ARTICLE Metrics

How Should CISOs Build a Secure AI Enablement Program?

SHARE
By Team CM · Oct 28, 2026, 8:51:53 PM
How Should CISOs Build a Secure AI Enablement Program?

How Should CISOs Build a Secure AI Enablement Program?

Most enterprises are discovering that AI adoption does not wait politely for governance to finish.

Employees experiment with approved tools, unapproved tools and whatever happens to be built into the software they already use. Business teams begin redesigning workflows before security has finished defining policy. Then agents arrive, bringing identity, permissions and autonomous execution into a conversation that was already moving quickly.

That leaves the CISO balancing two bad outcomes. Move too slowly and AI adoption develops around the control environment. Push too aggressively toward restriction and security becomes the function everybody is quietly trying to route around.

A Secure AI Enablement program needs to solve a different problem: how does the organization make useful AI adoption easier while ensuring that increasing capability, autonomy and authority are matched by appropriate human and technical controls?

At Cybermaniacs, we use Secure AI Enablement for this wider discipline: enabling people and organizations to use AI productively while maintaining the human capability, judgment, accountability and control conditions required to keep that use within acceptable risk.

That is broader than AI security alone. It brings governance, workforce capability, Human Risk Management, identity, data protection, operational control and organizational change into the same operating model.

Where should a CISO start with Secure AI Enablement?

Start with the work rather than the tool.

Organizations often begin with an inventory of approved AI platforms, followed by acceptable-use rules and perhaps a training module explaining what employees should and should not do. Those are useful components, but they tell the CISO surprisingly little about how AI is changing the business.

A better starting point is to understand where AI is entering work and what responsibility is moving with it. An employee using AI to improve the wording of an internal email creates one kind of risk. A claims analyst using AI to interpret information that affects a customer creates another. An engineering agent with permission to modify production systems changes the problem again.

The technology may be related while the operating conditions are completely different.

NIST's AI Risk Management Framework is helpful here because it treats AI risk as contextual rather than assuming one control posture fits every system. The framework is intended to help organizations govern, map, measure and manage AI risk across the lifecycle, while the Generative AI Profile recognizes that different Human-AI configurations may require different levels of oversight, review and management.

For Secure AI Enablement, that translates into a practical question: what is AI changing about the work, the decision and the authority involved? Once that is clear, the control design becomes much easier to reason about.

Secure AI Enablement needs a risk model that extends beyond the technology

Traditional security programs are very good at thinking about assets, identities, data flows, vulnerabilities and controls. AI introduces all of those familiar concerns and then complicates them with human judgment.

An employee may have legitimate access to a document and still place its contents into an inappropriate AI system. A model may produce technically plausible output that is badly suited to the business decision. A human-review checkpoint may exist but provide little protection because employees have become accustomed to approving what the system recommends.

Agents stretch the model further because AI can begin exercising authority rather than merely influencing a person. That means a Secure AI Enablement program has to consider several forms of risk at once: technical compromise, inappropriate data exposure, misuse, poor decisions, over-reliance, weak supervision, unclear accountability and the possibility that important human capability erodes as AI takes over more of the underlying work.

Our Guide to What Human Risks Should Be Included in an AI Risk Assessment? explores that human layer in more detail.

The important point for CISOs is that a good AI risk model needs to describe the whole socio-technical system around a use case.

This direction is increasingly visible in the standards environment as well. ISO/IEC 42001 establishes an AI management-system approach for governing AI-related risks and opportunities across the organization, while ISO/IEC 42105 focuses specifically on human control and monitoring of AI systems across the lifecycle.

The CISO does not need to own every dimension personally, but somebody has to make sure those dimensions connect.

Build guardrails around authority, not simply around access

Early enterprise AI governance focused heavily on which tools employees were allowed to use. Agentic AI makes that question insufficient because the more important issue becomes what the AI is allowed to do once it has access.

An agent may be able to read data, call tools, update records, send communications or make changes in connected systems. Those capabilities should not be treated as one generic block of access.

Microsoft's current AI agent shared-responsibility model separates agent identity, tool permissions, action authorization, human approval, memory controls and audit logging. It also makes clear that organizations retain responsibility for defining what agents are permitted to do, which sensitive actions require approval and who remains accountable for autonomous behavior.

For CISOs, this means Secure AI Enablement needs a graduated authority model. Low-risk assistance can operate with broad usability and relatively light intervention, while systems that gain access to sensitive data, consequential decisions or external actions should face stronger expectations around authorization, oversight, traceability and reversibility.

This is where Human-Agent Accountability becomes part of the security model. An agent should not receive authority merely because the technology can support it; the organization should be able to explain where that authority came from, who is permitted to delegate it and who owns the outcome.

Human oversight has to be designed before it can be relied upon

The phrase “human in the loop” has become reassuring enough that it is often accepted without much further inspection. Secure AI Enablement requires more skepticism.

A human checkpoint reduces risk only when the person involved can understand what is happening, recognize when something is wrong and do something useful about it. That depends on role clarity, competence, workload, information and authority.

Microsoft's current responsible-agent guidance recommends human approval for actions that are difficult to reverse or that affect people, money or compliance, while also stressing that sensitive or ambiguous cases need clear escalation paths and enough context for reviewers to make meaningful decisions.

That sounds straightforward until the system scales.

A human reviewer who receives five unusual cases a week may provide excellent judgment. Give that same person five hundred AI-generated approvals and the control can gradually become a queue-clearing exercise.

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.

Our Guides to How Much Human Oversight Does AI Actually Need? and What Should Trigger Human Intervention in AI Systems? look at how consequence, autonomy, reversibility and changing conditions should shape that design.

Secure AI Enablement needs to know where humans add judgment, when that judgment should enter and whether the operating environment allows them to exercise it.

Workforce readiness belongs inside the control model

Training usually enters AI programs after the policies are written, but that sequence deserves reconsideration.

If an organization's control environment depends on employees recognizing inappropriate AI use, verifying outputs, understanding escalation conditions and challenging unreliable recommendations, workforce capability is part of the control design rather than a separate adoption workstream.

A CISO should therefore want to know more than whether employees completed an AI literacy course. Can people apply judgment when the situation stops looking like the examples in the training module? Can they tell when an AI-generated recommendation requires independent verification? Do they understand which decisions remain theirs, and will they challenge AI when the output conflicts with domain knowledge or context?

Our Guide to AI Literacy vs AI Judgment makes this distinction because awareness and judgment solve different problems.

This is also where security intersects with organizational culture. A workforce can be technically authorized to challenge AI while operating inside a culture where speed, adoption or executive enthusiasm makes challenge uncomfortable. A company can spend heavily on AI enablement while accidentally teaching employees that success means using more AI rather than using it well.

Secure AI Enablement should create permission for intelligent skepticism without turning skepticism into reflexive distrust. The capability we want is the ability to decide when the technology deserves confidence and when the situation deserves a closer look.

Safe adoption needs to be easier than shadow adoption

Security programs lose leverage when the approved path is substantially harder than the unofficial one, and AI magnifies the problem because useful consumer tools are available to employees almost immediately.

A workable enablement program therefore needs more than restrictions. Employees should know which tools they can use, what data can be placed into them, which use cases require additional review and where to go when the approved environment does not meet a legitimate business need.

That final point matters because a governance model with no practical mechanism for saying “yes, but safely” eventually teaches people that “don't” is easier to ignore than to operationalize.

ISO/IEC 42001 is useful here because its management-system model treats responsible AI as a balance between managing risk and enabling organizational objectives rather than as pure restriction. Microsoft's agentic maturity guidance makes a similar point from an operational perspective: higher maturity comes from controls, observability, human escalation and lifecycle ownership being standardized enough to support adoption without forcing every use case through a bespoke approval process.

For the CISO, that suggests a useful measure of program quality: how often does the safe path help the business get where it is trying to go?

If every valuable use case becomes a custom negotiation, the program will struggle to scale regardless of how elegant the policy is.

The program needs continuous evidence, not an annual training report

AI systems change, and so do the people using them.

A workforce may be cautious during the first month of deployment and considerably more trusting six months later. Teams discover shortcuts, agents receive additional permissions, new models replace old ones, and workflows that began as experiments become dependencies.

Secure AI Enablement therefore needs feedback loops that combine technical and human evidence.

Security telemetry can show usage patterns, incidents, data-loss events, unauthorized applications, unusual agent actions and policy violations. The human system tells us something different: whether employees are becoming better at identifying unreliable AI output, whether verification is declining in high-consequence workflows, whether supervisors are receiving more escalations than they can reasonably investigate, and whether important capabilities are weakening as AI takes on more of the work.

Our Guide to How Do You Prevent AI Skill Atrophy in the Workforce? examines one of the less obvious consequences. AI can improve productivity while gradually reducing the practice through which human expertise was previously maintained.

That risk will not appear in a data-loss dashboard, but it still matters if the organization expects the same humans to supervise, validate or recover AI-enabled work later.

Cybermaniacs uses Human-AI Risk Intelligence for the wider evidence organizations need to understand how human capability, behavior, psychology, culture, authority and working conditions interact with AI risk.

The purpose is not workforce surveillance. It is to understand whether the human control environment is becoming stronger or weaker as adoption grows.

Agentic AI should trigger an upgrade in the operating model

Organizations can adopt copilots without fundamentally changing who is allowed to act. Agents are different because they introduce software actors that can receive objectives, maintain state, use credentials, call tools and perform sequences of work with progressively less human involvement.

That pushes Secure AI Enablement into areas traditionally associated with identity, privileged access, application security and incident response while simultaneously changing human roles around supervision and accountability.

Microsoft's 2026 agentic maturity guidance reflects this evolution. Its higher-maturity model combines identity and access controls with telemetry, human escalation, responsible AI governance, lifecycle ownership, training and change management.

The important point is not that Microsoft has produced another maturity model. It is that technical and human disciplines begin converging as autonomy increases.

A company cannot solve agentic risk with employee education alone, nor can it solve the problem solely through IAM if people do not understand the authority they are delegating or what they are expected to supervise.

Cybermaniacs' Agentic Readiness & Change (ARC) work focuses on this wider transition: changing roles, workflows, supervision, delegated authority, Human-Agent Accountability and the human capabilities required around autonomous systems.

The technology may be agentic, but the operating model still belongs to the organization.

Who should own Secure AI Enablement?

Probably no single function, which is inconvenient but realistic.

The CISO owns important pieces of the risk architecture. The CIO controls much of the technology environment. The CAIO or AI governance function may shape standards and adoption, while legal and privacy interpret obligations. HR and learning influence workforce capability, and business leaders determine how work actually changes.

Secure AI Enablement therefore needs explicit shared ownership rather than a vague cross-functional committee that meets occasionally and produces increasingly ambitious slide decks.

The CISO's most valuable role may be to help define the operating conditions under which AI use can scale safely: what requires governance, what can happen by default, where consequence requires stronger controls and which technical or human signals indicate that the risk posture has changed.

This is where Cybermaniacs' AI Enablement & Change (AIECM) approach connects with Human Risk Management. AIECM addresses workforce readiness, behavior, confidence, capability and adoption, while Human Risk Assessment & Assurance provides deeper evidence about whether the human controls relied upon by the organization are viable.

The CISO does not need to become the chief learning officer. They do need confidence that the human responsibilities embedded in the control framework are real.

What should a Secure AI Enablement program actually contain?

The exact design will vary by organization, but several capabilities need to connect rather than operate as separate initiatives.

The program needs governance that defines acceptable risk and decision rights; an inventory of meaningful AI use rather than merely licensed tools; technical controls around identity, data, access and agent authority; and risk assessment that accounts for the work and consequence surrounding each use case.

It also needs a workforce component capable of building more than generic awareness. Employees need practical guidance for the AI they use, higher-risk roles need stronger judgment and verification capability, and people supervising agents need explicit responsibility, escalation paths and intervention authority.

Finally, the program needs assurance. Controls should be tested, human behavior should be measured where it materially affects risk, agent permissions and workflows should be reviewed as they change, and lessons from incidents, near misses and everyday use should feed back into governance.

That is why we think of Secure AI Enablement as an operating cycle rather than a launch project.

Organizations govern how AI should be used, enable people to use it, measure what is actually happening and adjust the system as capability and risk evolve. The exact mechanisms will differ by environment, but the need for that cycle does not.

What does a mature Secure AI Enablement program look like?

It probably looks less dramatic than people expect.

Employees know where they can use AI and rarely need to work around policy. Business teams can move useful ideas into production without beginning governance from scratch each time, while higher-risk use cases receive more attention than trivial ones.

Humans understand the decisions they still own, and agents operate inside bounded authority with identifiable owners. Intervention occurs when consequence or uncertainty warrants it rather than because every action has been forced through a ceremonial approval step.

Security teams have enough technical telemetry to understand the environment, while leaders also have evidence about whether the human side of the control system is working. Policies change because the organization learns, training changes because the work changes, and controls become stronger or lighter as experience provides better evidence about the actual risk.

That is the opportunity in Secure AI Enablement.

The objective is not to slow AI down until it becomes safe. It is to build an organization capable of moving quickly because it understands where the risk lives, which controls matter and when human judgment still needs to carry weight.

TAGS: Metrics