AI governance does not become real when a board approves a policy. It becomes real when the codebase enforces it. Most organisations now have AI principles, risk frameworks, steering committees, and approval processes. Yet the systems that matter most are still built one pull request at a time — by engineers deciding how data is retrieved, how agents call tools, how permissions are scoped, how logs are written, and what happens when something goes wrong. Governing Intelligence: Technology is a practical guide for engineering, platform, security, and technology leadership teams responsible for turning AI governance from policy into architecture. This book is not about writing better prompts. It is not a theoretical ethics text. It is an implementation-level playbook for building AI systems that are secure, observable, auditable, and governable in production. Across thirteen focused chapters, the book shows how to translate governance commitments into technical controls: policy-enforcement gateways, scoped identities, access-controlled retrieval, vendor due diligence, prompt-injection defences, agent permissioning, evaluation pipelines, audit logs, incident runbooks, CI/CD gates, and internal AI platforms that make the governed path the default path. The central argument is simple: a guardrail that lives only in a policy document is not a guardrail — it is a hope. A real guardrail lives in the platform, the gateway, the schema, the permission broker, the deployment pipeline, or the audit log. Written for CIOs, CTOs, platform leaders, security engineers, ML engineers, SREs, architects, and technical governance teams, this volume assumes technical fluency and focuses on the practical decisions organisations face as AI systems move from pilots to production. Readers will learn how to: Map AI governance obligations to engineering ownership Design a reference architecture for governed AI systems Vet AI vendors and models before production use Protect sensitive data across prompts, retrieval, outputs, and logs Scope identities and tool permissions for models, agents, and workflows Defend against direct and indirect prompt injection Classify agent actions by reversibility, impact, and approval requirements Build golden test sets, red-team routines, and AI-specific evaluation gates Monitor drift, cost, usage, anomalous behaviour, and audit evidence Prepare incident-response runbooks for AI-specific failures Treat prompts, model versions, and tool manifests as deployable artefacts Build internal AI platforms where guardrails are inherited by default This is the third volume in the Governing Intelligence series. Nations focuses on public policy and regulation. Boards focuses on directors and enterprise accountability. Technology translates those expectations into the systems engineering practices that make them enforceable. Finance completes the series by examining AI risk, value, reporting, audit, and capital allocation. If your organisation is moving AI from experimentation into real workflows, this book is for the people who must make sure the system does not merely work — but works safely, accountably, and with evidence. Because engineers do not govern AI through policy. Engineers govern AI through architecture.
