ADG | Adopt · Defend · Govern

Why ADG

Enterprise AI is shipping into production faster than governance can absorb. Engineering teams are deploying agents that act on tools, consume retrieved context, run multi-step plans, and chain to other agents — often before the people accountable for risk have seen the system, let alone approved it. Boards are asking questions control departments cannot answer. Regulations are taking effect—the EU AI Act first, with others to follow. And the standards every organization has invested in — ISO/IEC 42001, NIST AI RMF, the EU AI Act itself — tell you what to think about, but not who decides, who builds, who breaks, who signs, or what evidence belongs in front of a regulator on Monday morning.

The architecture is three pillars — ADOPT (execute and deliver), DEFEND (secure and validate), and GOVERN (oversee, assure, decide) — that scale unchanged from the board to the engineer. The same three words frame a board paper, organize a release gate, label a red-team report, and structure an incident playbook. Beneath those pillars sit nine governance surfaces (the points where engineers instrument and attackers act); 12 Minimum Controls (MC-1 through MC-12), each with a named evidence artifact your auditor can hold; nine deployment overlays for the patterns AI actually ships in today — agentic orchestration, tools and Model Context Protocol (MCP), multi-agent interoperability, multi-modal and composite stacks, long-context architectures; three autonomy tiers (human in the loop [HITL], human on the loop [HOTL], human out of the loop [HOOTL]) tied directly to the controls; and a four-phase roadmap that takes an organization from inventory to continuous assurance in 24 months.

ADG is built for alignment, not competition. Every Minimum Control maps cleanly into NIST AI RMF functions (GOVERN, MAP, MEASURE, MANAGE) and ISO/IEC 42001 Annex A — the framework's §11.2 crosswalk does it explicitly. MC-1 produces the AI inventory ISO 42001 A.6 demands and EU AI Act Art. 49 will register. MC-9 produces the post-market log Art. 73 inspectors will request. MC-11 produces the bias evidence Art. 10 requires. Organizations pursuing ISO/IEC 42001 certification or EU AI Act conformity use ADG to operationalize what those frameworks describe abstractly. The work is the same; ADG removes the translation layer.

ADG also goes where the standards do not. Agentic systems, MCP-connected tools, multi-agent interoperability, long-context architectures, and composite multi-modal stacks each carry an additive deployment overlay that composes with the same twelve Minimum Controls. Move from one copilot to a fleet of agents without rewriting a single control.

One element of ADG sits where peer frameworks generally do not: the AI Governance Council. Every framework tells you to govern AI. ADG tells you who breaks the tie when delivery and safety disagree — the Council resolves ADOPT-DEFEND tension, sets go/no-go thresholds, owns the exception register, and ties decision rights to MC-3 (Separation of Duties) and MC-10 (Periodic Governance Review). The mechanism, not the abstract requirement, is what is rarely named in peer frameworks — and it is where most production AI risk actually lives.

ADG is built for boards, CISOs and AI red teams, platform and AI engineers, risk and compliance, and procurement. Each audience reads a different path; the artifacts are the same. The framework was forged with senior AI, security, and governance leaders running production AI inside top Fortune 500, Fortune Global 500, and Big Four firms — including Salesforce, Microsoft, Citi, JPMorgan Chase, NTT DATA, KPMG, ServiceNow, BNP Paribas, Prudential, GE Healthcare, BASF, and Jio — the practitioners who own the failure modes this framework governs.

Who's reading this?

ADG is dense by design. Pick the role that matches you and jump to the path written for it. Each path is 4–5 sections out of 13 — you don't have to read everything.

1.Background

1.1 Origin and Positioning

ADG began with a mission at EC-Council: the AI governance space was fragmenting into point standards, vendor playbooks, and academic frameworks — none of which gave enterprise boards and engineering teams a single, practitioner-tested operating model for deploying AI safely at scale. EC-Council set out to build that missing operating model: a framework built around three enduring functions—ADOPT, DEFEND, and GOVERN—that scales from the board to the engineer.

To avoid building in a vacuum, EC-Council convened an Advisory Board of practitioners from mature enterprises that have actually deployed AI in production. The initial draft was circulated to senior AI, security, and governance leaders across global organizations in financial services, technology, manufacturing, healthcare, telecommunications, energy, and consulting — covering both regulated and less-regulated sectors, and spanning North America, Europe, and Asia-Pacific.

Framework Leadership

Jay Bavisi

Chairman & CEO

EC-Council Group

Karthik S.

Framework Architect & Lead Author · Practice Head, SecureAI

EC-Council Global Services

Thank you to the Advisory Board

The practitioners below gave their time, scrutiny, and hard-won experience to review ADG. Their comments, corrections, and counterpoints are the reason this framework is field-tested rather than aspirational.

Adam Spearing

VP of AI GTM EMEA

ServiceNow

Andrei Son

Head of AI Transformation

AUMOVIO

Anish Mitra

Director

KPMG

Anita Lacea

Head of AI Transformation, Azure Hardware Infrastructure

Microsoft

Dinesh Bhogle

Head of AI/ML Platform

Black & Veatch

Dr. Sayed Peerzade

EVP, Cloud, AI & Government Initiatives

Jio

Edoardo Tealdi

Executive Head of AI Transformation, Business Engagement and Growth Units

NTT DATA, Inc.

Kathy Baxter

VP/Principal Architect, Responsible AI & Tech

Salesforce

Lewis V. Adams

VP, Enterprise AI & Capital Productivity Transformation

Citi

Lily Rachmawati

Director, Head of Applied AI

BNP Paribas

Malik Hussain

AI Enablement Lead, Data & AI Academy

BASF

Mark Ritcey

VP, AI & Automation Delivery

Latent Bridge

Naveen Upadhyay

VP, AI/ML Product Management – Machine Learning & Intelligence Operations

JPMorgan Chase & Co.

Oscar Jarabo

Global Head of AI Product & Strategy

TKE

Pavan Kristipati

Head of AI Engineering & Transformation | Enterprise AI Adoption, Governance & Platform

Huntington Bank

Raghunandan Mishra

AI & Data Engineering Leader

Raji Bhimireddy

VP Cloud, AI, Architecture, FinOps & Business Value

Prudential

Sanjoy K. Saha

Head of AI Portfolio and Governance & Chief of Staff CDAO

GE Healthcare

ShanShan Pa

Global Head of AI & Data Governance

GlobalLogic

Sophia Katrenko

VP of AI/ML

EcoVadis

Sruthi Pakanati

Head of AI & Data Transformation, National Quality & Risk

Deloitte Australia

Sudarson Roy Pratihar

Founder and Principal

A2IQ

Yashwinder Chhikara

Sr. VP, AI, Analytics, and Product Management

iSON Xperiences

This version is the result — the iterative output of that collaboration. ADG is not a one-and-done document. It is a living framework, designed to be enhanced as AI deployment patterns evolve and as new practitioners join the conversation.

1.2 Framework Purpose: The RE³ Trust Model

ADG is purpose-built to operationalize four non-negotiable properties of trustworthy enterprise AI — captured in the RE³ Trust Model: Responsible, Ethical, Efficient, Explainable AI.

1.3 Core Thesis

1.4 Scope and System Coverage

ADG governs all AI systems as defined by the OECD AI Policy Observatory and the EU AI Act, including LLMs and transformer-based generative AI, diffusion models, multi-modal systems, classical machine learning, composite AI systems that orchestrate multiple model types, and emerging architectures such as world models and neuromorphic systems. The governance surfaces, overlays, and controls in this framework are architecture-neutral by design — not limited to LLM-centric systems.

1.5 Distinction: Standards vs. Regulations

2.Executive Summary

Key framework components

Layer-by-Layer Reference

Concrete definitions and one example per element — the quick-reference key for the rest of the document.

3 Pillars

9 Governance Surfaces

12 Minimum Controls

4 Harm Classes

4 Industry Threat Catalogs

ADG does not invent its own threat list — it consumes the four canonical industry catalogs as input and maps each threat into the Minimum Control Set and the deployment overlays.

9 Deployment Pattern Overlays

3 Autonomy Tiers

4-Phase Roadmap

3.Document Usage

This section provides guidance on how to read, apply, and implement the ADG framework.

3.1 Target Audience

3.2 Foundational Principles

ADG is built on nine foundational principles — durable standards that apply across industries, deployment patterns, and regulatory regimes.

4.The ADG Triad and Operating Model

4.4 Operating Model

ADG uses a simple rule: Adopt builds and operates, Defend breaks and protects, Govern authorizes and oversees.

5.Governance Surfaces

Where ADG meets MITRE and OWASP

The nine surfaces above describe what must be governed. The four industry threat catalogs describe what attackers do. ADG consumes them as input — every surface inherits its detection rules, eval scenarios, and runtime guardrails from these sources, then maps the resulting controls to NIST AI RMF and ISO/IEC 42001.

6.Harm Taxonomy and Responsible AI Integration

6.1 Harm Classification

The four harm classes plotted by detection difficulty and impact scope. Each carries the threats it covers, the pillars that own it, and the Minimum Controls that detect it.

6.2 Responsible AI Principles

Every governance surface must account for seven responsible AI dimensions, embedded across all three pillars rather than appended as a separate track.

Integration rule: Responsible AI is not a separate track. It is embedded into every ADG pillar — ADOPT, DEFEND, and GOVERN — and must be reflected in all certification curricula, operating artifacts, and governance reviews.

7.Shared Responsibility Model

7.1 Deployment Responsibility Classes

7.2 Vendor AI Due Diligence Requirements

1. Require vendor disclosure of model provenance, training data policies, alignment methods, and known limitations
2. Establish contractual AI governance clauses covering incident notification, data handling, model change management, and liability allocation
3. Conduct independent evaluation of vendor-provided safety controls rather than relying solely on vendor claims
4. Maintain consumer-side telemetry and monitoring regardless of vendor monitoring capabilities
5. Include AI governance questions in procurement processes and vendor risk assessments
6. Define rollback and exit strategies for vendor AI dependencies

8.People, Process, Technology, and Data

8.1 People Layer

8.2 Process Layer — the Nine-Step Backbone

8.3 Technology Layer

1. Model gateway and routing layer
2. Retrieval, memory, and context management layer
3. Tool registry and MCP trust layer
4. Policy engine, guardrail layer, and harm detection layer
5. Runtime telemetry, replay, and fairness measurement layer
6. Training, tuning, and evaluation pipeline layer
7. Input validation and sanitization layer (covering prompts, RAG inputs, and multi-modal inputs)

8.4 Data Layer

Traditional enterprise data architecture separates data into distinct tiers — structured databases, warehouses, lakes, file stores, and APIs. AI systems fundamentally disrupt this separation: an agent consuming enterprise data through retrieval pipelines, MCP connections, or tool invocations does not distinguish a SQL record from a Slack thread — all of it collapses into a single text-token consumption surface.

Access control at the storage layer alone is no longer sufficient. Organizations must govern the full pipeline from source data through retrieval, embedding, context assembly, and AI consumption — across four dimensions:

9.Lifecycle Governance

9.1 Post-Deployment Continuous Governance

1. Configuration drift detection: verify that the system in production matches what was approved (models, prompts, tools, context sources)
2. Usage authorization monitoring: confirm that approved users are using the system for approved purposes within approved boundaries
3. Feature and capability change governance: new model versions, prompt updates, tool additions, and retrieval source changes post-deployment must go through change control
4. Continuous automated evaluation: scheduled adversarial testing, fairness benchmarking, and accuracy regression testing on production systems
5. Ongoing governance of the product roadmap: feature additions to deployed AI systems require re-evaluation against the original risk classification and approval conditions

10.Deployment Pattern Overlays

Overlays are additive — select all that apply.

10.8 Multi-Agent and Agent Interoperability

Multi-agent systems create emergent governance challenges that single-agent controls do not address:

• Agent-to-agent trust boundaries: each agent interaction must enforce explicit trust verification; no agent should implicitly trust another agent's outputs or instructions.
• Value alignment verification: agents with different guardrail configurations, different training, or different vendors may have incompatible safety policies. Cross-agent interactions must be tested for value alignment conflicts.
• Collusion and deception detection: monitoring for patterns where agents coordinate to bypass controls, produce misleading outputs, or exploit gaps between their respective guardrails
• Cascading failure and amplification risk: a failure or bias in one agent can be amplified through a chain of dependent agents. Circuit breakers must exist at agent-to-agent boundaries.
• Cross-agent audit trails: every agent-to-agent interaction must be logged with sufficient detail for forensic replay.
• Sycophancy degradation prevention: in agent-to-agent interactions, sycophantic behavior can devolve into harmful feedback loops. Detection and interruption mechanisms are required.
• Interoperability governance: when agents from different organizations or vendors interact, a shared governance protocol must define minimum safety, logging, and accountability requirements.

11.Controls, Artifacts, and Measurement

11.3 Required Operating Artifacts

11.4 Measurement and Evidence Framework

ADG requires measurable governance. Three measurement tiers ensure controls are not just documented but demonstrably effective.

12.Regulatory Alignment and Roadmap

12.1 Regulatory and Standards Alignment

Note: this mapping is indicative, not exhaustive. Organizations must conduct their own regulatory compliance assessment.

12.2 Implementation Roadmap

13.Workforce Capability

13.1 The AI Skills Gap

The defining constraint on enterprise AI is not the technology — it is the workforce. AI is scaling into production faster than the people who must run, secure, and govern it can be trained, credentialed, and deployed. IDC (2024) estimated that IT skills shortages could cost organizations USD 5.5 trillion in delays, quality issues, and lost competitiveness. Without workforce capability, every control in this framework remains theoretical.

13.2 The Capability Bridge

EC-Council's answer is the Enterprise AI Credential Suite — a portfolio of role-aligned credentials mapped directly to the ADG framework, structured as a four-part capability ladder:

1. Literacy baseline — a shared AI vocabulary for every practitioner, technical or not, so governance conversations can happen at enterprise speed. Anchored by the Artificial Intelligence Essentials (AI|E) program.
2. Role-aligned capability — three flagship credentials mapped 1:1 to the ADG pillars: C|AIPM for Adopt, C|OASP for Defend, C|RAGE for Govern. Each credential certifies execution-grade competence for the people accountable for that pillar.
3. Extended core skills — CEH v13 restructured to integrate AI modules across all five phases of ethical hacking, extending existing security practitioners into the AI attack surface without a role change.
4. Operational readiness — each credential crosswalks to the framework's Minimum Control Set (MC-1 through MC-12), turning certification into deployable evidence that auditors, regulators, and boards can accept.

13.3 Credential Portfolio

Each links to its official EC-Council program page for the syllabus, prerequisites, and enrollment.

Appendix A: Definitions

Executive Conclusion

This version turns the original ADG concept into a full enterprise AI security and responsible AI governance framework. It keeps the simplicity of ADOPT, DEFEND, and GOVERN, but adds the structure required to govern contemporary AI systems — across people, process, technology, deployment pattern, lifecycle, harm class, and regulatory environment.

This version was forged with senior AI, security, and governance leaders running production AI inside Fortune 500, Fortune Global 500, and Big Four firms — practitioners working from design through implementation across regulated and less-regulated sectors. It addresses eight major gap clusters identified through systematic review, and extends the framework to cover multi-agent systems, multi-modal architectures, shared responsibility, responsible AI, and measurable governance.

From a framework standpoint, this version is suitable as the basis for:

• Consulting assessments and maturity reviews
• Enterprise AI governance programs
• Certification architecture and training design across all three ADG pillars
• Board-facing AI security, responsible AI, and assurance discussions
• Regulatory preparation and compliance gap analysis
• Vendor AI due diligence and procurement governance