EW-AiRM™ Framework: 3-Layer Enterprise AI Governance (2026)

THE FRAMEWORK

EW-AiRM™: Enterprise-wide AI Risk Management, in three deliberate layers.

A practitioner-grade framework grounded in publicly licensed standards, built for the people doing AI risk work inside organisations under live regulatory scrutiny.

Dimension	EW-AiRM™	NIST AI RMF	ISO 42001	COSO ERM	EU AI Act
UNECE-grounded	Yes (ECE/TRADE/486)	No	No	No	No
Primary scope	Enterprise-wide AI Risk Governance framework & approach	AI Risk Management	AI Management System	(Traditional) Enterprise Risk Management	AI Regulation
Target audience	Boards, Risk functions, CTOs, CISO, Risk community	US Federal & Enterprise	Any organisation	CFO, Board, ERM teams	Operators & Providers (EU)
Layers	Strategic, Operational, Resilience + HAiPECR	Govern, Map, Measure, Manage	Plan, Do, Check, Act	Strategy, Performance	Prohibited, High-risk, GPAI
Certifiable	No (Open framework)	No (Voluntary)	Yes (ISO audit)	No (Guidance)	Public, EU-wide Regulation
Open / free	Yes (CC BY 4.0 base)	Yes	Paid standard	Paid guidance	Public regulation
UNESCO-aligned	Yes (HAiPECR)	No	Partial	No	Partial (GPAI)

ARCHITECTURE

Three layers. Each does a job the others cannot.

Most AI risk frameworks pick a level and stay there. EW-AiRM™ deliberately spans three, because the strategic questions cannot be answered with operational tools, and the operational layer cannot anticipate Black Swans.

LAYER 1

Strategic

Six pillars. Necessity, readiness, maturity, tolerance, governance and accountability, through-the-lifecycle monitoring. Sets the conditions any AI deployment has to meet before it goes live.

LAYER 2

Operational

MIT AI Risk Repository (CC BY 4.0): more than 1,000 risks across 7 domains and 24 subdomains. Mapped to 831 controls, primarily security. For control mappings in Primary, Secondary, and Tertiary tiers.

LAYER 3

Resilience

Eight AI Black Swan categories. Includes multi-agent emergence and the quantum cryptographic transition (NIST FIPS 203/204/205). For risks that do not fit the operational taxonomy.

STRATEGIC LAYER

The six pillars.

Each pillar is a question the organisation must answer in writing before a system can be deployed, and a check that must hold across the deployment lifecycle.

Strategic Alignment
+ Necessity Assessment

Is AI the right tool for this problem at all? If a deterministic system would do the job, AI is the wrong answer.

Organisational Readiness

Does the organisation have the people, the controls, and the decision rights to deploy this responsibly?

Technological Maturity

Is the underlying model, vendor, and integration pattern ready for this use case at this risk level?

Risk Tolerance

Has the board signed off on the residual risk, and is the tolerance documented at the right level of granularity?

Governance & Accountability

Who owns the decision, and who owns the incident? Named, not implied.

Through-the-Lifecycle Monitoring
+ Adaptability

What changes trigger a re-review? Who runs the re-review? What is the kill-switch SLA?

ETHICAL OVERLAY

HAIPECR: seven dimensions, mapped to UNESCO.

HAIPECR is the ethical filter that runs across all three layers. Mapped to the UNESCO 2021 Recommendation on the Ethics of AI (10 core principles, not 9). Listed on the OECD AI Policy Observatory since April 2023.

H

Human oversight Named accountability, override tested.

A

Accountability Mapped to UNESCO P5.

i

Inclusivity Embedded prerequisite, not a pillar.

E

Explainability Transparency, fairness, non-discrimination.

C

Conduct UN Universal Conduct Risk Paradigm.

P

Privacy Data protection, safety, security.

R

Resilience Sustainability, intergenerational rights.

UNESCO MAPPING: H → P7 Human Oversight A → P5 Accountability i → P4 Multi-stakeholder Governance + P9 Awareness & Literacy P → P3 Privacy + P2 Safety & Security (+ 2024 UNESCO Neurotech) E → P6 Transparency & Explainability + P10 Fairness C → P1 Proportionality / Do No Harm R → P8 Sustainability

RESILIENCE LAYER

Eight AI Black Swan categories.

For risks that sit outside the operational taxonomy. Each category is a scenario rehearsal, with named owners and a pre-positioned response.

Cat 1 — Foundation model integrity collapse

A frontier model vendor experiences a security or alignment incident compromising every system built on top. Owner: CISO + Chief AI Officer.

Cat 2 — Cascading agent failure across systems

An agent in one part of the organisation triggers downstream errors in dependent systems. Owner: Head of AI Operations.

Cat 3 — Adversarial misuse at scale

External actors industrialise misuse of deployed AI for fraud, impersonation, or social engineering. Owner: Fraud + CISO.

Cat 4 — Regulatory step-change

A jurisdiction reclassifies AI uses you depend on as prohibited or high-risk overnight. Owner: General Counsel + Compliance.

Cat 5 — Data poisoning of public training corpora

Public datasets you fine-tune on are intentionally poisoned, surfacing biased or unsafe outputs. Owner: Head of Data.

Cat 6 — Compute and supply chain disruption

Loss of access to compute, model weights, or critical inference infrastructure. Owner: CTO.

Cat 7 — Multi-agent emergence

Multiple deployed agents interact in ways that produce emergent behaviours none was designed for. Owner: Chief AI Officer + Risk.

Cat 8 — Quantum cryptographic transition

Cryptographic primitives underlying AI system identity, signing, and data integrity become breakable. Migration to NIST FIPS 203 / 204 / 205. Owner: CISO + Architecture.

FLOOR CONDITIONS

Five things every tier must do.

Named accountability for every deployed AI system

HAIPECR run as a pre-deployment filter, documented

Human override capability tested with a 4-hour SLA

Documented board or executive risk acceptance

Incident reporting pathway, with named recipient

WHERE THIS COMES FROM

The framework lineage.

EW-AiRM™ is grounded in publicly licensed standards: the MIT AI Risk Repository (CC BY 4.0), the UNESCO 2021 Recommendation on the Ethics of AI, NIST AI RMF, ISO 31000 (general risk), ISO 42001 (AI risk management systems), and the UNECE (ECE/TRADE/486, 2024). HAIPECR was originated by Prof. Markus Krebsz and has been listed on the OECD AI Policy Observatory since April 2023.

Want the practical version?

The framework is the conceptual picture. The three-tier comparison shows what it looks like in deployment.

See the three tiers →

Read the FAQ