AI governance must extend beyond approval into daily operations

AI governance needs to work beyond the approval stage

Enterprises already have AI policies and protocols, yet organizations continue to bypass them, widening the gap between rules and execution. Prominent, well-documented cybersecurity incidents show the risks linked to AI, but awareness alone has failed to close that gap. CIOs therefore need controls teams can use during development, deployment and operation, with clear authority to act when risks change.

Those controls can also help teams deploy faster. “Moving fast and applying appropriate governance are not mutually exclusive,” Jackson said. To make both possible, CIOs must turn governance requirements into decisions people can carry out throughout a system’s working life, while avoiding unnecessary obstacles to operations.

Ownership and expertise determine whether controls can work

Continuing oversight becomes especially important with agentic AI: systems or agents that can act autonomously. In a survey, around one in four respondents using this technology said responsibility for maintaining or monitoring it after deployment was undefined. Approval lets a team launch, but ongoing ownership establishes who must look after the running system. Without an owner, no one knows who must act when it needs attention.

Assigning an owner also means defining what that person or team must do. Autonomous activity needs observation, responsibility for changes in behavior and clear conditions for human intervention. These duties connect, but assigning someone to watch a system does not automatically give them responsibility or authority to respond. Teams need to know how an observation leads to a decision and who carries it out.

Doing that work also requires expertise. Around two-thirds of survey respondents expressed concern about insufficient internal expertise to design, implement or evolve AI governance controls. This finding concerns respondents generally, while the ownership finding concerns those using agentic AI. Together, they raise two separate questions about readiness: whether responsibility has been assigned and whether the people assigned can do the work.

Expertise matters throughout that work because each stage calls for a different judgment. Designing a control means deciding what it should require; implementing it means making the requirement usable; evolving it means adapting the control as deployments develop. A monitoring owner may still need help judging whether monitoring is adequate. So governance planning must address both the assignment of duties and the skills needed to fulfil them.

Build approval, exceptions and intervention into deployment

To make those duties workable, Jackson recommends building governance into AI development and deployment. Teams should define testing, approval, monitoring and escalation requirements as part of the work, including how a problem reaches someone authorized to act. Testing establishes what must be examined, and approval identifies who decides whether deployment may proceed. Monitoring and escalation carry those requirements into operation by connecting observed problems to a response.

Urgent deployments need equally clear decisions and responsibilities. Jackson recommends that CIOs specify who can authorize an expedited deployment, what additional controls apply and which reviews must follow once the system is running. An exception then becomes an authorized decision with explicit conditions. Those conditions preserve responsibility for the risk accepted and the follow-up required, while giving teams a permitted route to move quickly.

Responsibility must continue after a deployment or exception has been approved. Jackson recommends defining ownership throughout the system’s lifecycle, including who monitors autonomous systems, who takes responsibility when behavior changes and when human intervention becomes necessary. Teams should be able to trace an observed change to the person responsible for responding. This makes the handoff between monitoring and intervention an explicit operating decision.

For that handoff to work, the people responsible need to know what an agent is doing. Jackson calls for human safeguards, technology-based safeguards, visibility into agent activity and the ability to detect operation outside intended parameters. Visibility supplies the information, while defined limits help people judge whether action is needed. Intervention rules then connect that judgment to someone who can act.

Use assurance reviews to change operating decisions

Once those controls are in place, Jackson recommends regular assurance reviews: recurring checks of whether the arrangements work in practice. Reviews should examine the working controls and inform decisions to modify a system, pause it or stop it. Their findings should also guide changes to governance as AI use develops, so requirements remain relevant to the systems teams are running.

Acting on those findings takes both authority and capability. “Governance is what gives organizations the confidence to move quickly without losing control,” Jackson said. For a review to support that confidence, its findings must reach a responsible decision-maker and people able to modify, suspend or shut down the running system.

CIOs must connect responsibilities across teams

The route to action often crosses organizational boundaries, so leaders must coordinate the teams involved. “The role of a CIO and leadership is to establish cross-functional teams, bringing together legal, IT, security and compliance teams, and make sure governance across each team is keeping pace with AI deployment,” Jackson said. As deployments develop, those teams need to coordinate so a decision in one function leads to the necessary action in another.

Within that shared work, Jackson places particular weight on the CIO’s technical judgment. “The insight and expertise of the CIO has never been more critical to helping inform and enable governance across the organization,” Jackson said. The CIO can help connect approval, exceptions, monitoring and intervention to the technology being deployed and the people overseeing it. This work extends beyond the CIO’s own function because legal, security and compliance decisions also affect how a system operates.

Yet bringing those functions together still leaves the task of matching duties to capability. The concern about expertise matters when teams assign testing, monitoring and review work, because each assignment depends on people who can make the required judgments. An organization already bypassing protocols may also fail to carry out newly documented responsibilities. Leadership therefore needs to check whether assigned teams can do the work and whether their findings reach someone with authority to act.

Key takeaways for decision-makers

• Extend governance into operations: CIOs need controls that teams can use during development, deployment and daily operation, with clear authority to act when risks change.
• Match ownership with expertise: Around one in four respondents using agentic AI reported undefined post-deployment responsibility. Assign owners for monitoring and intervention, and verify they have the skills and authority to carry out those duties.
• Define deployment and exception rules: CIOs need explicit testing, approval, monitoring and escalation requirements. For expedited deployments, specify who authorizes the exception, which additional controls apply and when follow-up reviews occur.
• Connect assurance to action: Regular reviews need to inform decisions to modify, pause or stop running systems. Route findings to decision-makers with authority and people capable of carrying out the response.
• Coordinate responsibilities across functions: CIOs need legal, IT, security and compliance teams to connect their decisions as AI deployments evolve. Check that each handoff reaches someone equipped and authorized to act.