Podcast

Episode notes

Ethical AI: When the Model Imposes Values Your Organisation Did Not Choose

A foundation model arrives with a value system its provider built and the Board did not choose. The decision: accept it, reject it, or build.

An AI model is in production somewhere in the organisation, handling a difficult decision: a customer complaint, a redundancy query, a medical underwriter reviewing a claim. The model settles what to refuse, how much candour the moment can bear, where the customer’s interest gives way to the policy. It does this the same way each time, because the judgement was made long before the question arrived — not by the organisation running the model, but by the provider that built it, for a global product, before the organisation signed up to use it.

In this episode of The Board in the Machine, Mario Thomas — Chartered Director and Fellow of the Institute of Directors — examines the value system every foundation model carries into deployment, why the familiar controls only partly contain it, and the strategic choice a Board is left holding once it sees the problem clearly.

The argument draws on the 2026 arXiv paper “Alignment Drift in Multimodal LLMs”, which found large and persistent differences in how model families handle ethically sensitive questions; the 2025 withdrawal of a major model update after it became excessively agreeable; Stanford’s Foundation Model Transparency Index, which scored major providers at roughly 40 out of 100; and the disclosure obligations of the EU AI Act. Against that evidence the episode sets out the real decision — accept, reject, or build — and frames it through the Six Board Concerns, the AI Sovereignty Trilemma, and the discipline of Minimum Lovable Governance.

This episode is for Boards and directors who want to govern the ethics their AI runs deliberately, deployment by deployment, rather than inherit it by default. Read the full article at mariothomas.com

Episode notes

The Headroom Argument: Why AI Efficiency Means More Compute, Not Less

Architectural efficiency expands AI compute demand rather than reducing it. Three forces converge on more inference. Boards should read efficiency news as capability, not cost.

A new AI architecture lands on 5 May. Subquadratic launches SubQ: a 12-million-token context window on a sub-quadratic sparse-attention architecture that reduces attention compute by roughly 1,000 times at full context. The interesting question is not whether AI is about to get cheaper. It is what efficiency news actually says about compute demand. The day after SubQ launched, Anthropic announced a partnership at SpaceX’s Colossus 1 facility adding more than 300MW of new capacity and over 220,000 NVIDIA GPUs. Both kinds of news end in the same place: more inference, not less.

In this episode of The Board in the Machine, Mario Thomas — Chartered Director and Fellow of the Institute of Directors — examines why architectural efficiency expands AI compute demand rather than reducing it. The episode walks through the three forces that drive demand faster than per-unit cost falls, why every prior era of computing tells the same story, and how Boards should read efficiency news to fund the right opportunity rather than the wrong budget.

The argument draws on the SubQ launch, the Anthropic-SpaceX Colossus 1 partnership, Mozilla’s disclosure that Anthropic’s Claude Mythos Preview identified twelve times as many Firefox vulnerabilities as Claude Opus 4.6 had found earlier in the year, Goldman Sachs’ Powering the AI Era, Deloitte’s TMT Predictions 2026, and Jevons’ nineteenth-century observation that improving the efficiency of a resource raises its total consumption rather than lowering it. The takeaway is operational: the Six Board Concerns, AI Stages of Adoption, and the AI Sovereignty Trilemma frame the question, and Minimum Lovable Governance answers the design question that follows when cheaper inference accelerates probabilistic decision-making into the regulated decision space.

This episode is for Boards and directors revisiting AI strategy in light of efficiency announcements and capacity commitments, and who want a capability-first framing rather than a budget-first one. Read the full article at mariothomas.com

Episode notes

The Reasoning Gap: The Capability the Law Now Demands of Boards

UK law now requires four safeguards for solely automated decisions. Most Boards have approved probabilistic systems that cannot deliver them in operation.

A short statutory instrument lands on 12 May. It directs the Information Commissioner to prepare a statutory code on AI and automated decision-making. The interesting question is not what the code will say. It is what the law already requires. Since 5 February, UK law has required four safeguards for any significant decision taken solely by automated processing - information, representations, human intervention, and the right to contest.

In this episode of The Board in the Machine, Mario Thomas — Chartered Director and Fellow of the Institute of Directors — examines the capability gap that sits between those four safeguards and the systems most Boards have already approved. The episode walks through what the law actually asks for, why rule-based systems carry that capability on the surface and probabilistic systems do not, and where the gap will surface first when the first significant decision is contested.

The argument draws on the Data (Use and Access) Act 2025, the new Articles 22A to 22D of the UK GDPR, the CJEU’s SCHUFA judgment, the WP29 guidelines on automated decision-making and profiling endorsed by the EDPB, the IoD’s AI Governance in the Boardroom (2025), and practitioner analyses from Travers Smith, Bird & Bird, Debevoise, and Alston & Bird. The takeaway is operational: Minimum Lovable Governance is the operating principle through which a duty like this one actually gets delivered, and the Board’s job is not to build the capability but to refuse to approve systems that cannot deliver it.

This episode is for Boards and directors in financial services, employment, insurance, and any consumer context where significant decisions are being made by automated processing, and who want a capability-first framing rather than a compliance checklist. Read the full article at mariothomas.com

Episode notes

AI and the Chair: Governing the Board Through The Great Remaking

Existing chair responsibilities now require different execution as AI remakes both the Board's own work and the work the Board governs.

Chairs remain accountable for the Board’s effectiveness. They are no longer fully in control of how decisions are being formed. AI is remaking both the Board’s own work and the work the Board governs at the same time.

In this episode of The Board in the Machine, Mario Thomas — Chartered Director and Fellow of the Institute of Directors — examines how AI has changed the execution of the chair’s existing responsibilities. The episode walks through the two states of the duality the chair now sits between: AI in the preparation of board materials, and AI in the operations of the business the Board governs. Listeners will come away with a sharper view of where collective accountability is most at risk in their own boardroom, and what the chair’s existing responsibilities now require to keep it intact.

The argument draws on the Cadbury Report of 1992, the FRC’s 2024 UK Corporate Governance Code, the IoD’s NEDs Reimagined Commission of January 2026, and the 2026 Global Board Governance Survey from Protiviti and BoardProspects. The takeaway is operational: a chair who can name where the boundaries of agency and accountability are silently moving in their own boardroom is a chair who can hold them.

This episode is for chairs and senior independent directors operating in boards where AI has already entered both the preparation room and the operating environment, and who are looking for a constitutional framing rather than another tool list. Read the full article at mariothomas.com.

Episode notes

The Appreciating Ledger: When AI Capital Outgrows the CFO's Rulebook

AI capital appreciates, accumulates, and crosses functions. The four indicator types let CFOs see what conventional project-ROI models structurally cannot.

Only 31% of CFOs are satisfied with their AI outcomes, yet that figure rises to over 60% among CFOs whose organisations sit at the top of the AI adoption curve. Satisfaction tracks scale, and the gap is not about execution.

In this episode of The Board in the Machine, Mario Thomas — Chartered Director and Fellow of the Institute of Directors — explains why AI capital outgrows the CFO’s conventional measurement instruments. Listeners will come away understanding why AI capital appreciates rather than depreciates through use, why its returns accumulate across functions that did not fund them, and why project-level ROI cannot see the value the ledger is missing.

Drawing on Bain’s April 2026 CFO Survey, PwC’s 2026 AI Performance Study, BCG’s Widening AI Value Gap research, and the Brookings Institution’s work on AI in national statistics, the episode sets out a practical scorecard extension the finance function can install now, covering four indicator types, capability as an asset class, portfolio discipline, and the reinvestment default. It also shows why IAS 38 already permits the treatment AI capital requires, and why the audit committee conversation about AI on the balance sheet is about to become routine.

This episode is for CFOs, finance directors, audit committee chairs, and board members with financial oversight remits, particularly those whose organisations are scaling AI investment without a matching evolution in how that investment is measured. Read the full article at mariothomas.com.

Episode notes

Maximum Fidelity: How Four Indicator Types Strengthen Board Decisions

Four indicator types give boards progressively higher decision fidelity: lagging, leading, predictive, and reasoned. Together they represent the most accountable governance instrument available.

Boards have always made decisions under incomplete information. What has changed is not the volume of what reaches the table — it is the quality of what the Board can see before the vote is taken. Most Boards still rely on instruments built for a slower world.

This episode of The Board in the Machine, with Mario Thomas — Chartered Director and Fellow of the Institute of Directors, introduces a framework of four indicator types: lagging, leading, predictive, and reasoned. Listeners will leave with a clear way to assess what their own Board is currently using to make major decisions, and what is missing.

Drawing on Heidrick & Struggles’ 2026 CEO & Board Confidence Monitor, the Institute of Directors’ work on Board oversight, and three worked Board scenarios — a major acquisition, a market entry, and a significant capital allocation — the episode shows how each indicator type adds a different quality of input, and what changes when reasoned indicators prove what was previously only estimated. The takeaway is a question every director can put to their own Board after listening: which of the four indicator types informed our last major decision, and which were absent?

This episode is for chairs, non-executive directors, and senior executives sitting on Boards where the next major decision is already on the calendar — and where the information environment around it has not been examined recently enough.

Read the full article at mariothomas.com.