AI at Scale
From Pilot to Production: A Practitioner's Guide to Building AI That Lasts
Written from the field, not from a conference stage. Every principle, framework, and corrective discipline derives from direct operational experience deploying AI across government and infrastructure programmes in Asia and the Middle East.
The pilot trap is not a technology problem.
Global enterprise AI expenditure exceeded $300 billion in 2025. Yet 95% of enterprise AI pilots deliver zero measurable ROI. In 2025, 42% of organisations abandoned most of their AI initiatives — up from 17% in 2024. These numbers represent a systemic failure in how organisations conceptualise, scope, govern, and measure AI deployment.
The failure is not at the model layer. Most organisations have access to capable models, adequate compute, and sufficient capital. They are losing the AI race because they have not solved the harder problems: how to measure AI correctly, how to engineer its economics, how to govern it at scale, and how to build the human capabilities required to produce sustained value rather than isolated demonstrations.
AI at Scale is a practitioner's answer to those harder problems — not a survey of AI capabilities, but a production-grade guide to what makes the difference between a pilot that impresses and a system that lasts.
Most organisations are not losing the AI race because they lack models, compute, or capital. They are losing it because they have not solved the harder problems — how to measure AI correctly, how to engineer its economics, how to govern it at scale, and how to build the human capabilities required to make it produce sustained value rather than isolated demonstrations.From the Foreword
7 parts. 23 chapters. One through-line.
From deployment failure patterns to governance architecture to the human questions that models cannot answer.
The Deployment Reality
- Why Most AI Projects Fail — constraint mapping, four failure patterns, corrective disciplines
- The Wisdom Gap — diagnostic framework, data foundations, measurement baselines
- The Agentic Shift — from generative to autonomous, eight agent types, cascade risk
Economic and Cost Architecture
- Why Large Enterprises Are Measuring AI Incorrectly — model metrics vs operational baselines
- Local LLM vs API — break-even analysis, TCO, tiered routing architecture
- Token Optimisation: An Engineering Discipline — four cost drivers, optimisation stack
Technical Foundations
- Why Your AI Demo Works — But Your Production System Does Not — eight failure modes
- Data Is the Constraint — six quality disciplines, terminology standardisation
- Context Windows, Prompt Engineering, and Reliable AI — right-sizing, code generation governance
Governance Architecture
- A Risk Governance Framework for Scalable AI — four risk pillars, numeric risk scoring
- From Human-in-the-Loop to AI-on-the-Loop — three-layer oversight model, automation bias
- Governed RAG 2.0 — freshness, auditability, version citation, GraphRAG
Transformation at Scale
- AI Is Not an Add-On — operating model redesign, CoE failure modes, Singapore SME case
- The End of Billing Time — outcome-based commercial models, attribution methodology
- The Workforce Equation — five judgment muscles, NAIS 2.0 three tiers
- Use AI to Disrupt Yourself — self-disruption audit, four imperatives
- Outcome-Based Transformation — from adoption to impact, programme accountability
The Practitioner's Playbook
- Where AI Is Safe to Deploy in 2026 — high-confidence domains, Singapore framework
- Critical Tasks Suitable for Agentic AI — five-condition suitability test, TRACE framework
- Voice AI — The 2026 Inflection Point — field operations, production design, ROI evidence
The Human Question
- Navigating a World Where Thinking Is Outsourced — knowledge deflation, what remains valuable
- AI Fatigue — Why Verification Is Harder Than Creation — four design principles
- The AI-Augmented Practitioner — curiosity, domain depth, team design, performance frameworks
Three arguments the book makes.
The pilot trap is structural, not accidental.
A transport ministry in Southeast Asia ran fourteen AI pilots across three years. Each pilot delivered a working demo. Not one reached production. Each year the AI budget increased. Each year the board asked: where are the results? The question that was never asked — "can we operate AI that does this, at scale, reliably, for the next five years?" — is the question this book is written to answer.
Model accuracy is not a business result.
A language model that returns accurate outputs 94% of the time is a capability specification, not a business performance metric. The business results are a reduction in manual processing time, an improvement in decision consistency, a measurable change in cost per transaction. Most organisations measure the wrong layer — and approve scale-up budgets for programmes that have not demonstrated operational return.
Governance is an architecture decision, not a policy document.
The shift from human-in-the-loop to AI-on-the-loop is not a relaxation of governance — it is a redesign of it. When AI becomes the primary checker, the human role shifts from transaction-level approval to governance of thresholds, escalation logic, and exception handling. This redesign, done correctly, is more robust than manual review at scale — and it is what regulators in Singapore and Europe are moving toward.
Eight-Dimension AI Readiness Assessment
Production Readiness Gate — 12-Point Checklist
Regulatory Reference: Asia and the Middle East
Glossary of Production AI Terms
Be notified when the book is available.
Send a message via the contact form with the subject "AI at Scale" and you will be notified directly when the book is released.
Register InterestAlso available
The frameworks in this book are the basis for Terence's keynotes and panel contributions on AI governance, deployment, and measurement.
View speaking topics →