level-1-essentials/company/01-governance/README.md

01 — Governance

Governance before tools. Always.

A company with an AUP but no AI beats a company with AI but no AUP. Every incident in the past three years has come from an unstructured deployment.

Files in this folder

1. acceptable-use-policy.md — the one-page AUP employees actually read
2. data-classification-matrix.md — which data class goes to which tool
3. vendor-dpa-checklist.md — DPAs to sign with each vendor
4. incident-response-addendum.md — what to do when PII leaks
5. steering-committee-charter.md — who sits on it, what they decide
6. zero-retention-configuration.md — how to enable no-training on every provider

The sequence

Do these in strict order. Skipping = incident.

1. Convene the Steering Committee (day 1)
2. Draft and publish AUP (week 1)
3. Publish Data Classification Matrix (week 1)
4. Enable zero-retention on approved providers (week 2)
5. Sign DPAs with each approved provider (week 2)
6. Publish Incident Response Addendum (week 2)
7. Only THEN move to 02-accounts-procurement/

Definition of done

• [ ] AUP signed by all employees
• [ ] Every approved AI tool has a signed DPA in place
• [ ] Zero-retention verified on every enterprise tenant
• [ ] Steering Committee meets bi-weekly with published agenda
• [ ] Incident Response Addendum drilled at least once (tabletop exercise)

level-1-essentials/company/01-governance/acceptable-use-policy.md

AI Acceptable Use Policy (AUP)

Effective date: [DATE] Owner: Chief AI Officer / Head of AI / Steering Committee Chair Version: 1.0

Purpose

This policy defines what AI tools employees of [COMPANY] may use, what data they may input, and what obligations they carry.

Read this before using any AI tool for company work. Sign the acknowledgment at the end.

Scope

Applies to: - All employees, contractors, and interns - All AI tools — consumer, enterprise, API, open-source, embedded features - All company data — whether accessed at work, at home, or on personal devices

The four rules

Rule 1 — Only approved tools for work

The approved tool list is at company/02-accounts-procurement/platform-account-order.md. Anything else requires Steering Committee approval before use with company data.

Rule 2 — Data classification, then tool

Every piece of company data has a classification: Public, Internal, Confidential, Restricted. Match the classification to the tool tier per data-classification-matrix.md. When in doubt, treat as more sensitive, not less.

Rule 3 — You are responsible for AI output

• AI is a tool; you are the author.
• Review every AI-generated output before it is sent, published, or committed.
• You are accountable for accuracy, tone, and legal compliance of anything you approve.

Rule 4 — Report incidents immediately

If you paste sensitive data into an unapproved tool, receive suspicious output, or suspect a prompt injection, notify [SECURITY EMAIL] within 24 hours. No blame for reporting; blame for not reporting.

Explicit permissions

Employees may: - Use approved AI tools for drafting, summarization, research, code assistance, and analysis - Save prompt libraries locally or in company drives - Build automations with approved tools within their team's remit - Share AI-generated drafts internally for review

Explicit prohibitions

Employees may not: - Paste customer PII into a consumer AI tool (free ChatGPT, personal Claude, personal Gemini) - Paste source code with credentials or unreleased IP into any unapproved tool - Send AI-generated communications externally without human review - Represent AI-generated content as fully human-authored when accuracy matters (client deliverables, legal docs, medical/financial advice) - Use AI tools to circumvent company policy (fabricating evidence, drafting deceptive communications, evading review processes) - Build AI automations that send external communications without a human-in-the-loop approval step - Share their AI account credentials - Use personal AI subscriptions for confidential company work

Consequences

Violations follow standard disciplinary policy. Egregious cases (deliberate data exfiltration, deceptive use) may result in termination and legal action.

Amendments

This policy is reviewed by the AI Steering Committee quarterly and updated as the tool landscape evolves. Changes are communicated via all-hands and require re-acknowledgment.

Acknowledgment

I have read and understood this policy. I understand what I may and may not do, and I understand my obligation to report incidents.

Name: ___ Date: ___ Signature: ___

Plain-language summary (for orientation decks)

1. Only use tools on our approved list for work.
2. Confidential data goes only in Enterprise-tier tools with DPAs.
3. Never paste passwords, keys, SSNs, or client PII into consumer AI.
4. Review AI output before you use it.
5. If you mess up, tell Security. Reporting is safe.

level-1-essentials/company/01-governance/data-classification-matrix.md

Data Classification Matrix

The four classes

The matrix

Rows = data class. Columns = tool tier. Cells = allowed / not allowed.

Specific data types

When in doubt

Treat as more sensitive. Downgrading later is cheap. Cleaning up an incident is expensive.

Enforcement

• Technical: DLP scans block classified data from egressing to unapproved tools where possible
• Procedural: Regular training + spot audits
• Cultural: Report-safely policy in the AUP

Amendments

The Steering Committee reviews this matrix quarterly. Additions to the "restricted" list require CISO approval.

level-1-essentials/company/01-governance/incident-response-addendum.md

Incident Response Addendum — AI

This addendum extends the company's existing IR plan for AI-specific incidents.

Incident classes

Response by class

A1 — Data exposure

1. Reporter contacts Security via [SECURITY EMAIL] within 24 hours
2. Security confirms: what data, what tool, when, who else has access
3. Security requests deletion from vendor (if enterprise) or notes as unrecoverable (if consumer)
4. Legal assesses regulatory notification duty (GDPR: 72 hours to DPA; HIPAA: 60 days for large; state laws vary)
5. Affected parties notified per legal guidance
6. Employee coaching (not punitive on first report)

A2 — Model output leak

1. Recall or retract the leaked output where possible
2. Investigate: was it a hallucination, or actual data in the model context?
3. If context-sourced, treat as A1
4. If hallucinated, note as a quality issue; add to eval set

A3 — Prompt injection

1. Isolate the affected agent immediately (kill switch)
2. Determine blast radius: what tools did the agent have access to?
3. Preserve logs
4. Rotate any credentials the agent could have touched
5. Root cause: what input carried the injection? How did it evade defenses?
6. Publish an incident report to the Steering Committee within 5 business days

A4 — Agent action error

1. Stop the agent
2. Undo the action if reversible (recall email, restore deleted data)
3. Notify the affected party if action reached outside the org
4. Investigate the decision path
5. Add a HITL checkpoint or guardrail before re-enabling

A5 — Model misuse

1. Preserve evidence
2. Follow HR/legal escalation
3. Steering Committee reviews for policy gap

A6 — Vendor incident

1. Acknowledge vendor notification
2. Assess: is our data affected? What data?
3. Legal reviews vendor's breach report against contract commitments
4. Follow through on any downstream customer notifications required

A7 — Bias / harm

1. Suspend the affected workflow
2. Convene bias review (Steering Committee + Legal + Ethics if applicable)
3. Assess root cause: model, prompt, eval gap, data drift
4. Remediate before restart

Communication tree

• Reporter → Security (within 24 hours of discovery)
• Security → Legal, IT, Steering Committee (within 24 hours of confirmation)
• Steering Committee → CEO / GC (within 48 hours if A1–A3 with material impact)
• Company → Affected parties (per legal guidance)
• Company → Regulators (per statutory deadlines)

Tabletop exercise

Run a tabletop at least annually:

1. Choose an incident class
2. Announce the scenario to Security, Legal, IT, an AI Champion, and an executive
3. Time the response
4. Document gaps

Post-incident

Every material incident produces: - Incident report — timeline, root cause, blast radius, remediation - AUP update if a policy gap contributed - Eval set update — the incident scenario added as a regression test - Training update — the case incorporated into next quarterly refresh

Reporter protection

Reporters are protected from retaliation. First-time honest reports do not trigger discipline. Cover-ups do.

level-1-essentials/company/01-governance/steering-committee-charter.md

AI Steering Committee — Charter

Purpose

The AI Steering Committee sets policy, approves tools, oversees rollout, reviews incidents, and steers investment for AI at [COMPANY].

It is a governance body, not a build team. Build teams execute; the Committee decides scope, guardrails, and priorities.

Members

Size cap: 8 voting members. Rotating seats change every 6 months.

Cadence

• Weekly — 30-min triage of new requests during initial rollout (first 90 days)
• Bi-weekly — 60-min working session
• Monthly — 60-min executive review
• Quarterly — 2-hour full review + tool bake-off + policy refresh

Decision types

Standing agenda

Every meeting:

1. Incident review (5 min)
2. Approved-tool list changes (5 min)
3. Pilot progress (10 min)
4. Budget / cost review (5 min)
5. Policy / process items (15 min)
6. New requests intake (10 min)
7. Action items (5 min)

Intake

Anyone in the company can submit a request: - New tool approval - New use-case sanction - New data type classification - Policy question

Submit via [FORM URL]. Committee triages within 5 business days.

Public commitments

The Committee publishes internally: - Approved tool list (up to date) - Data classification matrix - Policy changelog - Quarterly summary (wins, incidents, spend)

Committee lifecycle

• Formed at kickoff of AI program
• Full charter (this doc) reviewed annually
• Committee itself is reviewed by Executive Sponsor annually — dissolved if AI becomes fully embedded across departments (steady-state governance moves to existing functions)

Escalation from Committee

• To Executive Sponsor — anything requiring exec approval, budget beyond charter cap
• To Board — material regulatory risk, or AI use with strategic implications

level-1-essentials/company/01-governance/vendor-dpa-checklist.md

Vendor DPA Checklist

A Data Processing Agreement (DPA) must be signed with every AI vendor that processes company data. Do this before rollout, not after.

What to require in every AI DPA

• [ ] Zero-retention clause — vendor will not retain inputs beyond the request lifetime, or will retain only for a defined short period (e.g., 30 days for abuse prevention) and only in encrypted form
• [ ] No-training clause — vendor will not train models on our inputs or outputs
• [ ] Sub-processor list — vendor identifies all sub-processors and notifies of changes
• [ ] Data location — where processing occurs; EU customers require EU residency options
• [ ] Breach notification — vendor notifies within 72 hours of a security incident
• [ ] Audit rights — right to request evidence of controls (SOC 2 Type II report at minimum)
• [ ] Deletion on termination — vendor deletes all data within defined period after contract ends
• [ ] Data portability — right to export data before termination
• [ ] Liability caps — appropriate to volume and sensitivity of data
• [ ] BAA addendum — if PHI is involved (HIPAA)
• [ ] DPA appendix compliant with GDPR — if EU personal data
• [ ] Standard Contractual Clauses (SCCs) — for cross-border EU-US transfers

Provider-specific notes

OpenAI (Enterprise / Team)

• API zero-retention available (x-openai-organization header + request); Enterprise ChatGPT no-training default
• Sign OpenAI DPA + BAA (if applicable)
• Verify: SOC 2 Type II, CSA STAR

Anthropic (Claude for Enterprise)

• No training on API traffic by default; verify in contract
• Zero-retention for API available under Enterprise contract
• Sign Anthropic DPA
• Verify: SOC 2 Type II

Google (Gemini for Workspace / Vertex AI)

• Existing Google Workspace DPA typically covers Gemini
• Vertex AI has separate terms
• Verify: SOC 2, ISO 27001

Microsoft (Copilot for M365 / Azure OpenAI)

• Existing M365 DPA covers Copilot for Microsoft 365
• Azure OpenAI has its own commitments (no data used for training)
• Verify: HITRUST, SOC 2, ISO 27001

Perplexity (Enterprise Pro)

• Zero-retention available under Enterprise agreement
• Sign Perplexity DPA
• Verify enterprise controls

xAI / Grok

• Read current terms carefully — training defaults have changed multiple times
• Confirm no-training in writing

Groq / OpenRouter / Together / Fireworks

• Read their terms; they route to underlying open-weight models
• Fewer certifications typically; use only for public/internal data unless verified

DPA request template

Send to vendor:

Subject: DPA request for [COMPANY] AI deployment

Hello [VENDOR],

We are preparing to deploy [PRODUCT] across [COMPANY]. Before rollout, we need: 1. Your current Data Processing Agreement (executable form) 2. Confirmation of zero-retention on [SERVICE TIER] 3. SOC 2 Type II report (under NDA) 4. BAA if we process PHI [yes/no] 5. Sub-processor list 6. Data residency options

Please share these and route to our legal team at [EMAIL].

Thank you.

Tracking

Maintain a register:

Owned by: Legal or Vendor Management. Reviewed quarterly by Steering Committee.

level-1-essentials/company/01-governance/zero-retention-configuration.md

Zero-Retention Configuration Guide

For each approved provider, this is how you configure "no training, no retention beyond request lifetime."

Verify after configuring; screenshot the settings for the audit file.

OpenAI

ChatGPT Enterprise / Team

• Default: no training on your prompts, no retention beyond 30 days for abuse monitoring (encrypted)
• Verify: Admin Console → Data Controls
• Contract: OpenAI Enterprise Agreement + DPA + BAA if applicable

ChatGPT Business (SMB tier)

• Default: no training; same 30-day abuse window
• Verify: Admin Console

OpenAI API

• Default: no training on API traffic (since March 2023)
• Zero-retention: enterprise customers can request x-openai-zdr header behavior or contractual zero-retention
• Verify: contract + openai.com/policies/api-data-usage-policies

Anthropic

Claude for Enterprise / Team

• Default: no training on Team/Enterprise conversations
• Zero-retention: configurable under Enterprise contract
• Verify: Admin Settings → Data & Privacy
• Contract: Anthropic MSA + DPA

Anthropic API

• Default: no training on API traffic
• Zero-retention available in Enterprise contract
• Verify: contract terms + Anthropic Trust Center

Google

Gemini for Google Workspace

• Default: no training on Workspace data
• Verify: Admin Console → Gemini App settings
• Contract: existing Workspace DPA (Gemini is covered)

Google AI Studio / Vertex AI (paid tier)

• Vertex AI paid tier: no training on your data
• AI Studio free tier: data may be used to improve models — DO NOT use for company data on free tier
• Verify: Vertex AI console → project settings
• Contract: Google Cloud DPA

Microsoft

Copilot for Microsoft 365

• Default: no training on your prompts / responses
• Data stays in Microsoft 365 service boundary
• Verify: Admin Center → Copilot governance settings
• Contract: existing M365 DPA covers this

Azure OpenAI Service

• Default: no training on customer data (contractually committed)
• Data stays in your Azure tenant / region
• Verify: Azure portal → OpenAI resource → Networking + Content filtering
• Contract: Azure Enterprise Agreement + AI Services Product Terms

Perplexity

Perplexity Enterprise Pro

• Default: no training on Enterprise Pro conversations
• Verify: Admin settings → Data & Privacy
• Contract: Perplexity Enterprise Agreement + DPA

Perplexity API

• Configure per contract — request zero-retention explicitly

xAI Grok

• Read current terms — default has changed multiple times
• If used: only for Public data unless zero-retention is contractually confirmed
• Verify quarterly

Groq / Together / Fireworks / OpenRouter

• These providers route to open-weight models
• Read each provider's data policy — most claim no training, but certifications vary
• Suitable for Public / Internal data by default; Confidential only with contract review

Local / self-hosted (Ollama, vLLM, TensorRT-LLM)

• Zero data leaves your infrastructure by definition
• Suitable for all classes, subject to local security controls

Verification checklist per provider

After configuring, save these to the audit file governance/audits/YYYY-MM/:

• [ ] Screenshot of the "no training" toggle set to off
• [ ] Screenshot of the retention setting
• [ ] Copy of signed DPA
• [ ] Date of last verification
• [ ] Owner (person responsible for re-verifying quarterly)

Quarterly re-verification

The Steering Committee's quarterly agenda includes: - Walk through this document - Confirm every provider still configured correctly - Note any changes to vendor defaults (which happen)

If a vendor changes defaults in a way that increases retention or training, they must renotify per DPA; escalate immediately.