The boundary separating the frontier AI labs (e.g., Anthropic, OpenAI) from the enterprise institutions deploying their foundation models has dissolved. All operate within an interdependent ecosystem where technical failure modes, supply-chain dependencies, and unmanaged autonomous behaviors cascade across institutional lines.
As artificial intelligence moves rapidly from back-office functions to customer-facing interactions, and increasingly into kinetic military applications such as automated target recognition and command-and-control (C2), ad hoc safety measures are wholly inadequate.
For Chief Risk Officers (CROs), board directors, and executive leadership, AI risk management must be reframed. It is not a bureaucratic tax or a regulatory compliance exercise. Instead, rigorous AI risk governance represents a core engine of enterprise valuation, capital preservation, and competitive advantage.
Institutions that master the ability to identify, aggregate, and treat complex AI risks will capture disproportionate market share because enterprise capital and government buyers inevitably flow toward reliable, resilient systems. Building this resilience requires an integrated, eight-component Enterprise AI Risk Management Framework that enforces strict individual accountability, risk-based prioritization, and alignment with national economic interests. The figure below, Manage Enterprise AI Risks Across Time, provides an overview.
1. Identify Risk Objects and Create Ownership. Every AI risk management strategy must begin by mapping its physical, digital, and human assets, termed "risk objects," across the business value chain. In the AI ecosystem, risk objects extend far beyond software code to encompass data centers, specialized compute clusters, high-voltage electrical grid infrastructure, base model weights, synthetic training pipelines, API endpoints, and top-tier engineering talent. Each risk object should be assigned a singular business owner within the organizational hierarchy. The risk owner holds ultimate decision-making authority and operational responsibility for the asset throughout its lifecycle.
2. Building a Standardized AI Risk Taxonomy. A primary driver of corporate failure during technological transitions is imprecise language. AI currently lacks these standard definitions. Executive leadership, quantitative researchers, and risk managers frequently talk past one another because basic technical terms lack rigorous operational definitions. Enterprise Risk Management (ERM) frameworks require a common taxonomy that explicitly defines critical concepts. Commonly misunderstood AI terms include such basic things as:
3. Distinguishing Quantifiable Risk vs. Knightian Uncertainty. Whether for systemic or idiosyncratic risks, traditional ERM relies heavily on historical data to build probabilistic models. However, CROs must distinguish between quantifiable risk, where outcome probabilities can be mathematically estimated, and Knightian uncertainty, where the probability distribution of failure modes is fundamentally unknown. Work on the types of uncertainty in AI model proceeds.
CROs cannot govern frontier AI solely through statistical backward-looking models. They must deploy qualitative scenario planning, stress testing, and kill switches designed explicitly for uncertainty.
Brenda Boultwood
4. Risk Identification Methods. Acknowledging systemic and idiosyncratic risks in complex systems requires active discovery processes. To identify AI idiosyncratic, or firm-specific, risks, organizations must implement continuous issue logging across all model inference calls, automated red-teaming scripts and other scenario analysis, data audits, and threat intelligence monitoring. To address national security and critical infrastructure concerns, risk identification must aggressively evaluate idiosyncratic physical supply chain integrity, such as hardware fabrication bottlenecks and energy grid dependencies, to prevent sudden single-point failures. Examples of AI systemic risks can be found here.
5. Risk Assessment, Key Risk Indicators, and Extreme Red-Teaming. Risk assessment augments identified risks with probability and impact scores. Uncertainties can be quantified based on simulations, sophisticated scenario analysis and expert judgment. AI labs and enterprise deployers must establish extreme quantitative and qualitative benchmarking evaluations to test model limits before deployment. Extreme red-teaming, employing specialized adversarial teams and automated AI agents to aggressively probe models for severe failure modes, is mandatory. In defense settings involving automated targeting, red-teaming must explicitly evaluate model failure under degraded sensor environments, outdated training data, adversarial data poisoning, and unexpected target misclassification thresholds.
6. Establishing AI Risk Appetite Framework. Board-approved risk appetite statements will align with the risk taxonomy and define the boundary between acceptable and unacceptable risk-taking. Risk appetite thresholds must set non-negotiable quantitative metrics, such as:
7. Risk Treatments. While systemic AI risks and uncertainties must be treated at a federal level, once an idiosyncratic risk is assessed and is found to exceed agreed risk appetite limits, the assigned Risk Owner must execute a risk treatment decision. The Risk Owner is directly accountable for selecting one of four formal actions:
Mitigate – Take deliberate steps to reduce the risk level to acceptable levels, including new sandbox guardrails, Retrieval-augmented generation (RAG) architecture to improve model accuracy, or real-time monitoring to lower risk likelihood or impact. Risk mitigation plans may require significant investment and have impacts that drive new risks.
Avoid – Halt model development, delay release, or refrain from deploying AI in specific high-hazard domains.
Transfer – Reallocate financial or operational exposure via contractual indemnities, master service agreements (MSAs) with cloud vendors, or specialized insurance instruments. Third-party insurance options may one day exist.
Accept – Formally signing off on residual risk when risk levels exceed agreed risk appetite. These risk acceptances are typically time bound.
8. Risk Aggregation, Prioritization, and Business Strategy Integration. The use of a common risk taxonomy and rating scales ensure that risks can be aggregated across granular models, hardware, and other risk objects into aggregate corporate risk profiles. The highest-rated risks can then be reported, along with any action plans describing risk treatment plans. These top risks should be linked to business strategy and annual performance objectives to get a realistic view of the potential impacts to profitability, investments, and capital. Through risk-based decision-making, executive leadership can prioritize resources toward the most critical risks.
Consider a real-world strategic decision faced by executive leadership at a AI lab. The lab completes training on a $10B next-generation foundation model. During pre-deployment evaluations, internal red-teams discover that the model exhibits a 3% emergent probability of autonomously executing multi-stage cyber-exploitation routines. It’s a capability that could compromise commercial cloud networks and military C2 infrastructure of some customers.
The commercial team faces intense competitive pressure to launch immediately to capture market share, arguing that delaying release will cost millions in daily API revenues.
Without an Integrated ERM framework, executive leadership yields to short-term commercial pressure and releases the model. Within weeks, external actors exploit the model's unmitigated cyber capabilities to breach domestic critical energy infrastructure. The resulting regulatory clampdown, third-party liability litigation, and brand destruction wipe out billions in enterprise value and permanently destroy institutional trust.
With the 8-Component ERM Framework:
1. The Velocity Paradox: Slowing deployment to execute rigorous pacing, red-teaming, and risk treatments dramatically increases long-term commercial velocity by preventing business-halting operational catastrophes and regulatory shutdowns.
2. The Precision Paradox: Over-relying on precise quantitative risk models creates a false sense of security. CROs must govern using qualitative frameworks designed for Knightian uncertainty rather than treating complex AI systems like traditional linear software.
3. The Boundary Paradox: Governing software code and model weights in isolation is futile if the physical risk objects such as power grid stability, hardware supply chains, and third-party vendor concentration are left unmanaged.
Far from hindering technological progress, a comprehensive AI risk management framework will enhance credibility, consistency and profitability. In an era where AI models are integrated into commercial product engines, healthcare diagnostics, and national warfighting architectures, operational failure carries existential costs. Institutions that enforce clear accountability through risk ownership, establish consistent risk taxonomies, and prioritize and treat risks effectively will not only protect their balance sheets against catastrophic tail events. They will win the long-term commercial race for market dominance, reputation, and public trust.
Brenda Boultwood is the Distinguished Visiting Professor, Admiral Crowe Chair, in the Economics Department at the United States Naval Academy. The views expressed in this article are her own and should not be attributed to the United States Naval Academy or the U.S. Department of Defense.
She is the former Director of the Office of Risk Management at the International Monetary Fund. She has previously served as a board member at both the Committee of Chief Risk Officers (CCRO) and GARP, and is also the former senior vice president and chief risk officer at Constellation Energy. She held a variety of business, risk management, and compliance roles at JPMorgan Chase and Bank One.