Where the Costs Land

The Information reported on 14 September that Nvidia, Palantir, and Booz Allen Hamilton are among the organisations limiting which work they send to the most capable frontier models. Not because of a breach, actual or feared, but because the work itself may be the leak: judgement an organisation has paid for over decades, passed through a tool, improving a model any competitor can buy.

The three pieces in this issue each find a cost of frontier AI that lands where nobody planned for it. Nothing Has to Leave the Building names the exposure first: an organisation's best-paid thinking now happens inside AI tools, and the risk falls between security, legal, and human resources, so nobody reports it upward. The Supplier Grades the Customer turns the relationship round. Access to the most capable models is already something the supplier allocates by grading the customer. A slowed frontier would make that grading universal, and the proposal says nothing about open-weight models, which leaves the customer one uncertified exit. Managers Own the Workflow brings the pattern home: the business case counts the hours agents save, but the saving becomes real, or does not, in a manager's job that was never redesigned to supervise them.

The thread is the gap between where a decision is made and where its cost settles. The supplier sets the pace, the executive approves the case, and the exposure, the dependence, and the supervision arrive unpriced somewhere else in the organisation. The AI Risk and AI Business Cases briefings hold the standing questions on both sides of that gap, and the Agentic AI signal frames the delegation decision that sits underneath it.

If your time is limited, I particularly recommend Nothing Has to Leave the Building. It closes with three questions a Board can put at its next meeting about which work goes into frontier tools and who decides, and it answers the strongest objection to the worry before a confident executive has the chance to raise it.

How is your organisation deciding which work goes into frontier models, and who would know if the answer is nobody?

— Mario

Nvidia, Palantir, and Booz Allen Hamilton are reported to be restricting which work reaches the most capable frontier models, and their reasoning deserves a Board's attention. The exposure is not a data breach but the slow transfer of an organisation's judgement into a model its competitors can also buy. This article explains why the risk falls unreported between security, legal, and human resources, answers the strongest objection to taking it seriously, and arms the Board with three questions for its next meeting.

Anthropic's chief executive wants the frontier laboratories to slow the rate at which their models improve, and OpenAI's agrees, yet the proposal says nothing to the customers who depend on those models. This article reads the argument from the customer's seat: access to the frontier is already something the supplier allocates by grading the customer, a checkpoint regime would make that grading universal, and the silence on open-weight models leaves one uncertified exit. Three questions for the risk register close the piece.

The business case for AI agents counts the time they save; whether that time becomes value is settled in the manager's workflow, a job that absorbs the supervision without any change to its description. Drawing on Gartner, Harvard Business Review, Gallup, and McKinsey, this article traces how the agent manager role is arriving by accretion while the routine work that once grew the next generation of managers goes to the agents, and closes with five questions for the executive before the next case is approved.

Board Briefings are my standing references for directors: each one takes a single governance topic and sets out what a Board needs to know, ask, and decide.

Seeing and pricing the risks AI actually creates

Investment cases that survive contact with the Board

Signals are my short readings of developments worth a Board’s attention: what changed, why it matters, and what to watch next.

Generative models are being given goals, tools, and the authority to act. The Board question is where to transfer agency, and under what limits.

Knowledge graphs make relationships and provenance explicit, helping enterprise AI produce answers that are more consistent, traceable, and easier to govern.

When AI changes how you work, lead it before it leads you. Remake is my framework for remaking a business around AI: the stages the change moves through, the assets the Board signs off, and the verdicts that keep it honest.