Skip to content
    Back to Insights
    AI Governance

    AI Governance Is Not Just a Policy Document

    Jaouhar Mosbahi10 January 20268 min read

    Every week, another organisation announces their AI Governance Framework, typically a polished document outlining principles like fairness, transparency, accountability, and human oversight. These principles are necessary, but they are wildly insufficient. The gap between an AI policy document and operational AI governance is vast, and it is precisely in this gap where real risks materialise. Organisations that believe publishing an AI ethics policy constitutes governance are making the same mistake that plagued information security two decades ago: confusing intent with implementation.

    The problem is familiar to anyone who has worked in information security governance. Policies without controls are aspirations, not governance. You would not consider your information security programme complete with just an acceptable use policy and no technical controls, monitoring, or incident response capability. AI governance deserves the same rigour. Yet the majority of organisations we assess have an AI policy, perhaps an AI ethics committee that meets quarterly, and little else. The models are deployed, the data flows, and the risks accumulate, all while the governance framework exists primarily as a PDF on the intranet.

    Operational AI governance starts with something deceptively simple: an inventory. You cannot govern what you do not know exists. Every AI model, every machine learning pipeline, every third-party AI service, every large language model API integration, and every automated decision system your organisation uses needs to be catalogued. This includes shadow AI, the tools your teams are using without IT or security awareness. In every engagement where we have conducted an AI inventory exercise, the organisation has been surprised by the extent of AI usage already embedded in their operations. Marketing teams using AI content generators, HR using AI screening tools, development teams integrating AI coding assistants, and customer support teams deploying chatbots are just the beginning.

    Risk classification follows the inventory. Not all AI applications carry the same risk. A recommendation engine for internal blog posts is fundamentally different from a model that influences credit decisions, medical diagnoses, or hiring outcomes. The EU AI Act codifies this intuition into law, establishing a risk-based classification system with four tiers: unacceptable risk (banned), high risk (subject to strict requirements), limited risk (transparency obligations), and minimal risk (largely unregulated). Whether or not your organisation falls directly under the EU AI Act's jurisdiction, this tiered approach provides a sound basis for proportionate governance. Higher-risk AI applications warrant more rigorous controls, more frequent monitoring, and more senior oversight.

    The model risk lifecycle is where governance becomes operational. From development through deployment to monitoring and eventual retirement, each stage needs defined controls and clear accountabilities. During development, governance should address training data quality and provenance, bias testing methodology, model validation against defined performance criteria, documentation of design decisions and limitations, and security testing including adversarial robustness. These are not theoretical requirements. They are practical controls that prevent the deployment of models that are biased, unreliable, or vulnerable to manipulation.

    Deployment governance is the gate between development and production. No model should reach production without a formal approval process that considers the results of testing and validation, an assessment of the model's potential impact on individuals and the organisation, confirmation that monitoring and alerting are in place, documentation of the model's intended use, known limitations, and failure modes, and assignment of ongoing ownership and accountability. This is analogous to change management in IT operations, and it serves the same purpose: ensuring that changes to production systems are deliberate, assessed, and reversible.

    AI impact assessments are a critical governance tool that most organisations have not yet adopted. Similar in concept to Data Protection Impact Assessments under GDPR, an AI impact assessment evaluates the potential effects of an AI system on individuals, groups, and the organisation. It should consider accuracy and reliability across different populations, potential for discriminatory outcomes, transparency and explainability to affected individuals, human oversight mechanisms, data protection implications, and environmental impact of model training and inference. The EU AI Act will require conformity assessments for high-risk AI systems, and organisations that build impact assessment capability now will be well-positioned when regulatory requirements crystallise.

    Monitoring for drift and bias is the ongoing operational challenge that separates genuine governance from paper governance. Models degrade over time. The data distribution in production diverges from training data, user behaviour changes, the world changes. A model that was fair and accurate at deployment can become biased or unreliable months later without any code changes. Governance must include defined metrics for model performance monitored continuously, statistical tests for bias across protected characteristics run at regular intervals, drift detection that alerts when input data distributions shift significantly, escalation paths when monitoring identifies degradation, and clear criteria for model retraining or retirement.

    The AI risk register is the governance tool that ties everything together. Just as an information security risk register catalogues threats, vulnerabilities, and their treatment, an AI risk register should catalogue the risks associated with each AI system: model bias, data quality issues, security vulnerabilities, regulatory non-compliance, reputational risk, and operational dependency. Each risk should have an assessed likelihood and impact, a defined treatment (mitigate, accept, transfer, or avoid), an assigned owner, and a review schedule. The AI risk register should integrate with, or be a sub-register of, your overall enterprise risk register to ensure visibility at the appropriate level.

    Accountability structures matter more than technology. Who owns the risk of a specific model? Who approves deployment? Who monitors performance? Who has the authority to pull a model from production when it misbehaves? These questions need clear, documented answers with named individuals, not just departments. The most common governance failure we see is diffused accountability: data science builds the model, engineering deploys it, operations runs it, and nobody owns the risk. Establishing clear model ownership, with an accountable individual who is responsible for the model's performance, compliance, and risk throughout its lifecycle, is a foundational governance requirement.

    ISO 42001, the international standard for AI Management Systems published in December 2023, provides a structured framework for operationalising AI governance. Built on the same Annex SL structure as ISO 27001, it is designed to integrate with existing management systems. For organisations that already have an ISMS, ISO 42001 offers a natural extension that applies familiar concepts like risk assessment, control implementation, internal audit, and management review to the specific challenges of AI. We expect ISO 42001 certification to become a meaningful differentiator for organisations deploying AI in regulated or high-trust environments.

    The EU AI Act, which entered into force in August 2024 with phased implementation through 2027, creates legally binding obligations for AI providers and deployers. High-risk AI systems must meet requirements for data governance, documentation, transparency, human oversight, accuracy, robustness, and cybersecurity. Organisations that treat these requirements as a future compliance exercise rather than a current governance priority will find themselves scrambling when enforcement begins. The regulatory trajectory is clear: AI governance is moving from voluntary to mandatory, and the organisations that build operational governance now will have a significant advantage.

    Practical steps to move from paper governance to operational governance follow a clear progression. First, conduct a comprehensive AI inventory across the organisation, including shadow AI. Second, classify each AI application by risk level using a methodology aligned with the EU AI Act or ISO 42001. Third, implement lifecycle controls for high-risk AI systems, including development standards, deployment gates, and monitoring requirements. Fourth, establish an AI risk register and integrate it with your enterprise risk management. Fifth, build an AI impact assessment process for new deployments. Sixth, define clear accountability structures with named model owners. Seventh, implement monitoring for drift, bias, and performance degradation. Eighth, establish an incident response process for AI-specific incidents.

    The cultural dimension of AI governance should not be underestimated. Governance frameworks that are perceived as bureaucratic obstacles to innovation will be circumvented. The governance programme needs to demonstrate value to the teams building and deploying AI, not just to the compliance function. This means involving data science and engineering teams in governance design, streamlining approval processes so they are thorough but not slow, providing clear guidelines that help teams make better decisions rather than just constraining them, and sharing examples of how governance has caught issues before they reached production.

    AI governance is not a destination. It is an evolving practice that must keep pace with rapidly advancing technology and an evolving regulatory landscape. The organisations that succeed will be those that treat AI governance as an operational discipline with the same maturity, resourcing, and executive attention as information security governance. The policy document is a starting point, not an end state. The real work, and the real value, lies in operationalising that policy into controls, processes, metrics, and accountability structures that govern AI throughout its lifecycle.

    AI GovernanceEU AI ActISO 42001Responsible AI