// Part Two · Chapter 4

Meet the AI Plumber

The Moment of Realization

In 2015, I was working with a major European bank. The technical team built a sophisticated ML model. The demo dazzled executives. Six months later it was a disaster — GDPR violations, no audit trail, regulators asking questions. The problem wasn't the AI. It was the plumbing.

What Is an AI Plumber?

An AI Plumber solves the ugly problems — infrastructure challenges, integration headaches, governance requirements that make AI work in the real world. We're not engineers who build impressive models; we're engineers who make models work in production. We design policy layers, build audit trails, implement controls, and create governance frameworks.

The AI Plumber Manifesto

Infrastructure beats algorithms — the best algorithm is worthless without infrastructure.
Governance is not a bottleneck — it's the foundation for success.
Compliance creates competitive moats — build advantages competitors can't replicate.
Production is harder than prototypes — the gap is where projects die.
Simplicity beats complexity — start simple.
Humans stay in the loop — AI augments, it doesn't replace.
Measure what matters — ROI matters more than accuracy.

The 6 Pipes

The framework is six pipes — the complete set of layers a regulated enterprise needs before AI enters production. The first four are the Non-Negotiables: the moment AI becomes agentic, they are mandatory. The other two make the system survive a real legacy enterprise and improve over time.

The 4 Non-Negotiables

Constrained Identities — every agent runs as a scoped, least-privilege service account; none inherit human privileges.
Attributable Actions — every decision logged with input, reasoning trace, and output; reversible and traceable in under 30 seconds.
Human-in-the-Loop Gates — high-stakes actions mechanically pause for a human authorization token.
Kill Threshold Monitoring — continuous telemetry; automatic suspension on breach.

The 2 Enterprise-Fit Pipes

Legacy Insulation — an API gateway wraps legacy systems so 40-year-old mainframes talk to agents without the core being touched.
Learning Loop — every production incident updates the architecture via ADRs; the system evolves, it isn't patched.

Six pipes; four are non-negotiable for any agentic deployment; the other two make it survive a real enterprise.