On 2 August 2026, the EU AI Act’s rules for standalone high-risk systems were supposed to take effect. They will not. Six days before the deadline, the EU brought an amendment into force deferring those obligations to December 2027, because the standards and implementation machinery needed to make them workable were not ready. In this article, I set out what the deferral reveals about Boards which anchored AI governance to regulatory deadlines, and where the accountability that never moved actually sits.

Every AI business case counts the hours saved. Almost none counts the hours added: the reviewing, correcting, and supervising that AI outputs demand before anyone can rely on them. That oversight labour is real, regulation increasingly mandates it, and people absorb it silently on top of existing roles. BCG research published in March 2026 shows the consequence: a distinct mental fatigue attaching to heavy AI oversight loads, while workers using AI only to replace routine work report less burnout, not more. I argue this fatigue is a control cost Boards must price, and the remedy is governance, not resilience training.

In the Redeployment Dividend I argued that AI’s real prize is releasing intellectual capital from undifferentiated work, not cutting headcount. The evidence has now caught up with the argument, and it is uncomfortable. Teams that deploy AI save the equivalent of five hours per person per week, yet most of that time drains into low-value work, and nine in ten executives report no measurable productivity impact at their own firm. The saving is real; the value is not arriving. In this article, I argue that the dividend leaks because nobody owns it. AI owns execution and managers own the workflow, but unless the Board owns whether freed capacity creates value, the hours AI recovers will simply refill with the work that was already there.

In The Great Remaking I established that businesses are being remade around AI. But “remake with AI” is not “put AI into everything”, and the difference between them is judgement. Asked recently how I decide which AI to use, I said that I do not start there. In this article, I argue for the three questions that come first, and that the one doing the real work is not the question about tools at all, but the one that asks how a piece of work is done today and whether it still needs doing, because a well-judged no is what makes every yes credible.

The market now treats visible AI adoption as proof a company is driving forward through innovation, and the chief executive is the one expected to show it. Being seen to adopt, not misleading the market, and choosing well are three demands held at once. In this article, I argue that AI does not rewrite the chief executive’s duties; it changes the conditions under which they are discharged. The task is to choose the few bets that matter, change how the company works around them, and answer for them without delegating the accountability.

Every position in the AI Sovereignty Trilemma carries a cost, but only one is shown to a Board before it is paid. The visible cost is that sovereign capability is dearer, which is where most sovereignty conversations stop. The hidden cost belongs to the convenient alternative, frontier capability bought cheaply and governed elsewhere, and it stayed invisible only because the control it surrenders had never been tested. On 12 June a directive tested it, forcing a provider to withdraw two frontier models from every customer overnight. In this article, I argue that model availability is a continuity risk a Board must own, and that the task is not to solve the Trilemma but to know which cost the organisation is paying, and to have chosen it.

The case for AI in the finance function is no longer in question. Commitment to it now runs well ahead of readiness, but accountability does not wait for that gap to close. The CFO answers for the integrity of the accounts and the stewardship of capital every reporting cycle, ready or not, and AI is already inside the work that produces both. In this article, I argue that AI changes how each of the CFO’s duties is carried out, not who signs for them, and I sort its effect into four groups: where it does the work, where it sharpens the judgement, where it operates out of sight, and where it changes almost nothing.

The strategic value of data is no longer in question. The next frontier is data whose meaning and relationships are explicit enough for machines to reason over rather than merely retrieve. Unstructured information must be interpreted by whoever consumes it, whether that is a person reading a report or an AI system generating an answer. Structured information makes explicit the relationships required for traceability, verification, and defensible reasoning at scale. In this article, I argue that ontologies and the knowledge graphs built upon them have moved from technical infrastructure into Board territory, because they increasingly determine what an organisation officially knows, what its AI systems can work with, and where durable advantage is created.

The company secretary’s role was built to maintain the conditions under which directors can apply judgement and the company can meet its governance obligations. Both are now being remade: by AI tools inside board administration that compose the materials directors will judge, and by AI deployments inside the business that shape the compliance position the secretary must disclose. This article works through how Cadbury, the Companies Act, the FRC’s 2024 Code, and the Chartered Governance Institute set out the secretary’s responsibilities, none dispensable, all now requiring different execution. The chair polices the boundary between agency transfer and accountability transfer. The secretary operates that boundary in practice.

A foundation model arrives carrying a value system its provider built: what it refuses, how it frames a sensitive subject, how it resolves a question with reasonable views on either side. That standard, not the organisation’s, is the one in force, and it changes with each model version without the Board’s consent. System prompts, retrieval, guardrails, and fine-tuning constrain the imported standard but cannot re-author it. Organisations can choose to accept the provider’s ethics, reject the deployments where it bears on people, or build alignment the organisation owns. This article sets out how a Board makes that choice, deployment by deployment.