Board altitudeChallenge the commitment, not every configuration

The G20/OECD Principles of Corporate Governance place strategy, risk, accountability, internal control and reporting integrity within board responsibility. The UK Corporate Governance Code 2024 gives a more specific framework within its own applicability. Neither turns directors into product owners. The board sets direction, determines or oversees risk appetite, tests management’s reasoning, allocates capital and holds executives to account. Management defines workflows, selects controls, operates systems and brings material departures back for decision.

That division matters. A board that asks only “Are we using AI?” is too remote. One that approves prompts, models or every low-consequence release is too close. The useful altitude is the consequential commitment: what changes for customers, workers, investors or the enterprise; how the full system produces that change; what evidence supports it; which exposure is accepted; and what conditions govern scale or exit.

The strong rival is the AI dashboard. Keep it. A concise recurring view of portfolio size, value, incidents, spend, concentration and action status helps directors notice movement. But management’s colour cannot substitute for the board’s reasoning. Aggregated use-case counts make a payment agent look like another writing assistant. Average accuracy can hide a harmed subgroup. “Compliant” can conceal an untested recovery path. The ten questions should supply the definitions and proof behind a dashboard, not create a second encyclopaedia beside it.

Reading disciplineInspect the proof object, the seam and the exception

A proof object is useful because it narrows an argument without pretending to settle every question. A decision statement can show the population and baseline. A system map can show where authority crosses an interface. An evaluation can show observed performance under stated conditions. None proves the next object automatically. Directors should therefore read the seams: whether the evaluation population matches the affected population, whether the control reaches the downstream action, whether the cost case includes the recovery design, and whether the named owner actually holds the relevant right.

Dates are part of the evidence. A model test completed before a retrieval change, supplier release or policy revision may no longer support the current boundary. A takeover drill run at small volume may not establish capability at peak load. An acceptable unit cost before review queues grow may not survive wider adoption. The pack should state the change that would stale each important object. That makes refresh an evidence decision instead of an automatic annual ritual.

Directors should also ask what management chose not to aggregate. A mean can conceal a tail. A total can merge reversible and irreversible authority. A single incident count can combine harmless user corrections with uncontained customer action. A portfolio view can distribute one shared supplier failure across dozens of rows, making concentration look diversified. Disaggregation is not always better, but the grouping rule should follow the decision consequence rather than the convenience of available reporting.

Finally, inspect exceptions. The organisation’s real control system often appears in overrides, informal workarounds, delayed reviews, unowned alerts and temporary supplier concessions. An exception register is not proof that operation is weak; it is evidence about where the designed boundary meets reality. The board needs the pattern, consequence and management response, not every ticket. Repeated exceptions that alter exposure should reopen the commitment even when headline performance remains stable.

Questions 1–4Move from strategic claim to contextual evidence

Evidence boundaryMake uncertainty governable without turning it into a colour

Management cannot remove every uncertainty before an AI system operates. The board’s task is to decide which uncertainty can be carried under which exposure. Three distinctions help. Epistemic uncertainty concerns what is not yet known about behaviour or value. Operational uncertainty concerns how people, integrations and suppliers will behave in use. Decision uncertainty concerns whether the proposed commitment remains preferable given the range of outcomes. Each may need a different condition, owner and response.

A disconfirming test gives uncertainty a direction. For a service assistant, it might ask whether improvement disappears for disputed bills or customers using assistive channels. For a payment agent, it might ask whether loss concentrates during identity degradation or a supplier delay. For a drafting tool, it might ask whether saved authoring time reappears as senior verification effort. The test should threaten the claim management most relies on, not merely add another metric that is easy to pass.

Conditional approval is therefore a decision form, not a softer green status. It names the permitted population, authority and duration; the evidence still missing; the threshold that prevents scale; and the executive who must return. When those terms are absent, a pilot can expand through local convenience without a new board decision. When they are explicit, incomplete evidence can produce disciplined learning while downside remains bounded and reversible.

Questions 5–7Make authority, capability and economics testable

Joined accountabilityKeep risk, value and authority in the same decision

Weak proposals often separate the upside owner from the exposure owner. A business sponsor presents gross capacity, technology presents model quality, risk presents a control inventory and finance sees an investment total. Each statement may be accurate within its page, yet the board still lacks one accountable commitment. The proof pack should reconcile them around the consequential decision: who captures value, who bears exceptions, which cost centre funds retained capability and who accepts the downside that remains after controls.

Decision rights also carry economic information. A human approval that cannot be exercised within the service promise is not a feasible control. A supplier exit that requires rebuilding evaluations and context has an option cost. A rollback that preserves customer obligations but consumes weeks of expert work has a recovery cost. Putting these constraints into the economics discourages proposals that appear attractive only because another function absorbs the work.

Observational research on board-level IT governance can inform the case for board involvement, but it cannot prove that a particular committee design causes performance. The stronger basis for the ten questions is the board’s established role in strategy, risk, internal control and accountability, combined with AI-specific proof domains from NIST and GAO. The Console and its routing rules are this article’s synthesis, not a requirement found in any one standard.

Questions 8–10Govern dependencies, intervention and the end

The board AI proof packTen answers, one connected decision record

The output is not ten polished essays. It is a compact proof pack that links each management claim to an evidence object, its date, the system and population boundary, an evidence owner, an independent challenger, a material assumption, a disconfirming test, the consequence if wrong, the decision requested and any open condition. Labels should distinguish observation, modelled estimate, assumption, legal interpretation and management judgement.

The chain matters more than volume. If Question 1 changes, the evidence population in Question 4 may no longer fit. If Question 2 reveals tool authority, the maximum consequence in Question 3 changes. If Question 5 exposes an assurance conflict, the apparent strength of Question 4 weakens. If Question 8 shows that context cannot move, the retirement evidence in Question 10 is incomplete. The pack should make these dependencies visible rather than bury them in appendices.

A committee can own preparation without fragmenting accountability. Audit and risk may scrutinise control and assurance; remuneration may consider incentives; technology or sustainability committees may examine specialist matters. The full board still needs an integrated commitment where strategy, risk and capital meet. Repetition across committees is not extra assurance when each receives a different boundary.

In the meetingAsk ten questions without staging a ten-hour hearing

The chair can organise discussion around breaks in the chain. Management begins with the decision requested and the strongest evidence for it. Directors then test the most consequential seam: perhaps whether the system boundary includes downstream action, whether the affected population matches the evaluation, or whether the stop role has authority. Questions whose proof has not changed can remain in the pack without consuming meeting time. Open conditions, disputed assumptions and changed exposure receive attention.

The board minutes should record the commitment, material evidence considered, assumptions that carried weight, conditions, delegations and the reason the route fits appetite. They need not reproduce technical debate. A concise decision record helps later directors distinguish what was known from what emerged and prevents an old approval being cited for a materially different system. It also makes retirement possible because the organisation can identify which authority is being withdrawn.

When management cannot answer, the next route depends on consequence. An incomplete portability test may justify a condition for a capped learning stage. An assurance conflict may require redesign before evidence can be relied upon. A live incident without stop authority requires containment, not another paper. Evidence that misses the affected population while credible loss exceeds appetite calls for stopping the system. “More information” is not itself a board route; the exposure carried while information is gathered must still be decided.

Interactive instrumentBoard AI Proof Pack Console

The Console stores the same eight fields for each question and then tests the cross-pack conditions that determine a board route. It does not produce a maturity, trust or traffic-light score. The output is a decision with an evidence reason and the next proof obligation.

Board cadenceUse the questions as gates, then govern exceptions

At initial commitment, management should answer all ten at the depth proportionate to consequence. A bounded learning decision may carry open evidence conditions with dates and an accountable executive. A scale decision should close the conditions that make wider exposure acceptable. A retirement decision should remain open until credentials, records, integrations and customer or staff reliance have been resolved.

Recurring reporting can then be short. Show changes in the consequential decision, system boundary, evidence, maximum consequence, authority, capability, economics, concentration, intervention state and exit conditions. Highlight which original assumption failed, which threshold moved and which decision is now requested. A stable answer need not be rewritten for every meeting. A material boundary change must not hide behind an unchanged dashboard tile.

The proof-pack approach also has a failure boundary. It becomes bureaucracy when every uncertainty produces another document request, when directors rehearse technical detail they cannot govern, or when committees duplicate management decisions. Missing evidence should route to a bounded learning condition when exposure permits it. It should not be treated as automatic proof of failure. Conversely, a large dossier cannot rescue an undefined decision or a consequence outside appetite.

Good board questions change management’s preparation. Teams define the decision before the model, trace the action before reporting accuracy, design stop authority before monitoring, test portability before depending on it and plan retirement before the system becomes infrastructural. The result is not certainty. It is a visible chain of claims, boundaries, evidence and authority strong enough for a board to make a bounded commitment and know when that commitment no longer holds.