The Board in the Machine
Podcast
Mostly the audio edition of the writing: articles on AI, governance, and the boardroom, read and expanded for the road. New episodes monthly.
Browse the episodes (44)
The Return of Traditional AI: Organisations Are Rethinking Their LLM-First Strategies
13 minute · The audio edition of the article.
Episode notes
The Return of Traditional AI: Organisations Are Rethinking Their LLM-First Strategies
Forty-two percent of companies abandoned most of their AI initiatives this year, often because generative AI was applied to problems traditional methods solve better.
Forty-two per cent of companies abandoned the majority of their AI initiatives this year, up from seventeen per cent in 2024. The pattern behind the number is consistent: organisations applied generative AI to problems that traditional machine learning or deterministic automation solve better, and they are now recalibrating.
In this episode of The Board in the Machine, Mario Thomas, Chartered Director and Fellow of the Institute of Directors, examines that recalibration. S&P Global supplies the abandonment figure, Forrester predicts generative AI will orchestrate less than one per cent of core business processes in 2025, the World Quality Report finds sixty per cent of organisations citing hallucination and reliability as top concerns, and BCG’s survey of more than 1,250 firms finds only five per cent achieving AI value at scale. He notes that Salesforce has been building deterministic automation into Agentforce, its Agentforce CTO observing that an LLM given more than about eight instructions starts dropping them. This is maturation, he argues, not failure.
Mario sets out where each approach belongs. Classification against fixed criteria, pattern recognition in structured data such as fraud detection and predictive maintenance, and precision-dependent calculation sit with rules engines and traditional machine learning. Contextual understanding, summarisation, semantic search, and creative generation are natural LLM territory. Hybrid architectures let deterministic systems execute while generative AI supplies context. For Boards, he draws three diagnostics from the Complete AI Framework: problem domain assessment, honest capability matching that includes inference cost at scale, and risk and reliability assessment. He closes with a maturity arbitrage mapped onto the AI Stages of Adoption: proven traditional AI builds confidence at the Experimenting stage, while hybrid architectures serve Adopting and Optimising organisations.
This episode is for directors and Boards facing AI budgets under scrutiny. The question is not whether to invest in AI, but whether each investment deploys the right AI for the specific challenge. Read the full article at mariothomas.com
Read the article →A New Grid Actor: AI Infrastructure Is Becoming Energy Infrastructure
14 minute · The audio edition of the article.
Episode notes
A New Grid Actor: AI Infrastructure Is Becoming Energy Infrastructure
AI infrastructure operators building their own generation become grid actors rather than consumers, and that changes energy economics, nowhere more sharply than the UK.
The USA faces a 19GW power shortfall by 2028, over 40 per cent of projected data centre demand, and with eight-year grid interconnection queues against 18-month GPU refresh cycles, hyperscalers and AI labs are building their own generation. Capacity built to secure AI workloads typically exceeds what those workloads need, and that surplus turns infrastructure operators from energy consumers into grid actors.
In this episode of The Board in the Machine, Mario Thomas, Chartered Director and Fellow of the Institute of Directors, examines the scale of the shift and what follows from it. OpenAI’s Stargate facility includes 361MW of on-site generation, Meta’s Prometheus cluster adds 200MW, and combined hyperscaler nuclear commitments now exceed 8GW. Wärtsilä and AVK project European data centre demand rising 250 per cent by 2030, with 40 per cent of existing AI data centres constrained by power availability by 2027.
Mario then works through the progression from consumer to prosumer to grid actor. A 1GW campus at 60 per cent average utilisation holds 400MW of dispatchable capacity off peak, functionally a power station, and that excess can be sold through demand response programmes. For the UK he weighs opportunity against risk: the AI Energy Council, Rolls-Royce SMR’s 470MW units arriving in the mid-2030s, and a 285 TWh system in which a single player at 5 per cent of capacity raises concentration and foreign-ownership questions. He names the emissions trade-off, gas turbines filling the gap until SMRs arrive, and closes with three questions: whether to assess generation potential alongside compute, where backup supply becomes grid export, and what frameworks belong in place before an AI investment makes the organisation an energy market participant.
This episode is for UK Boards and directors who can no longer separate energy strategy from AI strategy. The organisations that anticipate this transition will shape it; those that do not will adapt to rules others write. Read the full article at mariothomas.com
Read the article →The AI Maturity Mirage: Diagnosing the Gap Between Investment and Readiness
16 minute · The audio edition of the article.
Episode notes
The AI Maturity Mirage: Diagnosing the Gap Between Investment and Readiness
Boards overestimate AI maturity by counting tools and pilots rather than capability. Three patterns create the illusion, and each can be diagnosed before it misleads.
Boards reviewing AI progress see pilots underway, tools adopted across teams, and early efficiency wins. That view misleads, because visible activity bears little relation to genuine organisational capability. The gap between the two is the AI maturity mirage, and it derails transformation strategies before anyone notices.
In this episode of The Board in the Machine, Mario Thomas, Chartered Director and Fellow of the Institute of Directors, examines the disconnect and the three patterns that produce it. Larridin’s State of Enterprise AI 2025 report finds that 89% of enterprises have adopted AI tools while only 23% can accurately measure their return on investment, and IBM and Ecosystm’s 2025 APAC research shows 85% of organisations claiming data-driven or AI-First status against 11% demonstrating true readiness. The patterns are the tool-centric illusion, where deployments are counted as maturity; the pilot success trap, where isolated wins are read as systemic progress; and hype-driven metrics, where advertising agencies rate AI criticality at 8.1 out of 10 yet embed it in only 16% of operations.
Mario then sets out a three-step diagnostic. Map each function independently to one of the five AI Stages of Adoption, treating pilots that consistently fail to scale within six months as a sign of Experimenting-stage capability. Evaluate balance across the Five Pillars, since tool adoption without governance, people, and value realisation maturity is the signature of the mirage. Test with leading, lagging, and predictive indicators together, because over-reliance on lagging measures inflates perceived maturity. He matches remedies to the gaps: an AI Centre of Excellence with minimum lovable governance, infrastructure coherence before more tools, people programmes, the hub-and-spoke model, and Well-Advised scorecards.
This episode is for directors and the Boards who suspect perceived Optimising status may reflect Experimenting capability. The mirage persists for organisations that mistake activity for capability; for those willing to look clearly, the path to genuine maturity becomes visible. Read the full article at mariothomas.com
Read the article →Minimum Lovable Governance: The AI Operating Principle Boards Should Use
17 minute · The audio edition of the article.
Episode notes
Minimum Lovable Governance: The AI Operating Principle Boards Should Use
Minimum lovable governance replaces episodic compliance with continuous, embedded oversight people actually want to use: guardrails that earn adoption rather than enforce it.
AI governance is too often heavy where it should be light and light where it should be heavy: elaborate approval processes for low-risk experiments, minimal oversight of high-stakes autonomous systems, and policy documents nobody reads. The result is friction without assurance. Unloved governance gets routed around, and shadow AI shows how often.
In this episode of The Board in the Machine, Mario Thomas, Chartered Director and Fellow of the Institute of Directors, explains minimum lovable governance, a concept he has referenced across several articles, and why it makes AI governance work. It borrows Eric Ries’s progression from Minimum Viable Product to Minimum Lovable Product and applies it to oversight: the smallest system that achieves the necessary guardrails and that people actually want to use. UpGuard’s November 2025 report found more than 80 per cent of employees, and nearly 90 per cent of security professionals, using unapproved AI tools: governance that exists on paper but fails to govern.
Mario sets out what lovable means operationally: embedded in how work happens, continuous rather than episodic, proportionate to risk, and clear at the point of decision. He explains why the approach has only recently become viable, pointing to the EU AI Act’s risk-tiered architecture, ISO 42001 and ISO/IEC 42006:2025 as reference points, and AI-assisted governance making continuous oversight achievable while accountability stays with humans. He then works through five principles, each with a Board test question, such as whether an auditor calling tomorrow would be answered in seconds or weeks, and whether a marketing chatbot goes through the same approval process as a credit decisioning system.
This episode is for directors, chief executives, and the Boards whose governance reports say everything is under control while most of the organisation works outside it. The choice is to build governance people route around, or governance people want to use, and the outcomes will differ accordingly. Read the full article at mariothomas.com
Read the article →World Models: The Next Horizon in AI for Predictive Enterprise Intelligence
15 minute · The audio edition of the article.
Episode notes
World Models: The Next Horizon in AI for Predictive Enterprise Intelligence
World models move AI from reacting to anticipating: systems that simulate future scenarios, with aviation and finance already seeing the operational gains.
Current AI excels at analysing what has happened. World models, an emerging approach, build internal simulations of how reality works and predict what will happen next, moving AI from recognising patterns to anticipating futures. For Boards already tracking leading and lagging indicators, they promise the technical realisation of the predictive indicator.
In this episode of The Board in the Machine, Mario Thomas, Chartered Director and Fellow of the Institute of Directors, examines world models as the next horizon in predictive enterprise intelligence, drawing on Yann LeCun’s work on the Joint Embedding Predictive Architecture and on the gains current predictive AI already delivers: aviation pioneers reporting twenty to thirty per cent improvements in operational efficiency in BCG’s AI-First Airline report, and PwC’s 2025 finding that industries exposed to AI disruption achieve three times higher revenue per employee growth. He is candid about the barriers, which LeCun himself acknowledges as huge practical impediments: computational demands that may need quantum breakthroughs, vast real-time data requirements, and energy consumption that could make large-scale world models economically unviable.
Mario sets out grounded timelines, with narrow-domain world models perhaps two to three years away and enterprise-ready systems five to seven, and argues for a dual strategy: build foundational predictive capabilities with current AI now whilst monitoring world model developments for strategic timing. He imagines the transformation across aviation, financial services and manufacturing, then turns to governance: the ethical boundaries Boards must set when a simulation reveals a profitable but harmful path, the case for minimum lovable governance, transparency about the scenarios and assumptions behind each prediction, and risk management that simulates whole consequence chains rather than single events.
This episode is for directors and Boards preparing for predictive intelligence. Which operational predictions would transform the organisation’s competitive position when world models arrive? That answer determines where to build foundations now. Read the full article at mariothomas.com
Read the article →The Accountability Gap: When AI Delegation Meets Human Responsibility
19 minute · The audio edition of the article.
Episode notes
The Accountability Gap: When AI Delegation Meets Human Responsibility
Organisations are transferring decision-making agency to AI while accountability stays with people, and approving deployments without the verification capability that accountability needs.
Research from Tenable found that 63% of organisations lack any AI governance policies, yet McKinsey’s research shows 88% using AI in at least one function. Behind that gap sits a mistaken assumption: that delegating work to AI delegates accountability for its quality. Courts, regulators, and clients consistently reject that assumption, and organisations planning to replace junior roles with AI are reducing the capacity needed to verify what AI produces.
In this episode of The Board in the Machine, Mario Thomas — Chartered Director and Fellow of the Institute of Directors — argues that AI is best understood as a brilliant graduate trainee: highly productive, short on judgement, and needing expert review before its work reaches a client. He explains why AI amplifies rather than reduces the need for subject matter expertise, and why a Board that approves deployment without matching investment in verification is approving the conditions for the failures now documented in court cases worldwide.
Mario works through the evidence: Stanford research finding that general-purpose AI models produce false information on legal queries between 58 and 82% of the time, research based on payroll data suggesting a 13% decline in AI-exposed entry-level roles, a BCG experiment finding productivity gains of 20 to 40% when AI augmented junior and senior staff, and PwC’s AI Jobs Barometer showing three times the revenue growth per employee in AI-exposed industries, but only with skilled oversight. He sets out the five investments a Board should approve alongside any AI deployment and the questions to require answered first, including who has the expertise to verify the outputs today, and where the senior experts will come from in 2030.
This episode is for directors, chairs, and executives weighing proposals to capture AI efficiency gains by replacing junior roles, because the expertise pipeline those roles sustain takes a decade or more to rebuild once gone, if it can be rebuilt at all. Read the full article at mariothomas.com
Read the article →