AI Governance Framework · Version 2.0

Cognitive Systems Management 2.0
(CSM)

Govern AI through explicit decisions, evidence, handoffs and reassessment, not disconnected checklists.

CSM 2.0 preserves the original Enterprise, Project, Code and UX domains and adds an operational specification for how governance decisions are evaluated, evidenced, transferred between teams and reopened when conditions change.

Four governance domains. Six execution functions. Explicit decisions and evidence. Deterministic evaluation applies to objective governance rules; interpretive decisions require explicit human review.

Cognitive Systems Management (CSM) is a four-domain AI governance methodology spanning Enterprise, Project, Code and UX. It is designed to connect organizational governance, initiative execution, technical development and human interaction so governance decisions remain visible across the lifecycle of an AI-enabled system.

CSM is a four-domain governance methodology designed to connect organizational accountability, AI initiative execution, AI-assisted development and human interaction rather than managing those responsibilities as isolated problems.

Original 2025 Framework: Cognitive System Management: A Framework for Enterprise AI Project Governance” by Subodh KC, AI Governance on Medium, August 29, 2025. CSM 2.0 is based on and evolved from this original publication.

Read the original article

The Problem

Governance Fragments at Handoffs

Governance can become fragmented at organizational handoffs. Policy exists but project criteria do not reflect it. Project approval exists but implementation differs from the approved assumptions. Technical controls exist but users do not understand appropriate reliance. Operational feedback never returns to policy or project decisions. CSM makes those handoffs visible.

Policy exists but project criteria do not reflect it
Project approval exists but implementation differs from approved assumptions
Technical controls exist but users do not understand appropriate reliance
Operational feedback never returns to policy or project decisions

CSM makes those handoffs visible.

The framework’s value is not simply that it has four domains. Its value is maintaining governance context across them.

The Four Domains

Where Governance Responsibilities Operate

CSM-Enterprise

Establish the decision context before individual projects improvise it.

Who has authority, who owns the outcome and risk, and what organizational boundaries apply?

Original components:

Policy Framework

Risk Assessment

Data Stewardship

Strategic Mandate

CSM-Project

Make scaling an explicit governance decision.

What evidence should justify continuing, changing, scaling or stopping an AI initiative?

Original components:

Business Case Definition

Controlled Testing

Scale Decision Framework

Playbook Documentation

CSM-Code

AI-generated code is still organizationally accountable code.

How should software engineering governance change when AI contributes to implementation?

Original components:

Development Standards

Security Protocols

Human Oversight

Traceability Logging

CSM-UX

Governance reaches the people relying on the system.

What do humans need to understand, supervise, challenge and appropriately use AI-supported outcomes?

Original components:

Impact Analysis

Explainability Design

Capability Development

Adoption Measurement

Domain 1 of 4

CSM-Enterprise

Establish the decision context before individual projects improvise it.

Core question: Who has authority, who owns the outcome and risk, and what organizational boundaries apply?

Problem this domain addresses

Without an organizational decision context, each AI initiative must reconstruct governance independently. Projects define their own risk tolerance, data expectations and accountability structures, leading to inconsistent oversight and fragmented policy enforcement.

Verified original components

Policy Framework

AI ethics standards and organizational policies that account for system behavior which may differ from conventional software.

Risk Assessment

Risk evaluation that accounts for system behavior which may change over time, including model updates, provider changes, or data drift.

Data Stewardship

Governance for datasets that influence ongoing AI behavior, including training data, retrieval sources and operational data.

Strategic Mandate

Organizational authority and strategic alignment that defines why AI systems are being deployed and what boundaries apply.

Current implementation examples

Current implementation interpretation. These examples were not explicitly present in the original 2025 article.

AI use policy

Document defining permitted and prohibited AI uses.

System inventory

Catalog of AI systems, owners and risk classifications.

Accountable owner assignment

Named individual responsible for each AI system’s outcomes.

Risk classification

Tiered risk levels assigned based on use case and impact.

Data-owner/steward assignment

Named owners for data used in AI systems.

Prohibited-use boundaries

Explicit list of uses the organization will not pursue.

Risk-acceptance authority

Defined escalation path for accepting residual risk.

Strategic alignment record

Documentation linking AI initiatives to organizational objectives.

Vendor governance agreement

Contractual terms with AI providers covering data usage, model disclosure, liability and exit conditions.

Provider data-processing review

Assessment of how third-party AI providers handle organizational data, including training-on-input prohibitions.

Model card / system card requirement

Requirement that vendors provide documented model capabilities, limitations and training data provenance.

Incident response authority

Named role with authority to declare an AI incident, trigger shutdown and initiate escalation.

Incident escalation protocol

Defined escalation path from operational teams to governance leadership for AI-related incidents.

Breach notification procedure

Process for determining whether an AI incident triggers regulatory or contractual breach notification obligations.

Example artifacts

AI use policyAI system inventoryRisk classification matrixData stewardship assignmentProhibited-use listRisk-acceptance recordVendor governance agreementProvider data-processing assessmentIncident response planIncident escalation matrix

Typical roles involved

Executive sponsorAI governance leadRisk officerData stewardLegal/privacy counselCompliance leadVendor managerIncident response lead

Handoffs to other CSM domains

CSM-EnterpriseCSM-Project: Policy and risk boundaries become project requirements.
CSM-EnterpriseCSM-Code: Security incidents require code-level investigation and remediation.

Common failure modes

  • Policy exists but is not communicated to project teams
  • Risk classification is assigned but never revisited
  • Data stewardship is nominal without actual enforcement
  • Strategic mandate is implicit rather than documented
  • Vendor agreements lack AI-specific data usage and exit terms
  • No one has authority to declare an incident or trigger shutdown
  • Incident escalation path is undefined or bypasses governance leadership

Reassessment triggers

  • New regulatory or contractual obligations
  • Organizational restructuring
  • New AI use cases outside existing policy
  • Incidents from other AI initiatives
  • Vendor changes terms of service, data handling or model behavior
  • Vendor announces model deprecation or end-of-life
  • AI-related incident requiring escalation beyond operational teams

Intended organizational value

Provides projects and technical teams with an organizational decision context rather than requiring each initiative to reconstruct governance independently.

Limitations / proportionality

Enterprise governance is only effective if downstream domains actually receive and apply the decision context. Policy without communication does not create governance.

Domain 2 of 4

CSM-Project

Make scaling an explicit governance decision.

Core question: What evidence should justify continuing, changing, scaling or stopping an AI initiative?

Problem this domain addresses

AI initiatives may require learning during implementation. Performance, capability or success criteria may not be fully known at the start. Without explicit decision boundaries, experimentation continues indefinitely without governance gates.

Verified original components

Business Case Definition

What problem or value is being tested. The business case defines the hypothesis an AI initiative is evaluating.

Controlled Testing

What must be learned before scale. Testing designed to answer specific governance and performance questions.

Scale Decision Framework

What evidence justifies broader commitment. Defined criteria for deciding whether to proceed, change or stop.

Playbook Documentation

What decisions and learning need to survive beyond the pilot team. Documentation that transfers knowledge to operational owners.

Current implementation examples

Current implementation interpretation. These examples were not explicitly present in the original 2025 article.

Use-case charter

Document defining the problem, scope and success criteria.

Success/failure criteria

Explicit thresholds for what constitutes a valid outcome.

Pilot evaluation plan

Structured approach to testing before broader deployment.

Acceptance thresholds

Defined quality, safety and performance criteria for scale.

Decision record

Documented go/no-go decision with rationale.

Scale/no-scale decision

Explicit governance gate before operational commitment.

Lessons learned

Findings from the pilot that inform future initiatives.

Deployment playbook

Operational instructions derived from pilot experience.

Example artifacts

Use-case charterPilot evaluation planAcceptance criteriaScale decision recordLessons-learned documentDeployment playbook

Typical roles involved

Project sponsorProduct managerTechnical leadRisk assessorBusiness owner

Handoffs to other CSM domains

CSM-EnterpriseCSM-Project: Policy and risk boundaries become project requirements.
CSM-ProjectCSM-Code: Acceptance criteria and approved assumptions become implementation constraints.
CSM-ProjectCSM-Enterprise: Incidents, lessons and newly discovered risks may require policy or risk updates.

Common failure modes

  • Pilot continues indefinitely without a scale decision
  • Success criteria are defined after results are known
  • Playbook is never written and knowledge leaves with the pilot team
  • Scale decision is implicit rather than documented

Reassessment triggers

  • Material change in system behavior or performance
  • Change in use case or affected population
  • New regulatory requirements
  • Incident during pilot

Intended organizational value

Creates an explicit decision boundary between experimentation and operational commitment.

Limitations / proportionality

Project governance does not guarantee that operational teams will follow the playbook. Without handoff to operational owners, pilot learning may not persist.

Domain 3 of 4

CSM-Code

AI-generated code is still organizationally accountable code.

Core question: How should software engineering governance change when AI contributes to implementation?

Problem this domain addresses

AI coding tools contribute to software development. Traditional code review processes were not designed for AI-assisted code. The question is how to maintain engineering accountability when AI tools participate in implementation, including tools that can autonomously execute commands, create files and run tests without per-action human approval.

Verified original components

Development Standards

Engineering standards that account for AI-assisted development, including review requirements and quality expectations.

Security Protocols

Security practices that address AI-generated code, including vulnerability scanning and dependency verification.

Human Oversight

Human review of AI-assisted contributions proportionate to risk and consequence.

Traceability Logging

Records that provide appropriate provenance for AI-assisted changes where risk warrants it.

Current implementation examples

Current implementation interpretation. These examples were not explicitly present in the original 2025 article.

Approved AI coding-tool policy

Document defining which AI tools are permitted for which use cases.

Code review requirements

Review standards that apply to AI-assisted contributions.

Dependency verification

Verification of dependencies introduced or suggested by AI tools.

SAST/security scanning

Static analysis security testing applied to AI-assisted code.

Secret scanning

Detection of credentials or secrets in AI-generated code.

Testing expectations

Test coverage requirements for AI-assisted changes.

AI-assisted change declarations

Disclosure of AI assistance where proportionate to risk.

Provenance/traceability

Records linking AI-assisted changes to review and approval where risk warrants.

Agentic tool authority policy

Document defining what actions autonomous coding agents may take without per-action human approval, including file creation, command execution and test running.

Agent sandbox boundaries

Constraints on autonomous coding agents limiting filesystem access, network calls and production system interaction.

AI API key management

Controls governing API keys used by AI coding tools, including rotation, scope limitation and secret scanning.

Prompt injection defense

Review requirements for AI-generated code that may process untrusted input susceptible to prompt injection.

Example artifacts

AI coding-tool policyCode review recordsSecurity scan resultsTest coverage reportAI-assisted change logException recordAgentic tool authority matrixAgent sandbox configuration

Typical roles involved

Engineering leadSecurity engineerCode reviewerDeveloperDevOps engineerPlatform engineer

Handoffs to other CSM domains

CSM-ProjectCSM-Code: Acceptance criteria and approved assumptions become implementation constraints.
CSM-CodeCSM-UX: Actual system behavior and limitations shape user interaction and oversight.
CSM-CodeCSM-Enterprise: Security findings from AI-assisted code may indicate enterprise policy gaps.

Common failure modes

  • AI-assisted code bypasses normal review
  • Security scanning is skipped for AI-generated contributions
  • No record of which AI tools were used
  • Provenance tracking is required for every token, creating disproportionate overhead
  • Autonomous coding agents execute commands or access systems beyond their approved authority
  • AI API keys are committed to repositories or shared across projects without scope limitation

Reassessment triggers

  • New AI coding tool introduced
  • Security incident involving AI-assisted code
  • Change in approved tool list
  • Change in regulatory or contractual requirements for software provenance
  • Autonomous coding agent granted expanded tool authority or filesystem access
  • AI coding tool underlying model updated or changed by vendor

Intended organizational value

Extends normal engineering accountability into AI-assisted development rather than creating an exception to normal review and security practices.

Limitations / proportionality

CSM-Code does not require tracking every AI-generated token. Provenance and traceability should be proportionate to risk. AI-assisted code is not inherently insecure, but it should not bypass normal review and security practices.

Domain 4 of 4

CSM-UX

Governance reaches the people relying on the system.

Core question: What do humans need to understand, supervise, challenge and appropriately use AI-supported outcomes?

Problem this domain addresses

Users interact with systems whose outputs may be probabilistic, context-sensitive or otherwise require different expectations and oversight. Without governance reaching the user interface, technical controls exist but users do not understand appropriate reliance.

Verified original components

Impact Analysis

Assessment of how AI-supported outcomes affect individuals, groups and workflows.

Explainability Design

Design choices that help users understand system behavior, limitations and appropriate reliance.

Capability Development

Training and skill development that enables users to effectively supervise and interact with AI systems.

Adoption Measurement

Monitoring of how AI systems are actually used, including feedback and complaints.

Current implementation examples

Current implementation interpretation. These examples were not explicitly present in the original 2025 article.

User impact assessment

Evaluation of how outcomes affect users and affected populations.

Appropriate-use guidance

Clear guidance on what the system should and should not be used for.

Explanation/context design

Interface elements that communicate system behavior and limitations.

Confidence/uncertainty presentation

Communication of uncertainty where appropriate to the use case.

Escalation path

Defined process for users to escalate concerns or disputes.

Human review procedure

Process for human review of AI-supported outcomes where required.

Training

User training on appropriate use, limitations and oversight.

Feedback channel

Mechanism for users to report issues or provide feedback.

Adoption/usage review

Periodic review of how the system is actually being used.

Complaint/recourse process

Process for affected individuals to seek review or remedy where relevant.

AI processing consent

Mechanism for users to consent to AI processing of their data, with clear disclosure of what AI does and why.

Opt-out and human review request

User-facing option to opt out of AI processing or request human review of AI-supported outcomes.

Misuse detection monitoring

Proactive monitoring for users bypassing safety controls or using the system for prohibited purposes.

Manipulation guardrail review

Review of interface patterns to prevent deceptive, manipulative or addictive design in AI-driven interactions.

Example artifacts

User impact assessmentAppropriate-use guideInterface design specificationTraining materialFeedback logAdoption review reportConsent disclosure recordMisuse detection report

Typical roles involved

UX designerProduct managerUser advocateTraining leadSupport leadPrivacy counsel

Handoffs to other CSM domains

CSM-CodeCSM-UX: Actual system behavior and limitations shape user interaction and oversight.
CSM-UXCSM-Project: User feedback and operational behavior trigger product/project reassessment.
CSM-UXCSM-Enterprise: User-reported incidents may trigger enterprise-level incident response.

Common failure modes

  • Explainability is confused with perfect model interpretability
  • Users are not informed of system limitations
  • Feedback is collected but never reaches project or enterprise domains
  • Training is one-time and not updated as the system changes
  • Consent is buried or bundled with unrelated terms, preventing meaningful user choice
  • No mechanism for users to request human review of AI-supported outcomes
  • Misuse patterns go undetected because adoption monitoring only tracks volume, not behavior

Reassessment triggers

  • Change in system behavior or output characteristics
  • User complaints or feedback indicating mismatched expectations
  • Change in affected population
  • New regulatory disclosure requirements
  • Detected misuse pattern or users bypassing safety controls
  • Change in consent requirements under applicable privacy or AI regulation

Intended organizational value

Connects governance to actual human behavior and creates feedback that can trigger changes elsewhere in the system.

Limitations / proportionality

Explainability design does not guarantee perfect model interpretability. The goal is appropriate user understanding, not full transparency into model internals. CSM-UX is only effective if feedback actually returns to project and enterprise domains.

System View

How CSM Works as a System

CSM domains are not four sequential phases. They are interconnected governance lenses.

EnterprisePolicy, risk, data, mandateENT-POLICY · ENT-RISK · ENT-DATA · ENT-MANDATEProjectBusiness, testing, scale, playbookPRJ-BUSINESS · PRJ-TESTING · PRJ-SCALE · PRJ-PLAYBOOKCodeStandards, security, human, traceCODE-STANDARDS · CODE-SECURITY · CODE-HUMAN · CODE-TRACEUXImpact, explain, capability, adoptionUX-IMPACT · UX-EXPLAIN · UX-CAPABILITY · UX-ADOPTIONFeedback LoopExecution FunctionsEF1-PURPOSEEF2-MAPPINGEF3-RISKEF4-DELIVERYEF5-OVERSIGHTEF6-COMPLIANCECross-functionalacross all domains
1

Enterprise

Defines purpose, ownership and boundaries.

2

Project

Turns those decisions into business criteria, testing and a scale decision.

3

Code

Implements the system under engineering and security controls.

4

UX

Defines how people use, interpret, challenge and oversee outcomes.

5

Feedback

Operational learning returns to Project and Enterprise.

Handoff Model

Governance Handoffs

One of CSM’s most valuable explanatory concepts is making governance handoffs explicit.

CSM-EnterpriseCSM-Project

Policy and risk boundaries become project requirements.

CSM-ProjectCSM-Code

Acceptance criteria and approved assumptions become implementation constraints.

CSM-CodeCSM-UX

Actual system behavior and limitations shape user interaction and oversight.

CSM-UXCSM-Project

User feedback and operational behavior trigger product/project reassessment.

CSM-ProjectCSM-Enterprise

Incidents, lessons and newly discovered risks may require policy or risk updates.

CSM-EnterpriseCSM-Code

Security incidents require code-level investigation and remediation.

CSM-CodeCSM-Enterprise

Security findings from AI-assisted code may indicate enterprise policy gaps.

CSM-UXCSM-Enterprise

User-reported incidents may trigger enterprise-level incident response.

Getting Started

Implementation Guidance

Adapted from the original 2025 implementation guidance.

1

Assessment

  • Inventory existing AI initiatives
  • Review governance approaches
  • Identify gaps
  • Identify applicable legal/contractual requirements
  • Assess organizational readiness
2

Pilot

  • Choose an initiative with meaningful governance questions
  • Apply the relevant CSM domains and components
  • Document findings and lessons
  • Refine the approach
3

Scale

  • Integrate useful governance practices into existing operating processes
  • Develop organizational capability
  • Establish governance effectiveness measures appropriate to the organization

This is not a mandatory waterfall. Organizations may begin with any domain where governance gaps are most consequential.

Right-Sizing

Proportionality

Not every AI system requires equal CSM depth. Governance intensity should depend on factors such as intended use, consequence of failure, autonomy, reversibility, data sensitivity, affected population, scale, external exposure and regulatory or contractual obligations.

Intended useConsequence of failureAutonomyReversibilityData sensitivityAffected populationScaleExternal exposureRegulatory/contractual obligations

A low-risk internal productivity tool does not need the same governance as consequential employment, healthcare or financial decisions.

Boundaries

What CSM Is Not

Certification
Regulation
Legal advice
Compliance guarantee
NIST replacement
ISO replacement
SDLC replacement
Project-management replacement
Software product

CSM is: A governance methodology for connecting enterprise, project, engineering and human-interaction responsibilities around AI-enabled systems.

Current Architecture

CSM and the AI Governance Execution Framework

CSM

CSM answers WHERE governance responsibilities operate.
EnterpriseProjectCodeUX

AI Governance Execution Framework

The AI Governance Execution Framework answers WHAT operating functions should continuously occur across those domains.
  1. Purpose, Scope & Accountability
  2. System, Data & Dependency Mapping
  3. Risk, Evaluation & Monitoring
  4. Controlled Delivery & Change
  5. Human Oversight, Feedback & Learning
  6. Compliance, Evidence & Assurance

The AI Governance Execution Framework operationalizes and extends Cognitive Systems Management into six cross-functional governance functions.

Current framework architecture. This relationship was not part of the original 2025 publication.

Layered Architecture

CSM, Execution Framework, and HAIEC

CSM

Conceptual governance domains.

AI Governance Execution Framework

Operational governance functions.

HAIEC

Technology and workflow capabilities supporting selected activities.

HAIEC does not fully implement every CSM or Execution Framework responsibility. Some responsibilities remain with the organization.

Honest Assessment

Limitations

CSM is a practitioner-developed governance methodology. The original publication described it as an evolving framework with limited long-term validation data. It should be applied proportionately and alongside applicable organizational, technical, legal and assurance practices.

Do not call CSM peer reviewed. Do not call it an industry standard. It is a practitioner-developed governance methodology.

Resources

CSM 2.0 Framework Resources

CSM 2.0 Specification

Full versioned specification: domains, execution functions, governance contracts, state model, determinism boundary, evidence and decision schemas.

Governance Contracts

Browsable reference for all 16 component governance contracts with IDs, core questions, inputs, decisions, evidence and handoffs.

Framework Guide

Full versioned framework guide with all domains, components, handoffs, implementation guidance and provenance.

Quick Reference

Concise printable reference: four domains, six execution functions, governance contract model.

Reference Assessment

Interactive reference evaluator. Provide structured system facts and receive applicable CSM requirements, evidence gaps and human review items.

Machine-Readable Spec

JSON specification and schema for programmatic consumption. Generated from canonical TypeScript source.
FAQ

Frequently Asked Questions

Is CSM a standard or certification?

No. CSM is a practitioner-developed governance methodology. It is not an industry standard, regulation, or certification program. It is designed to complement standards like NIST AI RMF and ISO/IEC 42001, not replace them.

Does CSM replace NIST AI RMF or ISO 42001?

No. CSM provides a governance operating model that helps organizations implement those standards more effectively. The V2 specification includes an informative NIST/ISO crosswalk showing how CSM components map to NIST functions and ISO clauses.

Is CSM 2.0 backward compatible with V1?

Yes. CSM 2.0 preserves all four original domains, all sixteen original components, and their names unchanged. V2 adds an operational specification layer on top of the existing structure: governance contracts, execution functions, state models, and reassessment triggers.

How long does CSM adoption take?

CSM is designed to be adopted incrementally. The implementation guidance outlines three phases: assess current state, pilot on one system, then scale. Organizations can start with a single domain or system and expand. Proportionality factors allow smaller AI deployments to apply lighter governance depth.

What is the difference between CSM and the Execution Framework?

CSM defines WHERE governance responsibilities operate (the four domains: Enterprise, Project, Code, UX). The AI Governance Execution Framework defines WHAT governance activities happen (six cross-functional execution functions: Purpose, Mapping, Risk, Delivery, Oversight, Compliance). Together they form a two-axis operating architecture.

Does CSM 2.0 produce compliance verdicts or scores?

No. CSM 2.0 does not produce legal compliance verdicts, aggregate scores, or pass/fail ratings. The reference assessment identifies applicable requirements, evidence gaps, and human review items. It is a governance tool, not an audit instrument.

Can CSM be used for non-AI systems?

CSM was designed specifically for AI systems where model behavior, training data, and human-AI interaction introduce governance challenges not covered by traditional software governance. While some concepts transfer, the framework is optimized for AI and machine learning systems.

Provenance

Source and References

Cognitive System Management: A Framework for Enterprise AI Project Governance

Subodh KC · AI Governance on Medium · August 29, 2025
Read original article

Current Framework Guide revision: August 2026 (v1.0).This guide expands the original CSM publication with current implementation examples. Material labeled as implementation guidance should not be interpreted as having appeared verbatim in the original article.

Explore CSM 2.0

Dive deeper into the full specification, governance contracts, reference assessment, or explore how CSM 2.0 connects to the AI Governance Execution Framework and HAIEC.

Discuss AI →