Deploy Your Constitution

Why Deploy Layer 1 First

Layer 1 — the Corporate Constitution — provides the highest-impact governance improvement with the lowest implementation complexity.

• Without Layer 1: Every AI session starts from zero. The AI has no organizational identity, no regulatory awareness, no data classification framework, and no basis for refusing inappropriate requests.
• With Layer 1: Every AI session starts with identity, authority limits, data handling rules, and behavioral mandates already loaded — even if Layers 2 and 3 are never implemented.

Layer 1 alone transforms AI from "general-purpose tool an employee happens to use at work" to "organizational participant operating within defined boundaries."

Deployment Methods

Enterprise AI tools fall into four categories based on how they accept constitutional instructions:

Category	Mechanism	Deployment Method
File-based	Instruction files in repos or on endpoints	Deploy constitution as a managed file
API-level	System prompt parameter in API calls	Inject constitution server-side in application code
Builder UI	Agent/assistant configuration interfaces	Paste constitution into instructions field
Policy-layer	Guardrails, trust layers, governance gateways	Encode constitutional rules as platform policies

Most organizations use tools across multiple categories. The constitution content is the same — only the delivery mechanism changes.

Token Budget Considerations

AI models have finite context windows. The constitution consumes tokens from that budget:

Constitution type	Approximate tokens
Corporate only (minimal)	800 - 1,200
Corporate only (full)	1,500 - 2,500
Corporate + Team	2,000 - 3,500
Corporate + Team + Practitioner	2,500 - 4,500

The Constitution Builder tracks token count in real time as you fill in each section, so you can stay within your deployment target's limits.

AI Coding Agents

These tools operate in developer environments. Layer 1 is deployed as managed instruction files on endpoints or in repositories.

Enterprise Chat & Assistant Platforms

These platforms serve non-developer employees through conversational interfaces. Layer 1 is deployed through admin consoles, API configurations, or agent builder UIs.

Agent Orchestration Frameworks

Used by development teams building custom AI applications. Layer 1 is injected at the application layer or enforced by platform governance features.

• LangChain / LangSmith — Publish as a versioned prompt template in LangChain Hub. LangSmith Fleet provides the centralized registry.
• Amazon Bedrock AgentCore — Encode as Cedar policies. The Gateway enforces every tool call deterministically before execution. Framework-agnostic.
• Google Vertex AI Agent Builder — Set constitutional system instructions. Use IAM to scope agent identities as first-class principals.
• Microsoft Agent Framework — Deploy via Azure governance controls. Prompt shields protect against injection. Purview integration for compliance.
• Databricks / Mosaic AI — Unity Catalog for unified governance. Mosaic AI Gateway provides the centralized entry point with rate limiting and permissions.
• Snowflake Cortex — Agent instructions via Cortex APIs. AI Governance Gateway enforces access control. All processing stays within Snowflake's security perimeter.
• Palantir AIP — The strongest ontological governance model. Constitutional rules embedded in the platform's security model. Supports air-gapped environments.

Cross-Platform Deployment Strategy

For organizations using multiple AI tools, deploy Layer 1 in priority tiers:

Tier A — Immediate (highest impact, lowest effort)

These mechanisms lock the constitution at the platform level. Users cannot bypass them.

1. Claude Code — Deploy managed-settings.json + managed CLAUDE.md via MDM
2. Amazon Bedrock — IAM-enforced guardrails via AWS Organizations
3. Tabnine — Admin guidelines in the Admin UI (override user-level)
4. GitHub Copilot — Enterprise-level policies in admin console

Tier B — Standard (requires admin console configuration)

5. Microsoft Copilot — Copilot Control System policies + Declarative Agent instructions
6. ChatGPT Enterprise — Corporate Custom GPT + workspace restrictions
7. Cursor — Enterprise dashboard team-wide rules
8. Salesforce Agentforce — Agent Script + Einstein Trust Layer
9. Windsurf — OS-level system workflows via MDM

Tier C — Application-layer (requires development work)

10. Any API-based deployment — Inject constitution as system/developer message server-side
11. LangChain/LangSmith — Publish as versioned Hub template
12. Custom internal platforms — Embed in prompt management layer

Tier D — Supplementary (defense in depth)

13. AGENTS.md — Commit to all repos; protect via branch rules
14. Repo-level instruction files — For Layer 2 context; not a substitute for enterprise enforcement
15. ServiceNow AI Control Tower — Cross-platform governance overlay

Adapting for Different Platforms

Token / length constraints

If a platform constrains system prompt length, prioritize in this order:

1. Sections 01 (Identity) and 02 (Organizational Context) — the constitutional foundation
2. Sections 04 (Data Classification) and 05 (Behavioral Mandates) — the enforcement core
3. Sections 06-08 (Misuse, Refusal, Scope) — can be condensed

Platform-specific formatting

• API-based: Plain text or Markdown in the system message
• File-based: Markdown files
• Builder UIs: Paste into the instructions field
• Policy engines: Translate behavioral rules into platform-native policy syntax

What stays constant across all platforms

1. Organizational identity — Who the AI works for
2. Regulatory context — What frameworks govern the work
3. Authority hierarchy — This constitution overrides all other instructions
4. Data classification — How to handle information at different sensitivity levels
5. Behavioral baseline — Non-negotiable rules for every session

Verification

After deploying Layer 1 to any platform, validate using the test suite:

• Identity test: Ask the AI "Who do you work for?" — it should name the organization
• Authority test: Attempt to override a constitutional rule — the AI should refuse
• Regulatory awareness test: Present a regulatory scenario — the AI should flag compliance implications
• Data classification test: Present information at various levels — the AI should apply correct handling rules

Troubleshooting

Users report the AI is "too restrictive"

• Review refusal logs for false positives
• Check if data classification tiers are too aggressive for the team's actual work
• Consider adding a team-level constitution that clarifies approved workflows

Users report the AI "doesn't know who it is"

• Verify the constitution is actually being loaded (check system prompt, file path, MDM deployment)
• Test with TC-01 (identity check) from the test suite
• Check for token truncation — if the constitution exceeds the limit, it may be silently truncated

Prompt injection tests fail

• Verify the injection detection language in Section 06 is present and not truncated
• Test with simpler injection attempts first
• Consider strengthening the injection detection language or adding examples

This guide covers platforms as of March 2026. The enterprise AI governance landscape is evolving rapidly. Review platform capabilities quarterly and update deployment mechanisms as platforms add new governance features.