AI Governance Is an Operating Model, Not a Policy
The challenge is not simply deciding which AI tools employees can use. It is establishing how AI should be used, who remains accountable, and how organizations know it is working.
Ravinder Tulsiani, DBA · Enterprise Capability Executive
Executive abstract
An organization approves an AI assistant to help employees draft documents. The security review is complete, the tool is authorized, and employees receive guidance on acceptable use. Over time, the same tool starts being used for more consequential work. Employees ask it to interpret policies, summarize sensitive information, evaluate options, and recommend decisions. The technology may be the same. The organizational risk is not.
The central argument
AI governance becomes meaningful when an organization can translate its principles into operational decisions, responsibilities, controls, and evidence. That makes it an enterprise leadership responsibility involving technology, legal, privacy, risk, operations, human resources, and workforce capability.
Key findings
Start with the work, not the technology
The first governance decision should not be which AI platform to approve. It should be what organizational problem AI is expected to solve, whether AI is appropriate, and what consequences its use could create.
Accountability must exist before the incident
Every material AI use case needs an accountable business owner who understands the intended outcome and accepts responsibility for how the technology is incorporated into operations. Accountability must exist before the incident, not be negotiated afterward.
Match scrutiny to risk
A better approach is to classify use cases based on the risks they introduce. I saw this in the education technology portfolio I chaired. Requests to try a new AI tool had built into a backlog of about three months. We replaced a single approval route with tiered assessment by risk. Low-risk requests were those used to develop assets with no company-sensitive data exposed. Articulate 360 and Adobe AI features fell in that group, and cleared in days. Requests that involved sensitive data went for deeper review. Salesforce fell in that group in this case, because it involved patient data. IT and legal led that review to protect sensitive data.
Human review must work in practice
Can the person inspect the evidence supporting the recommendation? Do they have sufficient expertise to identify a mistake? Are they given enough time to review the output? Can they reject the recommendation without being pressured to accept it? If the answer to those questions is no, human review may exist on paper without providing meaningful oversight.
Measure performance and control effectiveness together
Each significant AI use case should have an agreed performance baseline and measures of both value and control effectiveness. Management also needs to distinguish capacity released from actual financial savings. Time saved does not automatically translate into reduced expenditure.
What the evidence supports
- NIST AI Risk Management Framework and Playbook: voluntary guidance for using the AI Risk Management Framework (AI RMF 1.0, released January 26, 2023; four functions Govern, Map, Measure, Manage).
- ISO/IEC 42001:2023, Artificial Intelligence Management Systems: international standard published December 2023, specifying requirements for establishing, implementing, maintaining and improving an Artificial Intelligence Management System.
- Government of Canada, Guide on Departmental AI Responsibilities: published January 2026. Guidance on multidisciplinary governance and business value.
- Government of Canada, Guide on the Use of Agentic Artificial Intelligence: federal guidance on agentic AI systems (Government of Canada Publications catalogue record, published May 18, 2026).
What this paper adds
Concepts, frameworks, models and decision principles developed by Ravinder — practitioner synthesis, distinct from the external evidence cited above.
- AI governance as an operating model, not a policy: use-case visibility, decision rights established before an incident, controls tested in the actual workflow, and measurement of performance and control effectiveness together.
- Tiered assessment by risk in the education technology portfolio: replace a single approval route with scrutiny matched to the risk, so routine work can proceed and consequential uses receive proper attention.
- Approval is not permanent. Governance must establish when a use case requires reassessment, what evidence triggers intervention, and who has the authority to suspend or retire it.
Implications for leaders
- The real test of AI governance is not whether an organization has published a policy or established a committee. It is whether leaders can explain what AI is permitted to do, who remains responsible for the consequences, how the controls work in practice, and what evidence justifies continued use.
- As AI takes over parts of established roles, organizations must consider which activities develop the expertise employees will still need to exercise judgment and oversight.
Recommended actions
- First, establish visibility. Identify AI uses already underway, including approved systems, informal experimentation, and AI functionality embedded in existing applications. Determine which uses involve consequential decisions or sensitive information.
- Second, clarify decision rights. Establish who owns each selected use case, who can authorize its operation, which specialists must review it, and who can suspend or restrict its use.
- Third, examine controls in the actual work. Follow a small number of workflows from beginning to end. Test whether the stated protections, human review responsibilities, and escalation mechanisms function as intended.
- Fourth, agree on performance evidence. Identify the baseline, expected improvement, potential harms, and monitoring requirements before making decisions about wider deployment. At the end of 30 days, leadership should have a named owner for each selected use case and a written record of the control gaps found.
Evidence boundaries and limitations
The proposed implementation sequence and executive recommendations are the author's synthesis, not prescribed steps or universal regulatory requirements.
- The proposed implementation sequence and executive recommendations are the author's synthesis, not prescribed steps or universal regulatory requirements.
- The Canadian government documents below are public-sector guidance, not universal requirements for private-sector organizations.
- The tiered-assessment example is from the education technology portfolio the author chaired, not an enterprise-wide program.
The complete strategy paper is available as a PDF — the full exhibit set, worksheets and source notes included.