Skip to content

The governance stack

Six layers of AI governance Most orgs fund the last one first

The layers are real. The order they get built in usually is not. Dependencies run top-down — budgets run bottom-up. That inversion is the most common failure pattern in enterprise AI governance.

Build firstSequence flexibleBuild last
Layers
Six
Build first
Inventory
Build last
Compliance
Failures concentrate
Top & bottom
The stack

Each layer assumes the one above it exists

↓ dependency·funding ↑
Layer 1AI InventoryDo you know what’s running?Build first
Shadow AI DetectionSystem ClassificationRisk TieringOwnership AssignmentModel Registry

The layer almost no one funds — and the one everything else depends on. Tools spread through SaaS features faster than any registry tracks them.

Layer 2Data FoundationCan you trust what feeds it?
Source TrackingLineage MappingQuality ValidationFreshness MonitoringData Bias Screening
Layer 3Data Security & AccessWho can touch it?
EncryptionAnonymizationRole-Based AccessLeast PrivilegeKey Management
⇅ layers 3 + 4 can run in parallel — reasonable teams sequence them either way
Layer 4Model AssuranceDoes it behave?
Model CardsPerformance BenchmarksFairness TestingRed-TeamingDrift Detection
Layer 5Human OversightWho answers for it?
Decision ReviewEscalation PathsOverride AuthorityOutput ValidationAccountability Mapping
Layer 6Compliance & AuditCan you prove it?Build last
EU AI Act MappingGDPR AlignmentPolicy EnforcementIncident ReportingAudit Trails

Compliance is not an input. It is an output. An audit trail only proves something if the layers above it produced something worth recording. Built first, it is a binder with nothing behind it.

Run one test this week

Ask for a complete inventory of every AI system operating in the company right now.

If no one can produce it, that is your real governance gap — everything below it is resting on a layer that does not exist yet.

appliedfutures.ioCopenhagen Claude Community · Practitioner-ledPractice over prescription