AI Governance
Your AI Safety Framework: Three Tiers
Not all AI tools are created equal. This framework helps you choose the right tool based on the sensitivity of your task — from public chatbots to fully isolated, on-premise systems.
This framework is for informational purposes only and does not constitute legal advice.
Framework originally developed by Michael James of learnmanagementonline.com. Adapted and presented here with appreciation for its practical clarity and applicability to enterprise data and CDP environments.
Public LLMs
"The Crowded Café" — An outsider. Treat it like a public social media post.
Public AI tools like ChatGPT (free), Gemini, and open chatbots operate over the public internet with no company-level privacy or security controls. Anything you type could be stored, reviewed, or used to train the model.
Examples: ChatGPT (free), Gemini, Claude.ai (free), open chatbots
How your data is used and retained
Stored and reviewed
Public AI tools may store and review everything you enter, including prompts and uploaded files.
Used for model training
Your data can be used to train and improve the model. Your inputs may influence future outputs seen by other users.
Logged and analyzed
Information may be logged or analyzed by the provider for quality assurance, safety, or product improvement.
No company-level controls
No company-level privacy, security, or control over your data. You have no DPA, no SLA, and no audit rights.
Appropriate use cases only
General knowledge queries
Look up publicly available facts, definitions, and concepts.
Learning & skill development
Practice new skills, study for exams, learn a language.
Public information research
Summarize news, explore open datasets, get general context.
Creative brainstorming
Non-confidential ideation, writing, and creative exploration.
What NEVER to input
Personally Identifiable Information
Names, addresses, SSNs, employee records, health data, or any data that identifies a specific individual.
Company data
Internal financials, strategy documents, customer lists, code repositories, or any proprietary business information.
Confidential information
Trade secrets, legal matters, M&A activity, board materials, unreleased product details, or classified data.
Enterprise AI
"The Secure Meeting Room" — A fellow colleague. Your go-to for work.
Enterprise AI platforms are company-approved tools with contractual security protections. They're the right choice for the majority of work scenarios — emails, reports, analysis, brainstorming, and content creation.
Examples: ChatGPT Enterprise, MS Copilot, internal AI assistants
Key security features
Zero retention
Enterprise agreements ensure your data is not stored or retained by the AI provider after processing.
Encryption
Data encrypted in transit and at rest. Your inputs and outputs are protected end-to-end.
Compliance
Covered by your company's enterprise agreement, DPA, SOC 2, and applicable regulatory frameworks.
When to use (majority of work scenarios)
Writing emails, summaries, or reports
Draft internal communications, meeting notes, status updates.
Drafting internal documents or presentations
Create decks, proposals, and internal-facing content.
Analyzing work data (non-highly sensitive)
Process spreadsheets, summarize reports, extract insights.
Brainstorming for projects or client work
Strategy sessions, idea generation, problem solving.
Creating content for teams or customers
Marketing copy, training materials, FAQ content.
How to access company-approved tools
Via your company login (SSO)
Authenticate through your corporate single sign-on to ensure audit trails and access controls.
Through approved apps
ChatGPT Enterprise, MS Copilot, internal AI assistants, and other company-sanctioned platforms.
Inside secure company platforms
Microsoft 365, internal dashboards, or company portals with enterprise security controls.
On-Premise AI
"The Confidential Meeting" — A fellow colleague with a security clearance.
On-premise AI (or Virtual Private Cloud) provides the highest level of security and data sovereignty. All processing happens on company-owned infrastructure with complete isolation from external networks. Reserved for your most sensitive work.
Examples: Self-hosted LLMs, VPC-deployed models, on-prem inference servers
Complete data control & sovereignty
Data stays on-site
All data stays inside company-owned servers. Nothing leaves your infrastructure.
No external vendors
No external vendors or cloud services involved in processing.
Company controls everything
Full control over storage, access, retention, and deletion policies.
Maximum isolation
Highest level of privacy and data isolation available.
Regulatory compliance scenarios
Strict legal and industry requirements
Meets mandates for HIPAA, SOX, ITAR, and sector-specific regulations.
Data residency and jurisdiction rules
Supports GDPR data localization, sovereignty requirements, and cross-border restrictions.
Mandatory isolation environments
Required for air-gapped networks, classified systems, and government work.
High-risk or regulated workflows
Critical infrastructure, defense, financial custody, and life-safety systems.
Appropriate use cases
Highly sensitive or classified data
Government secrets, defense materials, top-secret research.
Regulated industry tasks
Healthcare (HIPAA), finance (SOX), legal (privilege), government compliance.
R&D involving critical IP
Proprietary algorithms, trade secrets, pre-patent inventions, unreleased products.
Internal analytics requiring strict confidentiality
M&A modeling, compensation benchmarks, litigation hold data.
Quick Reference
Three tiers at a glance
| Criteria | Tier 1 — Public LLMs | Tier 2 — Enterprise AI | Tier 3 — On-Premise |
|---|---|---|---|
| Analogy | "The Crowded Café" — An outsider | "The Secure Meeting Room" — A colleague | "The Confidential Meeting" — Colleague with clearance |
| Data retention | May store, train on, and review your data | Zero retention; contractual protections | Full company control; nothing leaves your servers |
| Security level | None — public internet, no DPA or SLA | Encryption, SSO, enterprise DPA, SOC 2 | Maximum — air-gapped, company-owned infra |
| Use for | General knowledge, public research, learning, non-confidential brainstorming | Emails, reports, presentations, work data analysis, content creation | Classified data, regulated workflows, critical IP, high-risk analytics |
| Never use for | PII, company data, confidential info — anything sensitive | Highly classified, air-gapped, or regulation-mandated isolation scenarios | N/A — suitable for all data types |
| Access | Open internet — anyone can sign up | Company SSO, approved apps (ChatGPT Enterprise, Copilot, etc.) | Secure channels only, elevated permissions, designated teams |
| Examples | ChatGPT (free), Gemini, Claude.ai (free), open chatbots | ChatGPT Enterprise, MS Copilot, internal AI assistants | Self-hosted LLMs, VPC-deployed models, on-prem inference servers |
When in doubt, tier up.
If you're unsure which tier is appropriate for your task, always choose a higher tier. It's better to over-protect than to expose sensitive data.
Legal Disclaimer
This content is intended for educational purposes only and does not constitute legal advice. AI governance and regulatory requirements, including those under the EU AI Act, depend on specific facts and implementation details. Organizations should seek advice from qualified legal professionals to assess their obligations and risk exposure.
About this framework
This AI safety training framework was originally developed by Michael James of learnmanagementonline.com — an online learning platform focused on practical management education.
CdpHub has adapted and republished this framework for enterprise data and CDP practitioners, given its direct applicability to how organizations should think about AI tool selection in data-sensitive environments.
Visit learnmanagementonline.comExplore more from CdpHub
Interested in how AI governance frameworks intersect with your CDP selection and implementation strategy? We're here to help.
