FrameworkGovernance
The Board's AI Mandate
How boards of capital-intensive companies should oversee AI: risk appetite by consequence, the three lines model, the main frameworks and the questions to ask.
In brief
- ISet AI risk appetite by consequence and autonomy. A drafting tool and a system acting on physical operations need different rules.
- IIApply the three lines model to AI: business owners run systems, a second line sets standards and challenges, internal audit assures.
- IIINIST AI RMF, ISO/IEC 42001 and the EU AI Act are complementary. None substitutes for an inventory, named owners and controls.
AI governance is usually framed around data privacy, bias and reputation. Those risks matter in energy, oil and gas, real estate and infrastructure, but they are rarely the largest exposures. When AI influences how a pipeline is operated, how a grid is dispatched or how reserves are booked, the consequences are measured in safety incidents, environmental releases, outages and restatements.
Capital-intensive companies also start with an advantage that is easy to overlook. They already govern high-consequence decisions through process safety management, reliability standards and, in their trading operations, independent risk control. The board’s task is to extend disciplines the company already trusts rather than build a parallel bureaucracy for AI.
Set the appetite by consequence and autonomy
A board cannot oversee every model and should not try. It can define how much AI risk the company is willing to carry, and in which decisions. Two dimensions do most of the work, the consequence of an error and the autonomy the system is given, and together they produce four tiers:
- Productivity tools. Drafting, search and summarization, where a person reviews every output. Governed by acceptable-use policy and data controls.
- Decision support. Forecasts, rankings and recommendations that inform a human decision, such as a model that ranks pipe segments for replacement. Governed by validation before use and a named owner.
- Operational influence. Systems whose outputs change how physical assets are operated, inspected or dispatched, such as recommended compressor setpoints, even with a person in the loop. Governed as a change to the operation, through management of change and the safety program.
- Autonomous action. Systems, including language-model agents, that act on operations, markets or customers without a person approving each action. Permitted only within limits the board has explicitly approved.
The appetite statement then becomes concrete: which tiers are permitted in which parts of the business, what evidence moves a system up a tier and which deployments require committee approval. Oversight begins with an inventory, and an inventory begins with a definition. Management should be able to say what counts as an AI system, how many are running and at what tier.
The three lines, applied to AI
The Institute of Internal Auditors’ Three Lines Model translates well. The first line is the business: the operating, commercial and engineering leaders who deploy AI and own its performance and risk. The second line sets standards and monitors: risk management, compliance, process safety, cybersecurity and model validation. The third line, internal audit, gives the board independent assurance that the first two are working.
Two points of practice matter. Every system above the lowest tier needs a named first-line owner, an executive who answers for it, rather than the data science team that built it. The second line needs people who can challenge a model on its own terms, a lesson banks learned under SR 11-7, the 2011 U.S. supervisory guidance on model risk management, with its emphasis on independent validation and effective challenge. Operating companies can borrow that discipline without the banking bureaucracy.
Internal audit’s role changes too. Auditors need enough fluency to test whether the inventory is complete, whether validation happened before deployment and whether monitoring actually occurs. That usually means new skills, hired or contracted.
Oversight begins with an inventory, and an inventory begins with a definition.
The frameworks, and what each is for
Directors will hear about three frameworks and one regulator’s expectations. They are complementary, and none substitutes for an inventory, named owners and controls.
NIST AI Risk Management Framework
Voluntary guidance from the U.S. National Institute of Standards and Technology, organized around four functions: Govern, Map, Measure and Manage. It gives management a shared vocabulary for structuring a program, and NIST has published a companion profile for generative AI. It is not a certification.
ISO/IEC 42001
An international standard for an AI management system, built on the common structure ISO uses for management system standards, the same one behind ISO/IEC 27001 for information security. Organizations can be certified against it by an accredited third party, which helps when customers, lenders or partners want evidence rather than assertions.
EU AI Act
Binding law for organizations that supply or use AI in the European Union, with a risk-based structure: a short list of prohibited practices, obligations for providers of general-purpose AI models and the heaviest requirements for high-risk systems. That category includes AI used as a safety component in the management and operation of critical infrastructure, including the supply of electricity, gas, heating and water. Obligations phase in over several years and the high-risk timetable has been the subject of amendment proposals, so boards should ask counsel for the current dates. A company that builds or substantially modifies a system can also take on a provider’s obligations.
Securities disclosure
In the United States, the SEC has brought enforcement actions over misleading claims about AI, often called AI washing, and AI increasingly appears in risk factor disclosure. Statements about AI capability, use and risk are held to the same antifraud standards as any other disclosure. The SEC’s 2023 cybersecurity rules, which require registrants to describe board oversight of cyber risk, offer a reasonable template for how AI oversight may come to be described.
Committee structure and cadence
Most companies do not need a standalone AI committee. They need AI risk assigned to the committees that already own the consequences, with the full board owning strategy:
- Full board. AI strategy, capital allocation and the risk appetite statement, reviewed at least annually.
- Audit committee. Controls, disclosure, the inventory and internal audit’s findings.
- Safety, sustainability or operations committee. Tier three and four systems, reviewed with the same rigor as process safety performance.
- Compensation or talent committee. Workforce effects, reskilling and changes to roles that carry safety or fiduciary accountability.
Cadence matters as much as structure. A quarterly view should show the inventory by tier, systems that moved between tiers, incidents and near misses involving AI, validation exceptions and the status of regulatory obligations. Directors also need their own fluency, because a board that has never seen a model fail will struggle to judge management’s account of one.
Questions for your next leadership meeting
- Can management show us the AI inventory by tier, and who has attested that it is complete?
- Which systems can act on operations, markets or customers without a person approving each action, and within what limits?
- For each tier three or four system, who is the accountable executive, when was it last validated and what would tell us it had degraded?
- What have we said publicly about our use of AI, and can we substantiate every statement?
- Which AI obligations under the EU AI Act or other regimes apply to us in the next 24 months, and who owns readiness?