PDF · HTMLWhitepaper · ~12 pp
Your AI Pipeline Has No Kill Switch
The full thesis: why enterprise AI fails at the governance layer, and the four non-negotiables that decide whether a system ships.
The method, readable and indexed.
Every whitepaper, briefing and checklist is published here as crawlable, index-friendly HTML — not locked inside a PDF a search engine can't read. Skim the page, pull the source PDF, or hand any complex topic straight to an AI: every section carries an Ask-AI action that opens ChatGPT, Claude or Perplexity with the question pre-written.
HTML-first
PDF source attached
Ask-AI on every topic
FAQ structured data
The library
Read it indexed. Or take the PDF.
Four canonical artifacts. Each one is rendered in full as HTML below, with the original PDF attached for offline and board distribution.
PDF · HTMLBriefing · 4 pp
The 6 Pipes Briefing
A board-ready extract: the four mandatory pipes, what each one prevents, and the failure mode that triggers it — one page apiece.
PDF · HTMLChecklist · runbook
The AI Governance Checklist
The 5-Phase Runbook as an operational artifact — Assess → Build → Document → Test → Monitor — line items a delivery team can sign off.
BOOK~60–80 pp · 6 chapters
The AI Plumber — the book
The method long-form, each chapter anchored to one verified regulated deployment. In final consolidation.
Why this page exists
PDFs are invisible. This isn't.
A governance method buried in a downloadable file is unreachable to the people searching for it — and to the assistants they now ask first.
01
Crawler-readable
Every artifact is rendered as semantic HTML with headings, canonical URLs and FAQ structured data — so search engines index the substance, not a download button.
02
Answer-engine ready
The same text is what ChatGPT, Claude, Perplexity and Google's AI surfaces cite. The method shows up where the question is actually being asked.
03
Ask it directly
Every complex topic carries an Ask-AI bar. One tap hands the exact question to your assistant of choice — pre-written, source-aware, ready to interrogate.
Your AI Pipeline Has No Kill Switch
Your AI pipeline has no kill switch. When it fails in production — not if, when — you have no off switch and no audit trail for the regulator. You have a brilliant AI agent without an emergency brake. You have a model generating outputs at scale with no traceable decision rationale. You have an autonomous system making decisions that would normally require sign-off from three layers of management, and there's no way to explain to a regulator why it made the choice it did.
This is the reality facing most enterprises today. They have rushed to deploy AI systems without building the infrastructure around them that makes those systems safe, auditable, and compliant. They have invested millions in models, data pipelines, and proof-of-concepts — and almost nothing in governance, audit trails, and kill switches.
The result is a portfolio of AI systems that work beautifully in controlled environments and become liabilities in production.
Every week, somewhere in a corporate boardroom, a conversation like this takes place: "We need to do something with AI. Our competitors are doing it. The board is asking about it. Why don't we have an AI strategy yet?"
And then the scramble begins. Teams are assembled. Vendors are evaluated. Proof-of-concept pilots are launched. Six months later, the pilots have failed. The vendors were oversold. The technical team is frustrated. The compliance officer has flagged seventeen regulatory concerns. The budget is exhausted. And the word "AI" has become politically toxic in the C-suite.
The failures are almost never because of the AI itself. The models work. The technology is impressive. The potential is real. The failures happen because nobody built the system around the model. Nobody designed the governance framework. Nobody created the audit trail. Nobody implemented the kill switch.
After twenty years of building enterprise systems — from Kapaza (the Belgian classifieds platform acquired by Schibsted) to leading regulated-AI delivery at DevGap across Europe, the Middle East, and India — I've learned one fundamental truth: the organizations that succeed with AI aren't the ones with the best models. They're the ones with the best plumbing.
Topics · indexed & queryable
The complex matters, in brief.
Each is a crawlable summary of a deeper section. Read the gist, follow it into the framework, or interrogate it with an AI — the question is already written.
Constrained Identities
Agents run least-privilege service accounts with zero default write permissions. Destructive actions are blocked at the keyring, not by instruction — so a misbehaving agent can't reach what it was never granted. Anchored in production at Najm Insurance under SAMA.
Attributable Actions
Every decision, query and citation is logged on a read-only ledger. Any action is traceable and reversible in under 30 seconds — an audit trail built to survive a regulator, not satisfy a dashboard. Anchored at Govt of India / NMML.
Human-in-the-Loop Gates
High-stakes actions pause and wait for manual authorization. The agent cannot execute financial or legal tasks alone — gates that satisfy the EU AI Act without killing velocity. Anchored at De Lijn.
Kill Threshold Monitoring
Continuous telemetry watches speed, spend and error volume. On breach the agent auto-suspends — before damage, not after. Anchored in a LangGraph multi-agent system replacing a 200-person operation.
The 70% Pilot-Death Problem
Most enterprise AI pilots die before production — at the governance layer, not the model layer. The models work; nobody built the system around them. This is the gap the four pillars close.
EU AI Act · Annex III
High-risk (Annex III) obligations apply from 2 August 2026 — risk management, logging, human oversight and transparency for in-scope systems. (Operative date; subject to pending Digital Omnibus revisions — verify current status.)
Read it, then install it
Kleiber — the four pillars, enforced.
aiplumber.dev is the consultancy and the method. Kleiber is the open-source layer that enforces the four pillars in your agent runtime — the installable proof.
What it is
A runtime layer that wraps your agents in an Agent Accountability Envelope: least-privilege identities, a read-only action ledger, authorization gates and kill thresholds — the controls from the method, as code you can run.
Read the framework on this site, then drop the layer into your runtime. The method and the proof, in one place.
The shape of it
from kleiber import Envelope, Gate, KillThreshold
agent = Envelope(
identity = "claims.read_only", # least privilege
ledger = "append-only", # attributable
gates = [Gate("payout")], # human sign-off
kill_on = KillThreshold(spend="€500/h"),
)
Illustrative API shape — the four pillars expressed as runtime primitives.
Regulatory currency: EU AI Act high-risk (Annex III) obligations apply from 2 August 2026 (operative date; subject to pending Digital Omnibus revisions — verify current status before relying on it). Note: PDF downloads attach on deploy; on this preview the indexed HTML above is the canonical source.
Tool · your numbers
What is ungoverned AI costing you?
Three inputs, one transparent estimate of what governance recovers in a year — wasted model spend, engineering time, and the incident provision you're quietly carrying. Move the sliders; the maths is shown, not hidden.
AI agents / pipelines in production
Monthly AI / API spend
Eng. hours / week on manual audit & incident response
Assumes a €110/h blended engineering rate, 30% of model spend recoverable through caching & deduplication, 60% of manual governance hours eliminated, and a conservative €4,500/agent annual incident provision. Real numbers get set on a call.