI assessment into a culture survey. It is to recognize that a control dependent on challenge is only as strong as the environment in which challenge occurs.

If escalation is part of the control design, the organization needs evidence that people believe escalation is actually available to them.

How should CISOs distinguish risk indicators from activity metrics?

A useful test is whether the metric helps explain a risk decision.

“Most employees completed AI training” describes activity. It can be useful for understanding reach and program participation, but it says relatively little about whether people can recognize when an AI-supported recommendation requires independent verification.

Similarly, “employees report high confidence using AI” is interesting. Confidence increasing while independent verification behavior declines tells us something more useful because the two signals begin to explain a potential risk condition.

The distinction is not that completion, usage or sentiment measures have no value. They provide important context and can show how quickly AI adoption is moving through the organization.

The problem begins when those metrics are asked to prove outcomes they were never designed to measure.

Security awareness programs made this mistake for years. An organization could demonstrate excellent annual-training completion and still know remarkably little about how employees behaved when confronted with an actual threat.

AI gives us an opportunity not to repeat the whole experiment with shinier dashboards.

Does Human-AI Risk Intelligence require one Human Risk Score?

No.

Executive-level signals can be useful. Leaders need to know whether risk is moving in the right direction, which populations deserve attention and where intervention is required.

The danger appears when aggregation removes the explanation.

Two business units could show the same overall Human-AI risk score while facing completely different problems. One may lack sufficient understanding of AI limitations. The other may have capable employees supervising so much automated work that meaningful review has become unrealistic.

Those conditions should not lead to the same intervention.

The first may require learning or clearer guidance. The second may require workflow redesign, changes to exception handling, stronger supervision or a different level of agent autonomy.

Cybermaniacs' Human Risk Assessment & Assurance work therefore examines multiple dimensions of the human system rather than treating a single score as the measurement model itself. We deliberately do not publish the detailed scoring logic, diagnostic relationships or assessment methods behind that work because those mechanics are part of our proprietary methodology.

The public principle is more important anyway: good Human-AI Risk Intelligence should explain enough about why risk exists to help the organization decide what to change.

What should a Human-AI risk dashboard actually show?

A useful dashboard should help the organization move from asking “How much AI are we using?” toward asking “How well is the Human-AI system operating?”

The exact measures will depend on the role and use case, but the evidence usually spans several related areas. Capability asks whether people possess the AI understanding, domain expertise and judgment required by their role. Behavior looks at what they actually do around reliance, verification, escalation and intervention. Psychology helps explain confidence, risk perception, trust, self-efficacy and perceived control, while culture shows whether leadership behavior and team norms support appropriate challenge.

Authority and work design add another layer because responsibility only works when decision rights, workload and visibility are aligned. Human Control Effectiveness examines whether people relied upon as part of AI governance can actually perform the control expected from them, while Human-AI Workforce Resilience looks further ahead at whether critical human capabilities remain viable as AI assumes more execution.

Those dimensions are more useful than a universal shopping list of metrics because the purpose is diagnosis rather than dashboard decoration.

Different forms of AI-enabled work create different forms of human risk.

How often should Human-AI risk be measured?

The cadence should follow how quickly the work is changing.

Some workforce conditions move slowly enough for periodic deeper assessment. Others can change rapidly after a major AI rollout, a workflow redesign or an agent receiving additional permissions.

A sensible operating model therefore combines different forms of evidence.

Deeper assessment can establish competency, behavioral, psychological and cultural conditions. Operational data can show escalation patterns, overrides, exceptions and incidents. More targeted measurement can follow significant changes in systems, roles, authority or supervisory responsibility.

This is where NIST's emphasis on post-deployment monitoring becomes particularly useful. Pre-deployment testing happens under controlled conditions, while live environments introduce changing contexts, dynamic inputs and interactions that designers did not necessarily anticipate.

Humans change after deployment too.

They become more familiar with systems, discover shortcuts, build confidence, develop team norms and absorb new workloads. A baseline taken before launch does not tell the organization indefinitely how the resulting Human-AI system behaves.

Who should own Human-AI Risk Intelligence?

Probably nobody alone.

The CISO brings risk discipline and a direct interest in control effectiveness. The CAIO or AI governance function understands the systems and use cases. Technology teams can provide operational signals, while HR and learning functions understand workforce capability and development. Business leaders know how the work itself is changing.

The problem is not that several functions are involved.

The problem is when their evidence remains disconnected.

AI governance may know where a human control is required while L&D knows that everyone completed the training and neither can tell whether the control actually works. A security team may see unusual agent behavior while the business owner understands the process implications but lacks access to the same evidence.

Human-AI Risk Intelligence becomes useful when those signals can be interpreted together.

This is one reason Cybermaniacs sees Human Risk Management as an important operating layer in AI transformation. HRM already asks how competency, behavior, psychology, culture and organizational conditions affect risk. AI changes the decisions, tools and workflows surrounding those factors, but it does not make the human system disappear.

Where does Human-AI Risk Intelligence fit into Secure AI Enablement?

Secure AI Enablement is about helping organizations use AI productively while maintaining the human capability, judgment, accountability and control conditions needed to keep that use within acceptable risk.

That requires evidence.

Organizations need to know where the workforce is ready, where risk is emerging and which intervention is likely to address the underlying problem rather than merely generate another training assignment.

Cybermaniacs' AI Enablement & Change (AIECM) work helps organizations understand workforce readiness, capability, confidence, behavior and adoption as AI enters everyday work. Agentic Readiness & Change (ARC) extends that work into environments where agents are changing roles, workflows, supervision and accountability, while Human Risk Assessment & Assurance provides the deeper measurement layer when organizations need evidence about competency, behavior, psychology, culture and organizational conditions.

Human-AI Risk Intelligence connects those signals to decisions.

It helps an organization distinguish between a problem that requires education, one that requires workflow redesign, another that reflects cultural conditions, and one that should be addressed through clearer authority, stronger supervision or a different technical control.

That is considerably more useful than discovering that the workforce's AI risk score moved four points and scheduling a meeting to debate what the four points might mean.

What should CISOs ask next?

The first question is not whether the organization has enough AI metrics.

Most enterprises are likely to have more of those than anyone particularly wants.

The more useful question is whether the available evidence explains what is happening between the AI systems and the people expected to use, supervise and remain accountable for them.

Can the organization see when reliance is becoming inappropriate? Can it tell whether employees retain the judgment needed to review AI-supported work? Does it know whether people serving as human controls can recognize and act on the conditions governance expects them to catch? Can it distinguish a capability problem from a workflow problem, or a behavioral problem from a cultural one?

Those are the questions Human-AI Risk Intelligence is intended to answer.

Because once AI becomes part of normal work, understanding the technology is only half the risk picture. The other half is understanding what the technology is doing to the people around it, what those people are doing in response, and whether the resulting system is becoming more resilient or simply more dependent.

TAGS: Metrics