transition2.aiStart a project

Enterprise AI Governance

What Boards Need to Hear About AI Programs

A one-page AI report a board can act on: what is in use, what it can do, what went wrong, what is changing in regulation, and which decisions need the board.

Eric MacDougall ·

Answer first

Boards do not need a tour of model architectures. They need to know what AI the company relies on, what those systems are allowed to do, what has gone wrong, what external obligations are coming, and which decisions are theirs. Put that on one page, use the same structure every quarter, and attach detail for anyone who wants it.

Why AI reporting often misses

Management reports tend to fall into one of two traps. The first is a showcase: demos, pilot counts and adoption charts with no view of risk. The second is a compliance dump: long control matrices that nobody on the board can turn into a decision. Both leave directors unable to answer the basic accountability question that frameworks such as the OECD AI Principles put at the centre: who is accountable for these systems, and how would we know if they failed?

Management system standards make the same point from another direction. ISO/IEC 42001 follows ISO's common management system structure, which includes a leadership clause requiring top management to show commitment to the system. Leadership cannot do that without regular, decision-ready information about it.

A one-page structure

1. What we run

  • Count of AI systems in the inventory, by risk tier (assistive, decision support, acting, regulated).
  • New systems this period and systems retired.
  • Inventory coverage: systems found by discovery that were not registered.

2. What they can do

  • Systems that take actions without per-item human approval, and what actions.
  • Any expansion of permissions, data access or user groups approved this period.

3. What went wrong

  • Incidents by severity, with time to contain and the fix.
  • Releases shipped with accepted exceptions instead of passing evidence, and their age.

4. What is changing outside

  • Upcoming regulatory dates that affect the company. For example, the European Commission states that EU AI Act rules for Annex III high-risk uses are now scheduled to apply from 2 December 2027 and for Annex I product-embedded systems from 2 August 2028 (Commission AI Act page). See AI compliance dates to track.
  • Material changes from key AI vendors (new model versions, retired models, terms changes).

5. Decisions needed

  • Risk acceptances above management's authority.
  • Proposals to let a system act where it now only drafts.
  • Budget or staffing for controls that are behind plan.

Questions directors can ask

  • Which of our AI systems could cause harm to a customer, employee or our finances without a person checking first?
  • Who is the named owner of each of those, and when were they last reviewed?
  • How do we find AI tools that staff are using outside the approved list?
  • What evidence do we require before a system is allowed to take actions, and who signs it off?
  • What was our most serious AI incident this year, and what changed afterwards?
  • Which regulations apply to us, and what are the dates?
  • If our main model provider changed or withdrew a model tomorrow, what would break?

How to measure the reporting itself

A board report is working if it changes decisions. Track whether each report produced at least one decision or follow-up, whether the same metrics are comparable quarter to quarter, and whether incidents first reached the board through the report or through something else. The last one matters most.

Decision framing

Present AI value to the board as decisions it improves and work it removes, with the measurement method stated, rather than as projected dollar savings. Pair every request to expand a system's autonomy with the evidence that supports it and the control that limits the downside. Boards approve faster when the risk side is shown as clearly as the benefit.

Common mistakes

  • Reporting pilot counts as progress. A pilot that never reaches production is cost, not value.
  • Mixing vendor marketing claims into board material without your own measurements.
  • Changing the metrics every quarter, which hides trends.

Sources