Why AI Risk Questions Now Require Technical Specificity
Shareholder and investor scrutiny of AI risk has increased alongside regulatory activity. NIST published the AI Risk Management Framework (AI RMF 1.0) in January 2023, followed by a Generative AI Profile in July 2024 addressing risks specific to generative systems. ISO/IEC 42001, published in December 2023, established the first international standard for AI management systems. Separately, the SEC has taken enforcement action against companies for overstating AI use, charging two investment advisers in March 2024 for misleading statements about AI capabilities, an issue often referred to as AI washing. Together these developments signal that generic reassurance is no longer sufficient. Investors and regulators expect governance leaders to identify specific risk categories, reference a recognized framework, and support claims with evidence that internal audit or compliance functions can independently verify.
Categorizing AI Risk Before Responding
A credible answer starts with correctly categorizing the risk being asked about. Governance risk concerns whether policies, accountability structures, and risk tolerance definitions exist, which maps to the Govern function of the NIST AI RMF. Data exposure risk concerns how AI systems access, process, and retain information, particularly sensitive or regulated data. Model behavior risk, addressed in NIST's Generative AI Profile, covers content integrity, reliability, and dependency on third-party or open foundation models. Agentic runtime risk concerns how autonomous AI agents act during operation, including what systems and data they can reach and whether that activity is observable after the fact. Naming the correct category before answering prevents a common failure mode: giving a governance-level answer to a runtime-level question, or vice versa.
| Risk Category | What It Covers |
|---|---|
| Governance Risk | Whether accountability, policy, and oversight structures exist for AI use. |
| Data Exposure Risk | How AI systems access, process, and retain sensitive or regulated data. |
| Model Behavior Risk | Reliability, content integrity, and third-party model dependency. |
| Agentic Runtime Risk | How autonomous AI agents are permissioned, monitored, and constrained in operation. |
Distinguishing Implemented Controls from Planned Measures
SEC enforcement precedent has specifically targeted statements that overstated the current state of AI use or oversight, not merely the existence of AI-related disclosures. This makes the distinction between implemented and planned controls a material communication decision, not a stylistic one. Before any earnings call, proxy statement, or ESG disclosure, governance leaders should confirm which controls are currently operating and produce verifiable evidence, such as audit logs or policy enforcement records, versus which remain in design or pilot stages. Terms like agent identity, least-privilege permissioning, and runtime policy enforcement are not yet codified in NIST, ISO, or SEC guidance; they are enterprise governance practices that support the Manage function of the NIST AI RMF and should be presented as such, rather than as regulatory requirements the company has satisfied.
Building a Defensible Response Process
The following questions and answers reflect common shareholder inquiries and illustrate how to apply this framework in practice.
What AI risk categories should we be prepared to discuss with investors?
Governance risk, data exposure risk, model behavior risk, and agentic runtime risk are the categories most commonly raised. Each should be assessed against a recognized framework such as NIST AI RMF or ISO/IEC 42001 before any external statement is made.
What evidence supports a claim about AI governance controls?
Audit logs, policy enforcement records, and documented risk assessments provide the evidentiary basis for any statement about oversight. Statements without traceable internal evidence carry disclosure risk under SEC anti-fraud provisions.
How do we avoid AI washing when answering investor questions?
Limit statements to controls that are currently implemented and independently verifiable. Distinguish clearly between operating controls and planned initiatives, consistent with SEC enforcement precedent against overstated AI claims.
Should legal counsel review AI risk disclosures?
Yes. Legal or compliance counsel should review AI-related statements for consistency with SEC anti-fraud provisions and AI washing enforcement precedent before earnings calls or proxy filings.