a horizontal slice of the framework. like floors in a building. if one floor has no fire exits, the whole building has a problem.
A horizontal governance tier within the framework, numbered L0–L7. Each layer addresses a distinct domain of AI risk and is assigned to a plane. Layers provide separation of concerns: strategic governance (L1) is structurally separated from engineering governance (L5), which is separated from data governance (L3). This separation ensures that the failure or immaturity of one domain does not silently compromise the others.
Why It Matters
Without layers, AI governance is a flat list. A flat list treats board-level risk appetite and GPU infrastructure monitoring as equivalent governance concerns. They are not. They have different owners, different evidence requirements, different testing frequencies, and different regulatory implications. The layered architecture of the AI Control Index exists to prevent this collapse.
The OSI model demonstrated this principle for network architecture: by separating physical, data link, network, transport, session, presentation, and application concerns into distinct layers, engineers could reason about each domain independently while maintaining clear interfaces between them. The AI Control Index applies the same structural logic to governance. L1 (Strategy) defines what the organisation intends. L5 (AI Engineering) defines how that intention is built. The interface between them is explicit, auditable, and testable.
The practical consequence is accountability. When an AI system fails, the layered structure tells you where to look. A hallucination is an L5 problem (model evaluation) with L2 implications (ethics and transparency) and S4 visibility (observability). A data breach is an L3 problem (data governance) with S2 implications (security). Without layers, every incident is everyone’s problem, which means it is no one’s problem.
The Stress Test
You are asked to map your AI governance structure to the EU AI Act’s requirements for a risk management system. You have controls. You have policies. But you cannot demonstrate that your strategy layer (board risk appetite) is connected to your engineering layer (model evaluation gates). The risk appetite says “low tolerance for bias.” The engineering pipeline has no bias evaluation step. The two layers exist independently, with no interface between them.
The regulator does not ask whether you have governance. The regulator asks whether your governance is connected. Layers without interfaces are silos with better labelling.
In the Wild
Wells Fargo’s fake accounts scandal revealed a governance architecture where the strategic layer (board and executive risk appetite) was disconnected from the operational layer (branch sales practices). The board had risk policies. The branches had sales targets that incentivised fraud. The two layers had no effective interface. The Office of the Comptroller of the Currency imposed a $3.7 billion penalty, and the consent order specifically cited failures in the bank’s “risk management programme” and “internal controls.”
The strategy layer said one thing. The operational layer did another. The absence of a functioning interface between layers is itself a governance failure.
The MCAS system that contributed to two fatal crashes of the Boeing 737 MAX was designed at the engineering layer with assumptions about pilot training that were never validated at the operations layer. The engineering layer assumed pilots would recognise and counteract erroneous MCAS inputs. The training layer had minimal documentation of MCAS. The certification layer relied on Boeing’s self-assessment. Each layer operated with assumptions about the others that were never tested.
The layers existed. The interfaces between them did not. 346 people died in the gap between what engineering assumed and what operations knew.
Multiple NHS trusts piloted AI diagnostic tools for radiology and pathology. An NHS England review found that clinical governance structures governed the clinical decision but not the AI pipeline feeding it. The data governance layer (data quality, provenance, representativeness) was not connected to the clinical governance layer (decision accountability, patient safety). When model performance degraded on underrepresented populations, the clinical governance layer had no mechanism to detect it because it did not monitor the data layer.
Clinical governance without data governance is a liability. Each layer must be connected or neither works.
How to Govern It
Layers provide structure. Interfaces between layers provide governance.
Within the AI Control Index, layers are governed through explicit structural requirements:
- Layer Ownership — Each layer has a declared owner responsible for the controls within that tier. L1 is typically owned by the board or C-suite. L5 is typically owned by the AI engineering lead. Ownership cannot be shared between layers.
- Layer Interfaces — The framework defines explicit interfaces between adjacent layers. L1’s risk appetite must flow down to L5’s evaluation gates. L5’s model performance data must flow up to L1’s risk dashboard. These interfaces are auditable.
- Boundary Rules — Where two layers could plausibly own the same control, boundary rules resolve the ambiguity. The framework has 10 boundary rules that prevent governance dead zones.
- Shield Integration — Shields (S1–S5) cut vertically through all layers, ensuring that cross-cutting concerns like security and observability are not confined to a single tier.
- Plane Grouping — Layers are grouped into planes by operational function, enabling organisations to assess governance maturity at the plane level before drilling into individual layers.
When It's Relevant
Layers are relevant whenever an organisation governs more than one AI system, or when a single AI system spans more than one governance domain. In practice, this means every enterprise deployment. A customer-facing chatbot touches strategy (L1: risk appetite for customer interaction), data governance (L3: training data quality), AI engineering (L5: model evaluation), and applications (L4: system-level controls). Without layers, these concerns collapse into a single undifferentiated governance conversation.
Layer governance becomes critical when:
- Multiple teams share responsibility for AI governance
- The organisation must demonstrate separation of concerns to a regulator
- AI systems span strategic, operational, and technical domains
- Governance gaps appear between teams or departments
- Incident investigations need to locate the root cause in a specific domain
Related Terms
References
- [1] ISO/IEC 42001:2023. Information Technology — Artificial Intelligence — Management System. International Organization for Standardization.
- [2] European Parliament and Council (2024) Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act). Official Journal of the European Union.
- [3] NIST (2023) Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology, U.S. Department of Commerce.
- [4] OCC (2023) Consent Order: Wells Fargo Bank, N.A. Office of the Comptroller of the Currency, AA-EC-2023-61.
- [5] U.S. House Committee on Transportation and Infrastructure (2020) Final Committee Report: The Design, Development & Certification of the Boeing 737 MAX. 116th Congress.
- [6] NHS England (2024) Artificial Intelligence Governance Framework for NHS Trusts. National Health Service.
- [7] Zimmermann, H. (1980) ‘OSI Reference Model — The ISO Model of Architecture for Open Systems Interconnection’, IEEE Transactions on Communications, 28(4), pp. 425–432.
AI Control Index v6.0 · Glossary · June 2026 · i-DEPOT 158508 (BOIP) · CC BY-NC-ND 4.0
By Jeroen Janssen, Apparens