Governance
“Can we really trust what we're given from management?” A sitting board member asked us that question this autumn. It is the right question, and for AI it is harder to answer than for most topics.
Key takeaways
Board information on AI usually comes from the people who are building or buying it. That is not a criticism of management. It means the information is often written to show progress, and progress is easier to measure than risk. Counts of pilots, use cases and productivity gains are common in board packs. Evidence about how models perform once deployed, which incidents occurred, and whether controls caught them is much rarer.
AI also changes after approval. A model can drift, a vendor can update the underlying system, and a tool approved for one use can be extended to another. Information that was accurate at approval can stop being accurate without anyone deciding to change it.
Directors already compensate for this. In Director, Defined research, only 14% of directors said they begin with management and board materials when they face an unfamiliar issue. 37% research it themselves and 28% ask their network. Before a decision, 46% said they value evidence most, against 10% who value debate.
Independent research is valuable, but it does not replace reliable board information. A director's own reading tells them what is happening across the market. Only the company's own information can tell them what is happening inside it.
The structure most boards already use for other material risks applies to AI. The Institute of Internal Auditors' Three Lines Model separates management, who own and run the risk; specialist risk and compliance functions, who set standards and monitor; and internal audit, who provide independent assurance to the board. The NIST AI Risk Management Framework and ISO/IEC 42001, the international standard for AI management systems, give those functions a common reference for what good looks like.
For the board, the practical shift is to ask where each AI report sits in that structure. A management update on a deployment is first-line information. A risk function's view of whether testing was adequate is second-line. An internal audit review of AI controls is third-line. Most boards receive only the first.
These four questions help a board tell whether what it receives is decision-useful:
In the Vela Board Navigator℠, this is the element called Board information and assurance: the board receives decision-useful evidence on adoption, outcomes, incidents and control effectiveness. It is one of the elements boards most often rate themselves highly on and then struggle to evidence.
Questioning management information is a skill, and it improves with practice. Several of our cases are built around exactly this problem. In The Second Opinion, the board is presented with an AI-generated analysis and has to decide how much weight it can carry. In The Green Light, management recommends accelerating a rollout while independent testing has open concerns. You can work through either in three minutes in Rehearse the Moment.
We are also publishing what directors tell us. The Vela Barometer reports how directors approach consequential decisions, including where they look for information and what they trust, so boards can compare their own habits with their peers'.
Adoption and use, outcomes against the intended benefit, incidents and near-misses, changes since approval, and the results of independent testing of controls. Activity measures alone are not enough.
Where AI is material to the business, yes. Internal audit provides independent assurance to the board, and AI controls are within the scope of that work.
As often as the materiality of AI to the business requires. Many boards cover AI annually as a strategy topic, but material deployments need regular reporting through the relevant committee.
The Navigator is where every Vela engagement begins. Request access or a briefing to see what it surfaces for your board.