Stop Asking If AI Is Trustworthy. Start Asking If Your System Is Governable.

Watch the Webinar

Watch Cheri and Diane go deeper on AI Admissibility, system boundaries, consequential actions, human oversight, and what practical AI governance looks like inside an organization.

Listen to the Podcast

Hear Cheri and Diane’s original conversation on The Art of Cybersecurity: AI Is Still Just a System: A Practical Approach to Governance.

The Short Answer

AI governance should not depend on making an AI model perfectly trustworthy. Generative AI is probabilistic rather than deterministic, and increasingly AI-enabled systems do more than generate answers. They retrieve information, invoke tools, call APIs, trigger workflows, modify records, and take actions capable of creating real business consequences.

The practical governance question therefore becomes: can the surrounding system control what the AI is allowed to do? Before consequential action occurs, the organization should be able to evaluate who or what is acting, what action is being attempted, what information is influencing that action, and what evidence will show why the decision was permitted.

Many of the underlying cybersecurity principles are not new. What has changed is the speed, autonomy, changing context, and scale at which AI-enabled systems can apply them.

Key Takeaways

• AI is still part of a system. Existing cybersecurity concepts such as least privilege, authorization, boundaries, logging, evidence, architecture, threat modeling, and risk management still matter.
• Do not confuse a polished interface with an engineered system. AI may look highly competent while the permissions, data, workflows, identities, and other dependencies behind it remain poorly governed.
• Governance has to reach the point of action. Policies, assessments, testing, and monitoring matter, but they do not by themselves determine whether a specific AI-enabled action should occur right now.
• Consequential AI actions need explicit boundaries. Organizations should define what the system is allowed to do instead of trying to predict every possible thing that could go wrong.
• Human-in-the-loop is not the entire answer. Human oversight should be designed according to risk rather than requiring a person to approve every automated action.
• AI governance should enable responsible adoption. The objective is not to stop AI. It is to capture its value without surrendering accountability, evidence, and control.

Why Is AI Governance Different From Governing Traditional Systems?

Traditional software is generally built around deterministic logic. Under the same conditions, organizations can reasonably expect the same programmed logic to execute and produce the same result.

Generative AI behaves differently. It produces responses based on patterns, statistical inference, training data, and changing context. That flexibility is one of the reasons AI is so powerful, but it also means the model can produce different outputs or choose different paths depending on the circumstances around it.

“Don’t start by asking whether the model is trustworthy. Start by asking whether the system is governable.”

— Diane R Jones

That distinction becomes more important as AI moves from helping a person create content to making or proposing decisions inside systems that can actually act.

Why Can AI Look More Trustworthy Than It Really Is?

One reason organizations may move too quickly is that AI presents itself through an unusually convincing interface.

“Leaders see the interface, not the system.”

— Diane R Jones

A user enters a natural-language prompt and receives a polished answer within seconds. That experience can feel like competence. But underneath that interface are dependencies involving data quality, permissions, identities, retrieval sources, workflow design, tools, integrations, and system architecture.

AI is not magic. It is technology operating inside a system, and reliable use still requires engineering and governance.

What Changes When AI Can Take Consequential Action?

There is a meaningful difference between using AI to polish an email and allowing an AI-enabled system to send an email, change a record, invoke another tool, call an API, approve a payment, change a configuration, or trigger another workflow.

The output is no longer merely information. It can become consequence.

That is also where traditional governance silos can create gaps. An AI-enabled workflow may move from a model output to an identity, obtain a token, retrieve information from another tool, invoke an API, and initiate another action. Each individual component may be governed while the transitions between them are not clearly owned.

This is why Hotman Group approaches AI governance as part of the organization’s broader cybersecurity, risk, technology, governance, and business environment instead of treating AI as another isolated compliance program.

What Is AI Admissibility?

Diane’s AI Admissibility approach focuses on a practical question:

The approach evaluates four elements at the decision point:

Together, those conditions allow an admissibility boundary to determine whether an action should be permitted, denied, deferred, or escalated.

Why Is “Deny by Default” Useful for AI Systems?

Trying to predict every strange or harmful thing an AI-enabled system might do is nearly impossible.

A more manageable approach is to define what the system is supposed to do and constrain it to those approved paths. If an action falls outside the defined scope, the default response should not be silent approval.

That boundary can expand over time as the organization learns where additional AI capability creates value without introducing unacceptable risk.

What Does AI Admissibility Look Like in a Business Process?

The webinar uses a financial workflow to make the concept practical.

Imagine an AI-enabled finance assistant that is authorized to process routine payments of $500 or less. The system can confirm that the proposed payment falls within that permitted action, rely on approved internal financial information as context, and preserve evidence showing why the transaction was allowed.

Now imagine the same system proposes changing an employee’s compensation or sending money to new banking information.

The system may technically have access to financial and HR information, but that does not mean the proposed action belongs inside its authorized purpose. A payroll change may need to be denied or escalated. New banking information may require additional validation before the system relies on it.

The important point is that access alone does not equal authority, and capability alone does not mean an action should be permitted.

Why Does Evidence Need to Start Before the Action?

Traditional logging and after-the-fact monitoring remain important. But if evidence begins only after an AI-enabled action has occurred, the organization may simply have a record that something happened without enough information to understand why.

Evidence should begin at the decision boundary, where authority, action, and context are evaluated.

That creates the breadcrumbs needed to understand why the system allowed, denied, deferred, or escalated an action — whether an incident occurred or the organization simply wants to improve the system later.

Why Can AI Fail Without Looking Like a Failure?

Traditional systems often fail loudly. A service crashes, a backup job fails, or an expected process stops.

AI can behave differently. A model may continue confidently along what appears to it to be a plausible path even though that path is moving away from the organization’s intended outcome.

Continuing along a plausible path is not necessarily an error condition for AI.

That means drift may not immediately produce a clear failure signal. Changing context can then influence subsequent decisions, creating additional context and allowing the system to move farther from the original intent.

Organizations therefore need both ongoing monitoring and enforceable system boundaries designed around what the system is actually intended to do.

Does Human-in-the-Loop Solve AI Governance?

Human review is valuable, particularly where an action could create significant financial, legal, operational, security, customer, or human impact.

But a human cannot realistically approve every action taken by systems operating at AI speed.

“The human in the loop really needs to be at the top of the loop.”

— Diane R Jones

Humans should design what the system is expected to do, define its boundaries, establish escalation paths, determine where human review is mandatory, and continuously reassess the system as risk changes.

That is consistent with a risk-based cybersecurity approach: stronger controls and more human attention should be applied where potential consequence is greater.

Is AI Governance Different for Generative AI and Agentic AI?

The risk becomes more obvious with agentic AI because an agentic system can take model reasoning and connect it to tools, workflows, APIs, records, and other action capabilities.

But the core governance principle remains the same.

Agentic AI simply raises the stakes because autonomous or semi-autonomous actions can occur more quickly and with less direct human involvement.

Why Isn’t All Governance Friction Bad?

Threat modeling, context validation, runtime checks, evidence capture, escalation, and human review all introduce some overhead.

That overhead can look like friction, especially when the business is under pressure to move quickly.

But not all friction is waste.

Some friction creates evidence. Some prevents unauthorized action. Some forces validation of questionable context. Some ensures that a higher-risk decision reaches a person with the authority and judgment to make it.

The goal is not maximum control or zero friction. It is appropriate friction for the risk and consequence involved.

How Can Security Teams Enable AI Instead of Becoming the Department of No?

Organizations are under real pressure to adopt AI. Competitors are using it, employees are using it, third-party products increasingly contain it, and leadership wants the productivity and business benefits.

The answer cannot simply be “no.” It also cannot be “deploy it now and govern it later.”

Cybersecurity and GRC leaders need to help business leaders understand where AI can be enabled safely, where risk becomes meaningful, where stronger controls are necessary, and why.

A strong Cyber GRC operating model helps connect those decisions to ownership, architecture, risk, controls, evidence, escalation, and executive accountability instead of allowing AI governance to become another disconnected silo.

How Can Responsible AI Give People More Capacity?

The opportunity behind this work is substantial.

Cybersecurity and GRC teams are often overloaded with repetitive operational work. AI can help collect information, perform repeatable tasks, surface issues, and handle well-defined low-risk activity.

That can give skilled practitioners more capacity for the things humans are particularly valuable for: thinking, designing, evaluating exceptions, solving problems, communicating, exercising judgment, and continuously improving the system.

The goal is not to remove human responsibility. It is to use human attention where judgment creates the most value.

What Should Organizations Do Now?

1. Start where failure matters most. Choose a consequential business system rather than trying to govern every AI use case at once.
2. Inventory the AI already present. Determine where AI is embedded, what it is being asked to do, and which systems, data, identities, and tools it can reach.
3. Define system intent. Be explicit about what the system should be able to do and what falls outside its purpose.
4. Define authority and action boundaries. Identify who or what may act, which actions are permitted, and when an action must be denied, deferred, or escalated.
5. Evaluate context. Understand which data and external information can influence decisions and whether that context is trustworthy enough for the consequence involved.
6. Preserve evidence at the decision point. Be able to reconstruct what was proposed, what authority existed, what context was used, and why the action was allowed or rejected.
7. Apply human oversight according to risk. Keep humans focused on design, exceptions, high-consequence decisions, and continuous improvement rather than approving every routine transaction.
8. Build iteratively. Start with constrained, well-understood benefits and expand AI capability deliberately as the organization gains evidence that the system is operating as intended.

Frequently Asked Questions

Can an AI Model Be Made Completely Trustworthy?

Organizations should not design consequential systems around the assumption that an AI model will always behave predictably. AI is probabilistic and influenced by changing context. The surrounding system should therefore include enforceable controls governing what the AI is permitted to do.

What Is AI Admissibility?

AI Admissibility asks what must be true before an AI-enabled system is allowed to create a consequence. It evaluates factors such as authority, the proposed action, the context influencing that action, and the evidence needed to explain the decision.

What Is an AI Admissibility Boundary?

An admissibility boundary is the decision point where the system evaluates whether a proposed AI-enabled action should be permitted, denied, deferred, or escalated based on defined governance conditions.

Does Every AI Action Need Human Approval?

No. Human oversight should be proportionate to risk. Routine, low-risk activity inside a well-designed and constrained system may operate automatically. Higher-consequence actions may require validation, escalation, or direct human approval.

Is Agentic AI Harder to Govern Than Generative AI?

Agentic AI can create greater operational risk because model reasoning is connected to systems capable of taking action. The underlying governance principle remains the same: the organization needs to control what the surrounding system is allowed to do.

How Should Companies Start Governing AI?

Start with a consequential business system. Identify what AI capability already exists, what the system is allowed to do, what could create meaningful harm, and where authority, context, evidence, boundaries, and human oversight need to be strengthened.

Should AI Governance Be a Separate Compliance Program?

Not automatically. AI creates new risks, but organizations already have capabilities in cybersecurity, GRC, privacy, architecture, third-party risk, access management, change management, policy, and risk management. Effective AI governance should reuse those capabilities where they work and add new controls where the technology creates genuinely new requirements.

About This Conversation

This article brings together a June 12 podcast conversation and a subsequent July 1 live webinar between Cheri Hotman, Managing Partner of Hotman Group, and Diane R Jones, CISSP, CCSP on practical AI governance and the AI Admissibility approach.

Diane’s background spans software development, technical project management, cybersecurity, GRC, and AI governance. Her work on AI Admissibility applies established security-engineering principles to systems in which AI can reason, retrieve changing context, interact with tools, and create consequential action.

Together, Cheri and Diane explore how organizations can move from static, policy-centered AI governance toward practical system controls involving risk, architecture, authority, action boundaries, context, evidence, human oversight, monitoring, and continuous improvement.

About Hotman Group

Hotman Group is a cybersecurity and Cyber GRC professional-services firm that helps organizations diagnose, design, build, remediate, implement, operate and mature cybersecurity programs. HG provides hands-on vCISO and vGRC leadership, supports multi-framework environments, and helps organizations select and implement GRC technology while connecting cybersecurity decisions to business risk and strategy.

Hotman Group also helps organizations integrate practical AI governance into existing cybersecurity, risk, technology, third-party risk, policy, and governance structures instead of automatically creating another disconnected compliance program.

The objective is not to slow responsible AI adoption. It is to help organizations understand where AI creates meaningful risk, establish accountability and control, and use emerging technology in ways the business can explain, operate, and defend.

Talk With Hotman Group