AI risk assessments are getting better at asking about models, data, privacy, security, vendors and regulatory exposure.
They are less consistent when they reach the humans.
A questionnaire may ask whether users receive training or whether human oversight exists, then move on. That leaves a large part of the operating risk unexplored because people increasingly determine whether AI-generated information is trusted, challenged, escalated, overridden or allowed to become action.
The human contribution becomes even more consequential with agents. Someone decides what authority an agent receives, supervises activity that may be happening faster than they can personally review, handles exceptions and remains accountable for outcomes produced through a combination of human and machine decisions.
At Cybermaniacs, we think an AI risk assessment should therefore examine the human system around the technology, not merely confirm that people have been trained.
That means looking at capability, behavior, psychology, culture, authority, workload and the way work itself is changing. The objective is to understand whether the humans relied upon by the AI operating model can perform the roles governance assumes they will perform.
AI risk is rarely created by the model in isolation.
A technically sound system can still create poor outcomes when employees use it outside the intended context, accept recommendations too readily, misunderstand their responsibility or fail to intervene when something unusual happens. A well-designed human oversight control can deteriorate if reviewers become overloaded or lose the expertise needed to challenge the system.
NIST's AI Risk Management Framework takes a deliberately socio-technical view of AI risk, emphasizing that identifying and managing risk requires perspectives across the AI lifecycle rather than evaluation of the technology alone. NIST's newer ARIA evaluation approach makes the human dimension even more explicit by combining model testing and red teaming with user testing when assessing real-world AI applications. NIST's ARIA Evaluation Planning Manual is a useful signal that trustworthy AI evaluation increasingly has to include what happens when people interact with the system.
That does not mean every AI risk assessment needs to become an organizational psychology study.
It does mean that “users completed training” is rarely enough evidence to establish that the human side of the control environment is working.
Start with what people are actually expected to do.
A general employee using an approved assistant may need enough AI literacy to recognize common limitations, protect sensitive information and know when additional verification is warranted. A financial analyst using AI in a consequential decision needs domain expertise and judgment as well. Someone supervising an autonomous agent may need to recognize when the agent has moved outside its mandate and understand which intervention authority they possess.
The assessment should therefore distinguish knowledge from capability.
Knowing that AI can hallucinate is useful. Recognizing that a polished answer is resting on an inappropriate assumption requires something more.
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.
For higher-consequence use cases, an AI risk assessment should ask whether the people relied upon by the workflow possess the judgment and domain capability appropriate to the decisions they are being asked to make.
The standard does not need to be identical for every employee. It should follow the work and the consequence.
AI use changes behavior gradually.
An employee may begin by checking almost everything the system produces. As it performs well, they reasonably check less. Eventually, verification may become lighter without anyone formally deciding that the control should change.
That is one reason over-reliance deserves explicit attention in an AI risk assessment.
The useful question is not simply whether employees trust AI. Organizations should understand whether trust and verification remain appropriate to the task.
Our Guide to How Do You Know When Employees Are Over-Relying on AI? describes this as calibrated reliance. A low-risk system that performs consistently should earn trust. The concern appears when that same level of trust carries into situations where consequence, uncertainty or system performance requires more scrutiny.
Verification needs similar treatment.
An assessment should ask whether people know which outputs require independent evidence, whether they can identify assumptions worth checking and whether the workflow gives them a practical way to verify important claims.
Our Guide to How Should Employees Verify AI-Generated Work? goes deeper into the distinction between simply reviewing an output and establishing enough independent evidence to rely on it.
These behaviors matter because human review is only useful when the person remains capable of disagreeing with the system.
If the AI control framework relies on humans, yes.
Many AI governance models contain some variation of “human review required.” The phrase can hide an enormous range of operating conditions.
One reviewer may have deep expertise, clear decision authority and enough time to examine a handful of consequential cases. Another may receive hundreds of system-generated approvals each day with limited information and little reason to expect any of them to be wrong.
Both workflows contain human oversight.
They do not provide the same control.
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.
An AI risk assessment should therefore look beyond the existence of a checkpoint. It should consider what the human is expected to recognize, whether they possess the necessary capability, what information they receive, how much workload they carry and what authority they have when they decide something is wrong.
Our Guide to How Much Human Oversight Does AI Actually Need? explores why the required level of oversight should scale with consequence, autonomy, reversibility and the human judgment still needed.
The assessment question is whether the chosen model works under realistic conditions.
Agents shift the problem because people may no longer review every output before action occurs.
The human role moves toward setting objectives, defining boundaries, supervising exceptions and taking responsibility for activity that may span multiple systems.
That creates a different set of human risks.
Does the accountable person understand what authority the agent has been given? Do supervisors know which conditions should cause intervention? Can they interpret enough of the agent's activity to recognize when the situation has changed? Is there a clear path for escalation when the issue exceeds their own authority?
Identity and permissions matter enormously, but they are only one side of the problem.
Our Guide to What Is Human-Agent Accountability? describes the broader relationship among ownership, delegated authority, supervision, intervention and evidence. How Do You Assign Accountability for AI Agent Actions? takes that concept into the operating model.
An agent risk assessment should examine whether accountability remains usable once the technology begins acting autonomously.
A name in an inventory is useful. A person who understands what the agent can do and has meaningful authority when something changes is considerably more useful.
This is a slower-moving risk, which makes it easy to leave out.
AI may perform work perfectly well today while gradually reducing the human practice through which expertise was previously developed or maintained.
That matters when the organization still expects humans to evaluate, supervise or recover the work later.
IBM's September 2026 global workforce research found that 60% of employees worry AI is eroding their skills, while 71% of CHROs identified the ability to supervise, validate and override AI outputs as the workforce's most essential AI-era capability. The study also found growing concern around critical thinking and judgment. IBM's 2026 CHRO study illustrates the tension clearly: organizations want more sophisticated human supervision while employees worry some of the capabilities needed for that supervision are weakening.
An AI risk assessment should not assume every automated skill needs to be preserved.
It should ask which capabilities the organization will continue to depend on after automation.
Our Guide to How Do You Prevent AI Skill Atrophy in the Workforce? focuses on that distinction. If the human remains the fallback control when AI reaches an unusual condition, some independent human capability needs to remain viable.
That is a resilience issue, not nostalgia for manual work.
They should be included where the control model depends on people challenging, escalating or overriding AI.
An employee can possess excellent AI knowledge and still hesitate to question a system if doing so feels socially expensive.
This becomes particularly important when leadership is strongly promoting adoption. The formal message may be “challenge AI when necessary,” while the operating message is “we need everyone moving faster.”
IBM's 2026 study found that 41% of CHROs believe employees may not feel safe challenging or overriding AI outputs. It also found that employees reported substantially greater confidence in questioning AI where responsibility for which decisions remained human-led was more deliberately shared. That makes culture directly relevant to whether human controls are likely to function in practice.
Cybermaniacs' Human Risk Management approach therefore looks at factors such as psychological safety, leadership behavior, group norms and perceived control alongside competency and behavior.
If the organization depends on escalation as a safeguard, the assessment should consider whether employees genuinely believe escalation is expected and supported.
A control that people are technically allowed to use but socially discouraged from exercising is weaker than the governance document suggests.
Human attention has become one of the less glamorous constraints on AI scaling.
AI can increase output far faster than organizations can increase experienced reviewers.
IBM reports that 80% of CHROs believe AI creates “invisible” employee work such as validating recommendations, fixing errors, supplying context and managing exceptions. Forty-two percent of employees said AI had increased their workload or created work that goes unrecognized.
That matters for risk assessment because human review quality changes when volume increases.
A supervisor who can thoughtfully examine twenty unusual cases may behave very differently when the system begins generating two hundred. Review becomes faster, defaults become more attractive and unusual but subtle problems receive less attention.
An AI risk assessment should therefore ask whether the amount of human attention available matches the role assigned to people within the control model.
For agentic systems, our Guide to Supervision Sufficiency examines this relationship directly. The relevant question is whether the amount and quality of supervision remain appropriate for the autonomy, complexity and consequence of the work.
A human control designed without considering human capacity is eventually a workload problem wearing a governance badge.
If an organization expects humans to intervene, the assessment should establish what causes intervention to occur and what the person can actually do once it does.
Some triggers can be technical: a transaction crosses a threshold, an agent requests additional permissions or an action becomes difficult to reverse.
Others require human recognition. An employee notices that the situation is unusual, the AI's reasoning conflicts with important context or the agent is technically following instructions while drifting away from the intended outcome.
Our Guide to What Should Trigger Human Intervention in AI Systems? explores those conditions in detail.
The risk assessment should also distinguish intervention from override.
A reviewer may be able to raise a concern without possessing the authority to stop the system. Someone else may be able to reject an individual action while lacking the ability to change the agent's mandate.
Those differences matter.
An effective assessment asks whether the human authority available matches the responsibility the organization has assigned.
Carefully.
There is a strong temptation to compress human risk into a single red, amber or green score because executive reporting needs simplicity.
Summary signals are useful.
They become dangerous when the aggregation hides why the risk exists.
Two teams can receive the same apparent Human-AI risk rating while facing completely different problems. One may lack AI literacy, another may demonstrate excessive reliance, and a third may have excellent individual capability but work inside a culture where nobody feels comfortable challenging the system.
The interventions should be different.
Cybermaniacs' Human Risk Assessment & Assurance approach examines competency, behavior, psychology, culture and organizational conditions because the diagnostic structure underneath a score is what makes the measurement actionable.
We do not publish our detailed scoring models, weighting, diagnostic mappings or assessment engine as an open recipe. Those are part of the proprietary methodology we use with customers.
What buyers should expect, however, is a measurement approach capable of distinguishing why human risk exists rather than simply announcing that it does.
An assessment provides a point-in-time view.
Human-AI Risk Intelligence is the broader evidence layer that helps an organization understand how those conditions evolve as AI becomes part of normal work.
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.
That matters because people change after deployment.
They become more familiar with the technology. Trust grows or falls. Verification habits change. Teams develop norms. Agents receive new permissions. Workloads increase.
NIST's current AI evaluation work points in the same direction. Its TEVV-Athlon framework is designed to assess real-world AI outcomes across conventional, generative and agentic systems, while ARIA combines model, adversarial and user testing to create a more holistic evaluation approach. The common theme is that trustworthy AI cannot be understood solely through static model testing.
Human conditions should not be treated as static either.
Probably not one function alone.
Security and GRC bring risk methodology and control discipline. AI governance understands the use case and model. HR and learning functions understand workforce development. Business owners know the work and its consequences.
Human Risk Management can provide the connective layer because it is specifically concerned with how people, behavior, capability, culture and organizational conditions create or reduce enterprise risk.
That becomes increasingly important as AI governance places more responsibility on humans while simultaneously redesigning the work those humans perform.
Cybermaniacs' AI Enablement & Change (AIECM) addresses readiness, capability, confidence, behavior and adoption as AI enters the workforce. Agentic Readiness & Change (ARC) extends that work into changing roles, workflows, supervision and Human-Agent Accountability.
The AI risk assessment should connect those workforce questions to the larger control environment rather than leaving them in a separate change-management workstream.
Start with the human responsibilities embedded in the use case.
Where does the organization expect people to verify, supervise, decide, intervene, escalate or remain accountable?
Then examine whether the surrounding human system can support those expectations.
Do people have the required AI and domain capability? Is reliance appropriate to the task? Can they verify important outputs? Is workload compatible with meaningful oversight? Are authority and accountability aligned? Does the culture support challenge? Are critical skills being preserved as AI removes more of the underlying work?
Not every AI use case requires deep assessment across every one of those areas.
A low-risk productivity assistant should not receive the same treatment as an autonomous agent operating across sensitive enterprise systems.
The depth should follow the risk.
But the human questions should be present.
Because once AI becomes part of operational work, the human layer is no longer merely the population being trained on the technology.
It is part of the AI control environment itself.