AI Asset Inventory Register

AI Asset Taxonomy

[ORGANIZATION NAME] classifies AI assets into the following categories. Every AI asset must be assigned to exactly one primary category.

Category 1: AI Models

• Definition: Machine learning models, large language models, and other AI algorithms deployed or used by the organization
• Subcategories:

- Proprietary models (trained or fine-tuned by [ORGANIZATION NAME])

- Commercial models (licensed from vendors, e.g., GPT-4, Claude, Gemini)

- Open-source models (downloaded and self-hosted)

- Embedded models (bundled within vendor products, not directly managed)

Category 2: AI Platforms and Tools

• Definition: Software platforms, services, and tools that provide AI capabilities
• Subcategories:

- Enterprise AI platforms (Azure OpenAI, AWS Bedrock, Google Vertex AI)

- AI development tools (code assistants, testing tools, IDE plugins)

- AI-powered SaaS applications (AI features within business applications)

- AI infrastructure (GPU clusters, model serving platforms, vector databases)

Category 3: AI Integrations

• Definition: Connections between AI systems and other organizational systems
• Subcategories:

- API integrations (AI model APIs consumed by applications)

- Data pipeline integrations (data feeds into AI systems, RAG pipelines)

- Agent tool integrations (tools and functions callable by AI agents)

- Webhook and event integrations (AI systems triggered by external events)

Category 4: AI Data Assets

• Definition: Datasets used for training, fine-tuning, evaluation, or retrieval by AI systems
• Subcategories:

- Training datasets

- Fine-tuning datasets

- Evaluation and benchmark datasets

- RAG knowledge bases and vector stores

- Prompt libraries and template repositories

Inventory Record Schema

Every AI asset record must include the following fields. Fields marked with * are mandatory.

Core Fields

Technical Fields

Governance Fields

Shadow AI Definition

Shadow AI refers to any AI system, tool, or service used by [ORGANIZATION NAME] personnel that has not been formally approved, registered in the AI asset inventory, and subjected to security and governance review. Shadow AI represents a significant risk because it operates outside established controls.

Discovery Methods

Method 1: Network Traffic Analysis

• Approach: Monitor outbound network traffic for connections to known AI service endpoints
• Tool: CASB (Cloud Access Security Broker), web proxy, or firewall logs
• Frequency: Continuous monitoring with weekly review of new findings
• Target Domains: Monitor for traffic to known AI API endpoints including:

- api.openai.com, api.anthropic.com, generativelanguage.googleapis.com

- api.mistral.ai, api.cohere.ai, api.together.xyz

- Consumer AI interfaces: chat.openai.com, claude.ai, gemini.google.com, copilot.microsoft.com

- AI coding tools: api.github.com/copilot, codeium.com, cursor.sh

Method 2: Endpoint Agent Analysis

• Approach: Scan endpoint agents (EDR/UEM) for installed AI applications, browser extensions, and IDE plugins
• Tool: Endpoint detection and response (EDR) or unified endpoint management (UEM)
• Frequency: Weekly scan with monthly trend analysis
• Target: AI-related browser extensions, desktop applications, IDE plugins not on the approved list

Method 3: Expense and Procurement Review

• Approach: Review expense reports, corporate credit card transactions, and procurement requests for AI-related purchases
• Tool: Expense management system, procurement platform
• Frequency: Monthly review
• Target: Subscriptions to AI services, API credit purchases, AI training or certification not routed through IT

Method 4: Employee Self-Reporting

• Approach: Provide a low-friction mechanism for employees to report AI tools they are using that may not be registered
• Tool: Dedicated form or email alias (e.g., ai-inventory@[ORGANIZATION DOMAIN])
• Frequency: Continuous, with quarterly amnesty campaigns (no-penalty registration period)
• Incentive: No disciplinary action for self-reporting during amnesty periods; focus on bringing tools into governance

Discovery Response Process

When shadow AI is discovered:

1. Log: Record the finding in the shadow AI discovery log with: date, discovery method, tool/service identified, users involved, data classification assessment
2. Assess: Evaluate the risk level and determine if the AI tool serves a legitimate business need
3. Decide: Either (a) bring the tool into governance through the standard approval process or (b) block access and notify affected users with approved alternatives
4. Remediate: If sensitive data was processed through the shadow AI tool, initiate incident response procedures per the AI Incident Response Plan

Review Schedule

Validation Procedures

During each full inventory validation:

1. Completeness Check: Compare inventory against shadow AI discovery findings, procurement records, and network monitoring to identify unregistered assets
2. Accuracy Check: Asset owners confirm that all fields are current and accurate for each asset in their ownership
3. Status Check: Verify that assets marked as "Active" are still in use, and assets no longer in use are marked for decommission
4. Compliance Check: Verify that all required governance artifacts (DPAs, approvals, security reviews) are current and not expired
5. Risk Check: Reassess risk levels based on changes in the threat landscape, regulatory environment, or organizational risk appetite

Inventory Metrics

Change Management

Any material change to a registered AI asset requires an inventory update within 5 business days. Material changes include:

• Model version upgrade or swap
• Change in data classification of processed data
• Change in deployment type or hosting location
• Change in vendor or licensing terms
• Change in user population or use case scope
• Addition or removal of agent tool integrations

Changes to the inventory schema or review procedures require approval from the AI Governance Committee.

Inventory Owner: [ROLE TITLE], [DEPARTMENT]

Last Full Validation: [DATE]

Next Full Validation Due: [DATE]