ARC-010·ARC Track·Advanced·60–240 hrs saved

Stop AI From Scaling Faster Than Your Controls — Design an Enterprise AI Architecture That Governs Every Model, Agent, and Retrieval Path

A risk-based operating model, reference architecture, and governance framework for adopting AI at scale while preserving accountability, data protection, security, and financial control.

3Phases

18Quick wins

60–240Hours saved

11Deliverables

Executive Brief

Your Challenge

Your organization is adopting AI faster than it can govern it. AI enters through departmental experiments, employee use of public tools, vendor-embedded copilots, custom integrations, and acquired technology — and each entry point arrives without a shared inventory, owner, or control. The result is an enterprise capability forming by accident: you cannot reliably say which AI systems exist, which models they call, what data they reach, or who is accountable when one of them acts.

Common Obstacles

AI systems scale faster than controls, and the failure modes compound. Business units duplicate platforms and contract directly with providers. Retrieval systems index shared repositories without preserving source permissions, so users receive information they were never authorized to see. Agents run on broadly privileged service accounts with no transaction limits or human approval. Model evaluations rely on vendor benchmarks and demonstrations rather than representative enterprise data. Costs aggregate by cloud subscription, making unit economics impossible. Legal and compliance reviews arrive too late, and pilots become permanent production without ever passing a gate.

The ABME Approach

This workflow builds a durable control plane rather than a single-model deployment. You establish visibility and ownership first — reconcile the AI inventory, assign accountable owners, and classify every use case by risk tier and autonomy level. You then design the enterprise foundation: an AI gateway, model and provider strategy, governed retrieval, prompt and agent architecture, and an evaluation framework grounded in business-specific evidence. Governance is proportional to risk, human accountability stays explicit, and every system gets monitoring, incident response, cost controls, and a retirement path. The AI-assisted prompt pack accelerates assessment and design without ever approving its own risk, evaluation, or production release.

Insight Summary

Enterprise AI architecture should not be organized around a single model, provider, or generation of technology. It is a durable control plane for business outcomes, data access, model choice, evaluation, security, human authority, cost, observability, resilience, and lifecycle.

phase-1

Unknown risk is not low risk. An AI feature is not safe simply because it is embedded in a trusted SaaS product, and a missing inventory record is an ungoverned system, not an absent one.

phase-2

A model should never gain access to information the user is not authorized to see. Retrieval that indexes data without preserving source permissions is a data-exposure incident waiting for a query.

phase-2

Prompt injection cannot be solved by prompt wording alone, and prompt instructions are not a security control. Identity, authorization, isolation, validation, and monitoring are.

phase-3

Vendor benchmarks and successful demonstrations do not establish reliability, safety, or scalability. Enterprise use cases require representative evaluation against defined task thresholds.

tactical

Cost per successful outcome is more meaningful than cost per token, and human review cost is often a material part of the total. Use the smallest model that reliably meets the requirement.

The Journey

Three phases; each lists the tools you'll use there.

Establish Visibility and Control

Reconcile the AI inventory, assign ownership, and classify every use case by risk tier and autonomy level.

Define what qualifies as an AI system
Reconcile the AI inventory across procurement, identity, cloud, API, and SaaS evidence
Identify embedded SaaS AI and shadow AI
Assign business, technical, model, and data owners
Define risk tiers and autonomy levels
Establish governance forums and decision rights

ai-discovery-prompt AI Use-Case Inventory AI System Inventory risk-tier-model autonomy-levels

Build the Enterprise AI Foundation

Stand up the AI gateway, model and provider strategy, governed retrieval, prompt and agent architecture.

Implement an AI gateway with provider abstraction and routing
Establish model registry, prompt registry, and approved model catalog
Implement secure identity-aware retrieval patterns
Establish an agent tool registry with least-privilege permissions
Add use-case cost allocation
Publish reference architectures, standards, patterns, and anti-patterns

Primary AI Architecture Prompt prompt-pack-followups Prompt Governance Record Agent Tool Registry AI Model Registry

Evaluate, Operationalize, and Scale

Prove systems with business-specific evaluation, wire in observability and incident response, and govern the lifecycle.

Implement use-case-specific evaluation suites and red teaming
Add AI observability, monitoring, and incident playbooks
Establish production-readiness gates and material-change review
Test provider failover and kill switches
Measure cost per successful outcome and validate benefits
Retire unused pilots and reassess maturity

Validation Checklist AI Responsibility Matrix maturity-model twelve-month-roadmap

What's Inside the Execution Layer

Numbered deliverables grouped by phase. Membership unlocks every tool.

2. PHASE 2Prompt Packprotected

Primary AI Architecture Prompt

A single comprehensive prompt that assesses current state and designs a full enterprise AI architecture across strategy, governance, models, retrieval, agents, security, cost, and lifecycle.

Use this to

Assess and design an enterprise AI architecture end to end
Separate confirmed facts from assumptions and unknowns
Produce a 38-part deliverable set from executive summary to final recommendations

Included with membership — $20/mo

Primary AI Prompt

Start here with as much of the prerequisite material as you can supply.

You are a chief enterprise architect, chief AI architect, chief data officer advisor, chief information security advisor, responsible AI specialist, model risk manager, privacy architect, cloud architect, integration architect, platform architect, MLOps and LLMOps specialist, FinOps advisor, AI procurement advisor, and enterprise transformation consultant.I will provide some or all of the following:• Business strategy• Technology strategy• Enterprise architecture• Business capability model• Value streams• Application portfolio• Data architecture• Integration architecture• Security architecture• Cloud architecture• AI policies• AI use cases• AI systems• Models• Providers• SaaS applications• Data sources• Systems of record• Prompt libraries• Retrieval architecture• Vector databases• AI agents• Agent tools• Human approval processes• Model evaluations• Security assessments• Privacy assessments• Regulatory obligations• AI incidents• AI costs• Contracts• Identity logs• API logs• Source repositories• Monitoring data• User feedback• Business cases• Benefits records• AI standards• AI exceptions• Business continuity and disaster recovery plansYour task is to assess and design an enterprise AI architecture.Do not assume the AI inventory is complete.Do not assume an AI feature is low risk because it is embedded in a trusted SaaS product.Do not assume the largest, newest, or most expensive model is the best choice.Do not assume model benchmark performance proves business fitness.Do not assume prompt instructions alone provide security.Do not recommend autonomous action without evaluating permissions, reversibility, transaction impact, human oversight, monitoring, and kill-switch capability.Do not infer that an AI output is correct because multiple models agree.Do not treat missing evidence as a passing result.First:1. Summarize: • Business strategy • Technology strategy • Major capabilities • Major value streams • Existing AI use cases • Existing AI platforms • Models and providers • Data sources • Agent use • Current governance • Current security • Current privacy • Current evaluation practices • Current cost • Current maturity2. Separate: • Confirmed facts • Validated evidence • Stakeholder assertions • Inferences • Assumptions • Unknowns3. Identify: • Missing AI systems • Shadow AI • Embedded SaaS AI • Unapproved providers • Unapproved data access • Models without owners • Use cases without business owners • Agents without permission boundaries • Prompts without version control • Retrieval without source authorization • Systems without human oversight • Systems without evaluation • Systems without monitoring • Systems without incident response • Systems without cost controls • High provider concentration • Regulatory exposure • Lifecycle gaps • Retirement gaps4. For each AI use case assess: • Business capability • Business outcome • Users • People affected • Decision impact • Data • Model • Provider • Autonomy • Human oversight • Reversibility • Scale • Security • Privacy • Fairness • Transparency • Evaluation • Monitoring • Cost • Resilience • Vendor dependency • Risk tier • Evidence confidence5. For each AI system assess: • Architecture • Application • Model • Model version • Provider • Prompt • Retrieval • Embeddings • Vector store • Tools • Agent permissions • Identity • Data access • Human approval • Evaluation • Monitoring • Security • Privacy • Cost • Availability • Fallback • Incident response • Lifecycle6. Assess architecture across: • AI gateway • Model gateway • Provider abstraction • Model routing • Retrieval • Knowledge architecture • Prompt management • Agent orchestration • Tool registry • Human approval • Evaluation platform • Observability • Security • Privacy • Cost management • Resilience • Lifecycle7. For each model or provider recommendation provide: • Use case • Required capability • Candidate models • Quality • Security • Privacy • Latency • Availability • Cost • Portability • Provider concentration • Evaluation evidence • Recommended model • Fallback • Confidence • Validation required8. For each agent provide: • Goal • Owner • Model • Tools • Permissions • Data • Autonomy level • Transaction limits • Human approval • Monitoring • Kill switch • Failure behavior • Risk tier • Required controls9. Create: • Enterprise AI principles • AI operating model • AI governance model • AI risk tiers • AI autonomy levels • AI use-case inventory standard • AI system inventory standard • AI reference architecture • Model strategy • Provider strategy • AI gateway strategy • Retrieval architecture • Prompt governance • Agent architecture • Evaluation framework • Security architecture • Privacy architecture • Responsible AI framework • Observability model • Incident-response process • Cost-management model • Lifecycle model • Standards • Patterns • Anti-patterns • Metrics • Dashboards • RoadmapRequirements:• Use risk-based governance.• Keep human accountability explicit.• Include low-risk and high-risk paths.• Preserve “not assessed” when evidence is missing.• Include evidence confidence.• Require business-specific evaluation.• Include model and prompt versioning.• Include retrieval authorization.• Include agent least privilege.• Include transaction and spending limits.• Include kill switches.• Include non-AI fallback where appropriate.• Include provider-failure scenarios.• Include data retention and deletion.• Include embedded SaaS AI.• Include open-weight and self-hosted models.• Include model concentration risk.• Include unit economics.• Include production monitoring.• Include incident response.• Include material-change review.• Include retirement.• State when available evidence does not support a conclusion.• Do not recommend AI where a deterministic or simpler solution is more appropriate.• Do not allow AI-generated analysis to approve its own risk, evaluation, or production release.Then produce:1. Executive summary.2. Current-state AI maturity assessment.3. AI inventory completeness assessment.4. Business capability and use-case map.5. AI risk-tier assessment.6. AI operating model.7. AI governance model.8. Enterprise AI principles.9. Target AI reference architecture.10. AI platform strategy.11. Model and provider strategy.12. AI gateway and routing strategy.13. Data and retrieval architecture.14. Knowledge architecture.15. Prompt governance.16. Agent architecture.17. Human oversight model.18. Evaluation framework.19. Security architecture.20. Privacy architecture.21. Responsible AI framework.22. Observability model.23. Reliability and resilience strategy.24. Incident-response process.25. Cost and unit-economics model.26. Vendor and procurement strategy.27. AI lifecycle and change management.28. AI standards.29. Patterns and anti-patterns.30. AI risk register.31. Metrics.32. Executive dashboard.33. Operational dashboard.34. Quick wins.35. Twelve-month roadmap.36. Three-year evolution roadmap.37. Open questions.38. Final recommendations.

3. PHASE 2Prompt Packprotected

AI Architecture Follow-Up Prompts

Thirty-two targeted follow-up prompts that drill into inventory, risk, model selection, retrieval, agents, evaluation, security, privacy, cost, vendor, and lifecycle work.

Use this to

Run focused assessments after the primary architecture prompt
Design specific components such as an AI gateway, RAG system, or bounded agent
Produce readiness, retirement, and roadmap artifacts on demand

Included with membership — $20/mo

Discover the Enterprise AI Inventory

Reconcile shadow, embedded, and confirmed AI systems.

Discover and reconcile the enterprise AI inventory using:• Procurement• Expenses• Cloud accounts• API logs• Identity logs• OAuth grants• SaaS inventory• Source repositories• Model registries• Data science notebooks• Browser extensions• Department surveys• Vendor questionnaires• Network evidenceIdentify:• Confirmed AI systems• Shadow AI• Embedded SaaS AI• Duplicate records• Unknown owners• Unapproved providers• Inactive pilots• Production systems• Confidence• Validation actions

Create an AI Use-Case Inventory

Build the use-case inventory standard.

Create an AI use-case inventory.For each use case include:• ID• Name• Description• Business capability• Business owner• Product owner• Technical owner• Users• People affected• Decision or action• Data• Model type• Provider• Autonomy• Human oversight• Risk tier• Environment• Status• Cost• Benefits• Evaluation status• Review date

Create an AI System Inventory

Build the system inventory standard.

Create an AI system inventory standard.Include:• System• Application• Use cases• Owners• Platform• Model• Provider• Version• Hosting• Prompt• Retrieval• Embeddings• Vector store• Tools• Agent framework• Permissions• Human approval• Evaluation• Monitoring• Data classification• Regulatory scope• Cost• Lifecycle• Review date

Classify AI Risk

Assign a risk tier to a specific use case.

Classify the supplied AI use case by risk.Assess:• Decision impact• People affected• Customer impact• Employee impact• Financial impact• Health or safety• Regulatory scope• Data sensitivity• Autonomy• Action authority• Reversibility• Scale• Public exposure• Fraud potential• Security impact• Fairness riskProvide:• Risk tier• Rationale• Required governance• Required evaluation• Required human oversight• Required monitoring• Approval authority• Confidence

Assess Whether AI Is Appropriate

Compare AI against simpler alternatives.

Assess whether AI is appropriate for the supplied business problem.Compare:• No technology change• Process improvement• Rules engine• Workflow automation• Search• Analytics• Traditional machine learning• Generative AI• AI agentEvaluate:• Business value• Predictability• Data• Risk• Cost• Explainability• Maintenance• User need• Time to valueRecommend the simplest reliable solution.

Select an AI Model

Choose a primary and fallback model for a use case.

Select an AI model for the supplied use case.Compare candidate models based on:• Task quality• Business evaluation• Security• Privacy• Data terms• Latency• Availability• Context• Modality• Structured output• Tool use• Cost• Portability• Provider concentration• Hosting options• Regulatory requirementsRecommend:• Primary model• Fallback model• Routing rules• Evaluation plan• Change controls

Assess Open-Weight Model Use

Decide on self-hosting or open-weight models.

Assess whether an open-weight or self-hosted model is appropriate.Evaluate:• Data control• Offline requirements• Latency• Specialization• Fine-tuning• Infrastructure• GPU capacity• Security• License• Support• Skills• Cost• Evaluation• Lifecycle• Portability

Design an AI Gateway

Design the enterprise AI gateway.

Design an enterprise AI gateway.Include:• Authentication• Authorization• Provider abstraction• Model routing• Model allowlists• Rate limiting• Quotas• Logging• Redaction• Content controls• Data-loss controls• Cost allocation• Provider failover• Policy enforcement• Resilience• Administration• Metrics

Design a RAG Architecture

Design secure retrieval-augmented generation.

Design a secure retrieval-augmented generation architecture.Include:• Approved sources• Ingestion• Parsing• Chunking• Metadata• Embeddings• Indexing• Access control• Retrieval• Reranking• Prompt assembly• Citations• Freshness• Deletion• Monitoring• Evaluation• Security• Privacy• Cost

Assess Retrieval Security

Review retrieval authorization and leakage risk.

Assess retrieval security.Evaluate:• User identity• Service identity• Source authorization• Document-level access• Metadata filtering• Tenant isolation• Sensitive data• Query logging• Result logging• Revocation• Deletion• Prompt injection• Poisoned content• Cross-user leakage

Create a Prompt Governance Standard

Establish prompt governance.

Create an enterprise prompt governance standard.Include:• Prompt ID• Purpose• Owner• Version• Model• Risk tier• Approved use• Data sources• Tools• Test results• Change control• Effective date• Review date• Retirement

Design an AI Agent

Design a bounded agent.

Design a bounded AI agent.Include:• Goal• Owner• Users• Model• Memory• Tools• Permissions• Data• Planning• Action limits• Transaction limits• Spending limits• Human approval• Logging• Monitoring• Kill switch• Failure behavior• Escalation• Evaluation• Risk tier

Review Agent Permissions

Audit an existing agent's permissions.

Review the supplied AI agent permissions.Identify:• Excessive privileges• Shared credentials• Destructive actions• Missing transaction limits• Missing human approval• Missing destination controls• Missing logging• Missing kill switch• Irreversible actions• Cross-system escalation pathsRecommend least-privilege permissions.

Design Human Oversight

Design meaningful human oversight.

Design meaningful human oversight for the supplied AI system.Define:• Reviewer• Review stage• Evidence shown• Decision authority• Override• Escalation• Response time• Reviewer qualification• Audit record• Appeal• Sampling• Monitoring

Create an AI Evaluation Plan

Build a use-case evaluation plan.

Create an AI evaluation plan.Include:• Business task• Test dataset• Common cases• Edge cases• Adversarial cases• Sensitive cases• Metrics• Thresholds• Human review• Model-based evaluation• Red teaming• Regression testing• Production feedback• Failure actions• Approval

Evaluate a RAG System

Measure a RAG system's quality.

Evaluate the supplied RAG system.Measure:• Retrieval recall• Retrieval precision• Relevance• Freshness• Access correctness• Citation accuracy• Groundedness• Unsupported claims• Answer completeness• Latency• Cost• Security• User satisfaction

Red-Team an AI System

Create a red-team plan.

Create a red-team plan for the supplied AI system.Test:• Direct prompt injection• Indirect prompt injection• Jailbreaks• Data exfiltration• Unauthorized retrieval• Tool abuse• Privilege escalation• Fraud• Harmful output• False citations• Poisoned knowledge• Cross-user leakage• Agent loops• Denial of wallet• Provider failure• Monitoring failureDefine evidence, severity, remediation, and retesting.

Create an AI Threat Model

Build an AI threat model.

Create an AI threat model.Include:• Assets• Actors• Trust boundaries• Entry points• Data flows• Model endpoints• Retrieval• Vector stores• Tools• Agents• Vendors• Telemetry• Threats• Controls• Residual risk• Monitoring• Incident scenarios

Assess AI Privacy

Assess privacy risk.

Assess AI privacy risk.Evaluate:• Personal data• Sensitive data• Purpose• Minimization• Consent• Retention• Training use• Memorization• User disclosure• Data-subject rights• Automated decisions• Cross-border transfer• Vendor processing• Deletion• Logging

Assess AI Fairness

Assess fairness for a use case.

Assess fairness for the supplied AI use case.Evaluate:• People affected• Protected groups• Data representation• Historical bias• Proxy variables• Outcome differences• Error differences• Accessibility• Human review• Appeal• Monitoring• Remediation

Design AI Observability

Design observability with privacy controls.

Design AI observability.Capture:• User• Request• Application• Model• Provider• Model version• Prompt version• Retrieval sources• Tool calls• Agent steps• Output• Latency• Tokens• Cost• Errors• Safety events• Policy decisions• Human approvals• FeedbackDefine privacy, access, retention, and redaction.

Design AI Resilience

Design resilience for a system.

Design AI resilience for the supplied system.Evaluate:• Provider outage• Model outage• Rate limits• Quotas• Latency• Retrieval failure• Tool failure• Context overflow• Malformed output• Model degradation• Regional outageRecommend:• Fallback model• Non-AI fallback• Caching• Queueing• Circuit breakers• Human escalation• Recovery testing

Create an AI Incident-Response Plan

Build the incident-response plan.

Create an AI incident-response plan.Cover:• Data exposure• Harmful output• Unauthorized action• Prompt injection• Tool abuse• Model compromise• Bias event• Privacy complaint• Vendor outage• Model degradation• Runaway cost• Regulatory issueInclude:• Detection• Containment• Kill switch• Evidence• Impact assessment• Notification• Correction• Reevaluation• Restoration• Lessons learned

Calculate AI Unit Economics

Compute unit economics.

Calculate AI unit economics.Include:• Model tokens• Embeddings• Retrieval• Vector storage• Reranking• GPU• Network• Evaluation• Monitoring• Human review• Engineering• Support• SecurityCalculate meaningful measures such as:• Cost per successful task• Cost per user• Cost per case• Cost per customer interaction• Cost per hour saved• Cost per agent actionSeparate confirmed costs, estimates, and unknowns.

Control AI Costs

Build a cost-control plan.

Create an AI cost-control plan.Include:• Budgets• Quotas• Rate limits• Model routing• Smaller models• Caching• Context limits• Output limits• Retrieval optimization• Batch processing• Spending limits• Anomaly detection• Chargeback• Showback• Alerts• Executive reporting

Assess AI Vendor Risk

Assess a vendor or model provider.

Assess the supplied AI vendor or model provider.Evaluate:• Security• Privacy• Data retention• Training use• Region• Subprocessors• Model updates• Availability• Support• Evaluation transparency• Intellectual property• Indemnification• Incident notification• Cost• Portability• Exit• Concentration risk

Review Embedded SaaS AI

Review AI embedded in a SaaS product.

Review embedded AI within the supplied SaaS product.Determine:• Whether AI can be disabled• Models and providers used• Data accessed• Training use• Prompt retention• Processing regions• Administrative controls• User disclosure• Logging• Security• Privacy• Cost• Contract impact• Exit

Create an AI Production Readiness Review

Gate a system for production.

Create an AI production-readiness review.Validate:• Business owner• Technical owner• Model owner• Data owner• Risk tier• Architecture• Evaluation• Security• Privacy• Responsible AI• Human oversight• Monitoring• Incident response• Resilience• Cost controls• Support• Change management• Lifecycle• Retirement

Assess a Material AI Change

Determine whether a change is material.

Assess whether the proposed AI change is material.Changes may include:• Model• Model version• Provider• Prompt• Retrieval• Data source• Embedding model• Tool• Agent permission• Autonomy• User population• Geography• Decision impactDefine required reevaluation, review, testing, and approval.

Create an AI Retirement Plan

Retire a system safely.

Create an AI system retirement plan.Include:• User migration• Model access removal• API-key revocation• Prompt archival• Retrieval shutdown• Vector-data disposition• Tool removal• Agent removal• Contract changes• Data retention• Records retention• Monitoring closure• Cost validation• Inventory update• Business approval

Build an Enterprise AI Roadmap

Produce a three-year roadmap.

Create a three-year enterprise AI architecture roadmap.Include:• Governance• Inventory• AI gateway• Approved models• Data access• Retrieval• Agent platform• Evaluation• Security• Privacy• Responsible AI• Observability• Cost management• Workforce enablement• Use-case waves• Benefits• Dependencies• Investment• Maturity targets

4. PHASE 1Matrixprotected

AI Use-Case Inventory

A standardized register that captures every AI use case with its ownership, risk, autonomy, cost, and evaluation status.

Use this to

Record every AI use case in a consistent format
Assign business, product, and technical owners
Track risk tier, autonomy, cost, and review date per use case

Included with membership — $20/mo

Use-Case ID	Name	Business Capability	Business Owner	Model / Provider	Autonomy Level	Risk Tier	Lifecycle State	Review Date

RubricCapture for each use case: Use-case ID, Name, Description, Business capability, Business owner, Product owner, Technical owner, Users, Customers affected, Decision or action supported, Data used, Model type, Provider, Autonomy level, Risk tier, Human oversight, Environment, Status, Cost, Benefits, Evaluation status, Security status, Privacy status, Lifecycle state, Review date.

5. PHASE 1Matrixprotected

AI System Inventory

A standardized register that captures every AI system with its models, retrieval, tools, permissions, and lifecycle state.

Use this to

Record every AI system with its full component stack
Track model versions, providers, and hosting models
Maintain owners, regulatory scope, and last-review dates

Included with membership — $20/mo

System ID	System Name	Model Provider	Model Version	Hosting Model	Permissions	Data Classification	Lifecycle	Last Review

RubricCapture for each system: System ID, System name, Application, Use cases, Business owner, Technical owner, Model owner, Data owner, Platform, Model provider, Model name, Model version, Hosting model, Prompt version, Retrieval sources, Embedding model, Vector database, Tools, Agent framework, Permissions, Human approval, Evaluation suite, Monitoring, Data classification, Regulatory scope, Vendor contract, Cost center, Lifecycle, Last review.

6. PHASE 2Matrixprotected

Agent Tool Registry

A registry that records every agent tool with its permissions, allowed actions, transaction limits, and revocation process.

Use this to

Document each tool's purpose, owner, and permissions
Set transaction limits and human-approval requirements per tool
Define logging, failure behavior, and revocation for every action

Included with membership — $20/mo

Tool ID	Name	Purpose	Owner	Permissions	Allowed Actions	Risk Tier	Transaction Limits	Human Approval

RubricEach tool should record: Tool ID, Name, Purpose, Owner, API, Permissions, Allowed actions, Data accessed, Risk tier, Transaction limits, Human approval requirements, Logging, Failure behavior, Revocation process.

7. PHASE 2Matrixprotected

AI Model Registry

A registry that records every approved model with its license, approved and prohibited use, evaluation results, and lifecycle.

Use this to

Track approved, restricted, and prohibited model use
Record evaluation, security, and privacy review status
Manage model owners, cost, effective dates, and replacements

Included with membership — $20/mo

Model ID	Provider	Model Name	Version	License	Approved Use	Prohibited Use	Owner	Review Date

RubricCapture: Model ID, Provider, Model name, Version, Type, License, Hosting, Approved use, Restricted use, Prohibited use, Data classifications, Evaluation results, Security review, Privacy review, Cost, Owner, Lifecycle, Effective date, Review date, Replacement.

8. PHASE 2Templateprotected

Prompt Governance Record

A labeled record for governing a material prompt through versioning, approval, testing, and review.

Use this to

Version-control material prompts
Document approved use, model, and risk tier
Track test results, change history, and review dates

Included with membership — $20/mo

Prompt ID

Unique identifier for the prompt.

[...]

Purpose

What the prompt is designed to accomplish.

[...]

Owner

Accountable owner for this prompt.

[...]

Version

Current version; material prompts should be version-controlled.

[...]

Model

Model the prompt is approved against.

[...]

Risk Tier

Risk tier of the use case this prompt serves.

[...]

Approved Use

Where and how this prompt may be used.

[...]

Test Results

Evaluation and regression test results.

[...]

Change History

Record of prior changes.

[...]

Effective Date

Date this version takes effect.

[...]

Review Date

Next scheduled review.

[...]

Related Tools

Tools associated with this prompt.

[...]

Related Data Sources

Data sources accessed by this prompt.

[...]

9. PHASE 3Matrixprotected

AI Responsibility Matrix

A RACI-style matrix assigning accountability across activities from AI strategy through retirement.

Use this to

Assign accountable, responsible, consulted, and informed roles per activity
Clarify decision rights across architecture, security, privacy, finance, and procurement
Confirm no activity lacks an accountable owner

Included with membership — $20/mo

Reference rows from the blueprint — downloads ship as an empty skeleton

Activity	Executive Sponsor	Enterprise Architecture	Business Owner	AI Architect	Model Owner	Data Owner	Security	Privacy/Legal	Finance	Procurement
AI strategy	Accountable	Responsible	Consulted	Consulted	Consulted	Consulted	Consulted	Consulted	Consulted	Informed
Use-case approval	Accountable for material use	Consulted	Responsible	Consulted	Consulted	Consulted	Consulted	Consulted	Consulted	Informed
Risk classification	Informed	Consulted	Consulted	Responsible	Consulted	Consulted	Responsible	Responsible	Informed	Informed
Architecture approval	Informed	Accountable	Consulted	Responsible	Consulted	Consulted	Consulted	Consulted	Informed	Informed
Model approval	Informed	Governs	Consulted	Responsible	Accountable	Consulted	Consulted	Consulted	Consulted	Consulted
Data-source approval	Informed	Consulted	Consulted	Consulted	Informed	Accountable	Consulted	Responsible where applicable	Informed	Informed
Evaluation	Informed	Governs	Validates outcome	Responsible	Accountable	Consulted	Consulted	Consulted	Informed	Informed
Security approval	Informed	Consulted	Consulted	Consulted	Consulted	Consulted	Accountable	Consulted	Informed	Informed
Privacy approval	Informed	Consulted	Consulted	Consulted	Consulted	Responsible	Consulted	Accountable	Informed	Informed
Cost approval	Accountable for material spend	Consulted	Responsible	Consulted	Informed	Informed	Informed	Informed	Accountable	Consulted
Vendor selection	Informed	Consulted	Consulted	Consulted	Consulted	Consulted	Consulted	Consulted	Consulted	Accountable
Agent permissions	Informed	Governs	Approves business scope	Responsible	Consulted	Consulted	Accountable for security	Consulted	Informed	Informed
Production approval	Accountable for critical systems	Responsible	Responsible	Responsible	Responsible	Consulted	Responsible	Responsible	Consulted	Informed
Incident response	Informed	Consulted	Responsible for business impact	Responsible	Responsible	Consulted	Accountable for security incidents	Responsible for legal/privacy incidents	Consulted	Consulted
Retirement	Informed	Consulted	Accountable	Responsible	Responsible	Responsible for data	Consulted	Consulted	Validates cost closure	Responsible for contract closure

🔒 The full execution layer — every checklist, matrix, and the prompt pack — is included with ABME membership.

Unlock Full Blueprint

Full Playbook

Overviewpublic

Enterprise AI architecture defines how artificial intelligence is selected, built, integrated, secured, governed, operated, monitored, funded, and retired across the organization.

It covers more than models.

A complete AI architecture includes:

Business use cases
AI products
Machine learning models
Foundation models
Large language models
Small language models
Multimodal models
Embedding models
Rerankers
Retrieval systems
Knowledge sources
Vector stores
Prompt libraries
AI gateways
Agent frameworks
Tools and actions
Human approval
Evaluation
Monitoring
Security
Privacy
Data governance
Model risk
Cost controls
Resilience
Vendor management
Lifecycle management

The objective is not simply to deploy AI.

The objective is to create an enterprise capability that produces useful, trustworthy, secure, governable, economically sustainable, and measurable outcomes.

A mature enterprise AI architecture should answer:

“How can the organization use AI at scale while preserving accountability, data protection, security, reliability, human authority, and financial control?”

Business Problempublic

Organizations frequently adopt AI through:

Departmental experimentation
Employee use of public AI tools
Vendor-embedded AI
SaaS copilots
Custom model integrations
Data science projects
Executive mandates
Customer-facing pilots
Robotic process automation
AI agents
Unapproved browser extensions
Shadow APIs
Embedded third-party models
Acquired technologies

This often creates:

Unknown AI inventory
Inconsistent model selection
Sensitive data exposure
Unapproved vendors
Duplicate AI platforms
Unbounded agent permissions
Weak human oversight
Poor prompt governance
Uncontrolled retrieval
Hallucinated outputs
Inconsistent evaluations
Model drift
Cost volatility
Vendor concentration
Weak observability
Unclear accountability
Inadequate incident response
Regulatory exposure
Customer trust risk
AI pilots that never become production capabilities
Production systems that remain governed like pilots

Without an enterprise AI architecture:

AI systems scale faster than controls.
Business units duplicate platforms.
Data access becomes inconsistent.
AI-generated decisions lack accountability.
Models change without impact assessment.
Costs become difficult to forecast.
Security teams lack visibility.
Legal and compliance reviews occur too late.
Human users overtrust or underuse AI.
Valuable use cases remain blocked by unclear governance.

Expected Outcomepublic

After completing this workflow, the organization should have:

Enterprise AI vision
AI architecture principles
AI operating model
AI governance model
AI use-case inventory
AI system inventory
AI risk-tier model
AI reference architecture
Model strategy
Provider strategy
Model gateway strategy
AI platform strategy
Data and retrieval architecture
Knowledge architecture
Agent architecture
Human oversight model
Prompt governance
Model evaluation framework
AI security architecture
AI privacy architecture
Responsible AI framework
AI observability model
AI incident-response process
AI resilience strategy
AI cost-management model
AI vendor-management model
AI lifecycle model
AI standards
AI patterns and anti-patterns
AI portfolio roadmap
Metrics
Dashboards
Maturity model
Twelve-month implementation roadmap
AI-assisted assessment prompts

🔒 The complete playbook — reference models, worked examples, and operational guidance — is included with ABME membership.

Unlock Full Blueprint

Objectivesprotected

The enterprise AI architecture should determine:

Which business capabilities should use AI?
Which use cases should not use AI?
Which AI systems already exist?
Which models and providers are approved?
When should commercial models be used?
When should open-weight or self-hosted models be used?
When should smaller models be preferred?
How should model selection be performed?
How should enterprise data be accessed?
How should retrieval be governed?
How should prompts be managed?
How should AI agents be constrained?
Which actions require human approval?
How should outputs be evaluated?
How should AI systems be monitored?
How should hallucination and error risk be managed?
How should AI costs be forecast and controlled?
How should vendor concentration be managed?
How should model changes be governed?
How should AI incidents be handled?
How should AI systems be retired?
How should regulatory obligations be met?
How should business value be measured?
How should AI architecture integrate with enterprise architecture?

Enterprise AI Scope and Definitionprotected

Enterprise AI Scope

The AI architecture should cover:

Traditional machine learning
Predictive analytics
Classification
Recommendation systems
Computer vision
Speech recognition
Speech synthesis
Natural language processing
Large language models
Generative AI
Multimodal AI
Retrieval-augmented generation
AI copilots
AI assistants
AI agents
Autonomous workflows
Embedded SaaS AI
Customer-facing AI
Employee-facing AI
Decision-support systems
High-impact automated decisions
AI-enabled security tools
AI-enabled development tools
AI-generated content
Synthetic data
AI evaluation tools
AI monitoring tools

AI System Definition

An AI system may include:

Model
Model provider
Hosting platform
Application
Prompt
System instructions
Retrieval source
Embedding model
Vector store
Tools
Agent orchestration
User interface
Human reviewer
Data pipeline
Monitoring
Evaluation
Feedback loop
Business process
External consumer

The system boundary should include all components that materially influence output or action.

AI Architecture Principlesprotected

Recommended principles include:

Business value before novelty
Human accountability remains explicit
AI risk determines governance depth
Data access is governed
Least privilege applies to agents
Models are interchangeable where practical
Evaluation is continuous
Security is built in
Privacy is built in
Outputs are observable
Costs are transparent
Model and provider changes are controlled
AI systems have owners
High-impact actions require appropriate human authority
AI-generated content is treated according to use and risk
AI systems have retirement plans
Unknown risk is not treated as low risk

Principle 1 — AI Must Serve a Defined Business Outcome

Statement: Every production AI system must support a measurable business, customer, employee, operational, risk, or regulatory outcome.

Rationale: AI experimentation without defined outcomes produces fragmented investment and weak accountability.

Implications:

Use cases require named business outcomes.
Success metrics must be defined before production.
Novelty alone is insufficient justification.
AI should be compared with non-AI alternatives.
Business owners remain accountable for outcomes.

Principle 2 — Human Accountability Cannot Be Delegated to a Model

Statement: A model may assist, recommend, summarize, classify, or act within approved boundaries, but accountable authority remains with designated human roles.

Rationale: AI systems cannot accept legal, fiduciary, ethical, operational, or executive responsibility.

Implications:

Decision ownership must be explicit.
Human review must match consequence and reversibility.
Escalation paths must exist.
Users must understand when AI is involved.
High-impact decisions require defined oversight.

Principle 3 — Governance Is Proportional to AI Risk

Statement: AI systems should receive governance, testing, monitoring, and oversight proportional to their potential impact.

Rationale: A low-risk drafting assistant should not require the same controls as an autonomous clinical, financial, employment, or safety decision system.

Implications:

Risk tiers must be defined.
Higher-risk systems require stronger evidence.
Autonomy, data sensitivity, and reversibility influence risk.
Governance should remain lightweight for low-risk use cases.
Unknown risk should trigger additional assessment.

Principle 4 — Enterprise Data Is Accessed Through Governed Channels

Statement: AI systems must access enterprise data through approved, observable, least-privilege interfaces and repositories.

Rationale: Uncontrolled AI access creates security, privacy, data-quality, and lineage risk.

Implications:

Data sources require approval.
Retrieval must be logged.
Access should follow user and service identity.
Sensitive sources require additional controls.
Direct unrestricted data access should be prohibited.

Principle 5 — Agent Permissions Are Explicit and Bounded

Statement: AI agents may act only through approved tools, permissions, data boundaries, transaction limits, and human approval rules.

Rationale: An agent with broad authority can create operational, security, financial, or legal harm.

Implications:

Each tool requires documented purpose.
Permissions must follow least privilege.
Destructive actions require stronger control.
Transaction thresholds should be defined.
Actions must be auditable.
Kill switches must exist for material agents.

Principle 6 — Evaluation Is a Lifecycle Requirement

Statement: AI systems must be evaluated before release and continuously after deployment using use-case-specific evidence.

Rationale: Model performance varies by task, data, provider, configuration, and time.

Implications:

Generic benchmark scores are insufficient.
Test sets must reflect business use.
Quality thresholds must be defined.
Regression testing is required.
Model changes trigger reevaluation.
Production feedback must inform testing.

Principle 7 — Model Choice Is Based on Fitness, Risk, and Economics

Statement: Models should be selected based on task performance, security, privacy, latency, reliability, portability, and total cost rather than brand preference or raw model size.

Rationale: The largest or newest model is not always the best enterprise choice.

Implications:

Multiple models may be appropriate.
Smaller models should be considered.
Model routing may be used.
Cost per successful outcome should be measured.
Provider concentration should be assessed.

Principle 8 — AI Systems Are Observable by Design

Statement: AI systems must generate sufficient evidence to understand requests, data access, model use, output quality, actions, cost, and failure.

Rationale: AI systems cannot be governed when their behavior is invisible.

Implications:

Requests and responses may require logging.
Sensitive content must be protected.
Model and prompt versions must be recorded.
Tool actions require traceability.
Evaluation and cost metrics must be available.

Principle 9 — AI Security Is End-to-End

Statement: Security controls must cover the complete AI system, including model, prompt, data, retrieval, tools, plugins, infrastructure, vendors, outputs, and users.

Rationale: AI risk is not limited to the model endpoint.

Implications:

Threat modeling is required.
Prompt injection must be addressed.
Data exfiltration controls are required.
Tool invocation must be authorized.
Model supply-chain risk must be assessed.
Output handling must be secured.

Principle 10 — AI Has a Managed Lifecycle

Statement: Every AI system must have an owner, inventory record, approved lifecycle state, change process, monitoring plan, and retirement path.

Rationale: AI systems can become unsupported, inaccurate, uneconomic, or noncompliant as models, providers, data, and business needs change.

Implications:

Systems must be inventoried.
Model versions must be tracked.
Review dates are required.
Retirement includes data, prompts, integrations, and access.
Transitional pilots must have expiration criteria.

Operating Model and Governanceprotected

AI Operating Model

The operating model should define:

Executive sponsorship
Business ownership
Product ownership
Architecture ownership
Model ownership
Data ownership
Platform ownership
Security responsibilities
Privacy responsibilities
Risk responsibilities
Legal responsibilities
Financial ownership
Procurement responsibilities
Human oversight
Incident response
Lifecycle management

AI Operating Model Options

Centralized — A central AI team controls platforms, models, governance, delivery, and standards. Advantages: consistency, concentrated expertise, strong control. Risks: bottlenecks, weak domain alignment, slow adoption.

Federated — Business units independently select and implement AI. Advantages: speed, domain relevance, local ownership. Risks: duplication, inconsistent control, vendor proliferation, weak enterprise visibility.

Hub-and-Spoke — A central AI enablement function provides platforms, standards, model access, governance, reference architectures, and evaluation tooling. Domain teams own use cases, products, workflows, business outcomes, and domain evaluations. This is often the most scalable enterprise model.

AI Governance Structure

Possible governance forums include:

Executive AI Council — AI strategy, enterprise priorities, risk appetite, major investments, high-impact AI, public trust, regulatory exposure.

AI Architecture Council — Platforms, models, reference architectures, integration, data access, agents, lifecycle, standards.

Responsible AI and Model Risk Review — Fairness, impact, explainability, evaluation, human oversight, regulatory requirements, model risk.

AI Security Review — Threat modeling, prompt injection, tool security, data exfiltration, vendor security, identity, logging, incident response.

AI Product Review — User need, outcome, experience, adoption, quality, feedback, benefit realization.

AI Risk and Autonomy Modelsprotected

AI Risk-Tier Model

Risk should consider:

Decision impact
Customer impact
Employee impact
Financial impact
Health or safety impact
Regulatory scope
Data sensitivity
Autonomy
Action authority
Reversibility
Scale
Public exposure
Model opacity
Dependency
Fraud potential
Security impact
Potential for discrimination
Potential for misinformation

Example Risk Tiersprotected

Tier 1 — Low Risk

Examples: drafting internal text, summarizing non-sensitive documents, brainstorming, coding assistance in controlled environments, low-impact search support.

Approved platform
Data-use rules
Basic logging
User education
Periodic review

Tier 2 — Moderate Risk

Examples: employee productivity copilots, internal knowledge assistants, customer-service drafting, operational recommendations, classification of business records.

Use-case approval
Data-source review
Evaluation
monitoring
feedback
ownership
escalation

Tier 3 — High Risk

Examples: customer-facing recommendations, fraud detection, financial decisions, employment support, clinical decision support, security response recommendations, material contract analysis.

Formal risk assessment
documented human oversight
independent evaluation
security testing
privacy review
monitoring
model-change control
incident response

Tier 4 — Critical or Restricted

Examples: autonomous high-impact decisions, life-safety control, fully automated employment decisions, unreviewed legal determinations, autonomous financial transfers, unbounded production administration, systems prohibited by law or policy.

Executive and legal approval
formal assurance
strict human authority
strong technical constraints
continuous monitoring
independent audit
prohibition where appropriate

Autonomy Levelsprotected

Level 0 — Informational

The AI provides information without recommendation or action.

Level 1 — Assistive

The AI drafts, summarizes, or organizes content.

Level 2 — Advisory

The AI recommends an action, but a human decides.

Level 3 — Supervised Action

The AI prepares or initiates an action that requires human approval.

Level 4 — Bounded Autonomous Action

The AI acts within predefined limits, permissions, and monitoring.

Level 5 — Broad Autonomous Action

The AI acts across systems with minimal intervention. Level 5 should be rare and subject to the strongest governance.

AI Reference Architectureprotected

AI Reference Architecture

A typical enterprise AI architecture may include:

User channel
Application layer
Identity and access
AI gateway
Policy enforcement
Prompt management
Model router
Model providers
Retrieval orchestration
Search
Vector database
Knowledge graph
Enterprise data services
Tool registry
Agent orchestration
Human approval service
Evaluation service
Safety filters
Logging
Monitoring
Cost management
Incident controls
Model registry
Feedback service

User and Channel Layer

Channels may include:

Web
Mobile
Collaboration tools
Contact center
Voice
Developer tools
Internal portals
Customer portals
APIs
Embedded applications

Each channel should address:

Authentication
User identity
disclosure
accessibility
data handling
session management
feedback
escalation

AI Gateway

An AI gateway may provide:

Central model access
Authentication
Authorization
Provider abstraction
Model routing
Rate limiting
Quotas
Logging
Redaction
Content filtering
Cost allocation
Prompt policies
Provider failover
Model allowlists
Data-loss controls
Usage analytics

An AI gateway should not become a single point of failure without appropriate resilience.

Model Gateway and Routing

Model routing may consider:

Use case
Quality
Risk tier
Data sensitivity
Region
Latency
Cost
context length
modality
availability
provider policy
model version
fallback

Routing policies should be tested and governed.

Model Strategyprotected

Model Strategy

The model strategy should determine:

Approved model categories
Approved providers
Preferred models
Restricted models
Self-hosting criteria
Open-weight criteria
Fine-tuning criteria
Small-model criteria
Multimodel strategy
Routing strategy
Fallback strategy
Version management
Portability
Exit

Commercial Model Use

Commercial models may be appropriate when:

Time to value is important.
Managed security and availability are sufficient.
Model capability is strong.
Data terms are acceptable.
Portability risk is manageable.
Cost is justified.

Assess:

Data retention
training use
region
subprocessor use
security
availability
model change
intellectual property
indemnification
incident notification
exit support

Open-Weight and Self-Hosted Models

These may be appropriate when:

Data control is critical.
Offline operation is required.
Specialized fine-tuning is valuable.
Latency requires local execution.
Provider dependency is unacceptable.
Unit economics favor internal hosting.
Edge deployment is required.

Assess:

License
model provenance
security
infrastructure
GPU capacity
patching
evaluation
support
operational skills
lifecycle
total cost

Small Language Models

Small models may provide:

Lower latency
Lower cost
Easier local hosting
Better task specialization
Reduced data exposure
More predictable output

Use the smallest model that reliably meets the business requirement.

Fine-Tuning Strategy

Fine-tuning may be appropriate when:

The task is stable.
Training examples are available.
Behavior cannot be achieved reliably through prompting or retrieval.
Latency or cost improves materially.
Evaluation demonstrates value.

Avoid fine-tuning merely to inject current enterprise knowledge that would be better managed through retrieval.

Retrieval and Knowledge Architectureprotected

Retrieval-Augmented Generation

RAG architecture should define:

Source approval
content ingestion
parsing
chunking
metadata
embeddings
indexing
access control
retrieval
reranking
prompt assembly
citations
freshness
deletion
monitoring
evaluation

Retrieval Security

Controls should include:

Identity-aware retrieval
document-level access
source-level authorization
metadata filtering
tenant isolation
restricted collections
sensitive-data detection
query logging
result logging
access revocation
deletion propagation

A model should not gain access to information the user is not authorized to see.

Retrieval Quality

Measure:

Recall
precision
relevance
freshness
citation accuracy
answer grounding
unsupported claims
source coverage
access correctness

Knowledge Architecture

Knowledge sources may include:

Documents
databases
data products
APIs
knowledge graphs
search indexes
ticketing systems
collaboration platforms
records repositories
policy libraries
product catalogs
customer records

The architecture should define:

Source authority
ownership
metadata
freshness
retention
access
provenance
quality
deletion

Vector Database Architecture

Assess:

Data sensitivity
access model
tenant isolation
encryption
region
index strategy
metadata filtering
deletion
backup
recovery
cost
portability
vendor lock-in

Embeddings may reveal sensitive semantic information and should be protected accordingly.

Prompt and Output Architectureprotected

Prompt Architecture

Prompt architecture includes:

System instructions
user instructions
task templates
contextual data
retrieved content
tool descriptions
output schemas
safety instructions
escalation rules

Prompt Governance

Capture:

Prompt ID
Purpose
Owner
Version
Model
Risk tier
Approved use
Test results
Change history
Effective date
Review date
Related tools
Related data sources

Material prompts should be version-controlled.

Structured Outputs

Structured outputs should be used where:

Downstream automation depends on the result.
Validation is possible.
Data contracts are required.
Agent actions are triggered.
Consistency is important.

Controls may include:

JSON schema
type validation
allowed values
required fields
confidence
error handling
rejection
retry

Agent Architectureprotected

Agent Architecture

An AI agent may include:

Goal
model
memory
tools
planning
execution
observation
approval
constraints
stopping conditions
escalation
logging

Agent Identity

Agents should use:

Unique service identity
Least privilege
short-lived credentials
scoped tokens
approved secrets management
explicit impersonation controls
user delegation where appropriate

Agents should not use shared privileged accounts.

Agent Memory

Memory may include:

Session memory
user preferences
task state
long-term knowledge
prior actions
retrieved context

Assess:

Purpose
consent
sensitivity
retention
accuracy
deletion
access
poisoning
cross-user leakage

Agent Guardrails

Guardrails may include:

Tool allowlists
action allowlists
transaction thresholds
approval gates
spending limits
rate limits
time limits
step limits
scope limits
destination restrictions
content validation
policy checks
kill switch
rollback
sandboxing

Human Oversight Designprotected

Human-in-the-Loop Design

Human oversight should define:

Who reviews?
What is reviewed?
At what stage?
What evidence is shown?
Can the reviewer override?
Is the reviewer qualified?
Is there adequate time?
Is review recorded?
What happens after disagreement?
How is escalation handled?

A nominal approval button is not meaningful oversight if the user cannot understand or challenge the output.

Human-on-the-Loop Design

For bounded autonomous systems, humans may supervise rather than approve every action.

Requirements may include:

Real-time monitoring
anomaly alerts
sampling
transaction limits
kill switch
retrospective review
escalation
incident handling

AI Evaluation Frameworkprotected

AI Evaluation Framework

Evaluation should include:

Functional quality
task success
groundedness
factuality
relevance
completeness
consistency
safety
bias
privacy
security
robustness
latency
cost
user experience
business outcome

Evaluation Dataset

The dataset should include:

Representative cases
common cases
difficult cases
edge cases
adversarial cases
sensitive cases
failure cases
multilingual cases where relevant
accessibility cases
prohibited cases
historical incidents

Evaluation Methods

Methods may include:

Deterministic checks
Human review
expert review
model-based evaluation
pairwise comparison
rubric scoring
benchmark comparison
simulation
red teaming
production A/B testing
shadow deployment

Model-based evaluation should itself be validated and should not be the sole evidence for high-risk systems.

Evaluation Metrics

Possible metrics include:

Task completion
precision
recall
accuracy
false positive rate
false negative rate
grounded response rate
citation accuracy
hallucination rate
harmful response rate
policy violation rate
escalation rate
reviewer agreement
user acceptance
cost per successful task
latency
abandonment

Evaluation Thresholds

For each metric define:

Baseline
minimum threshold
target
owner
test frequency
failure action
exception process

Regression Testing

Trigger reevaluation when:

Model changes
provider changes
prompt changes
retrieval changes
data changes
tool changes
policy changes
user population changes
business process changes
incidents occur

Red Teaming

Red-team scenarios may include:

Prompt injection
indirect prompt injection
data exfiltration
unauthorized retrieval
jailbreaks
tool abuse
privilege escalation
fraud
harmful content
hallucinated authority
false citations
adversarial files
poisoned knowledge
hidden instructions
cross-user leakage
denial of wallet
denial of service
agent loops

AI Security Architectureprotected

AI Security Architecture

Security should cover:

Identity
access
data
prompts
model endpoints
model files
retrieval
vector stores
tools
plugins
agents
infrastructure
vendors
outputs
telemetry
supply chain

AI Threat Model

Threat actors may include:

External attackers
malicious users
insiders
compromised vendors
compromised data sources
malicious documents
model supply-chain attackers
abusive customers
autonomous agents behaving unexpectedly

Prompt Injection

Controls may include:

Treat external content as untrusted.
Separate instructions from data.
Restrict tool permissions.
Validate intended actions.
Require human approval.
Filter retrieved content.
Use structured tool calls.
Confirm transaction context.
Monitor unusual behavior.
Red-team indirect injection.

Prompt injection cannot be solved by prompt wording alone.

Data Exfiltration Controls

Use:

Least-privilege retrieval
output filters
data classification
destination controls
tenant isolation
logging
anomaly detection
restricted tools
redaction
policy enforcement
human approval

Model Supply-Chain Security

Assess:

Model source
license
provenance
signatures
checksums
dependencies
training claims
known vulnerabilities
malicious behavior
update process
hosting
access
vendor controls

AI Privacy Architectureprotected

AI Privacy Architecture

Evaluate:

Personal data
sensitive personal data
consent
purpose
minimization
retention
training use
model memorization
user disclosure
deletion
data-subject rights
cross-border transfer
automated decision requirements
vendor processing
synthetic data

Privacy-Preserving Techniques

Possible techniques include:

Data minimization
redaction
tokenization
pseudonymization
aggregation
differential privacy
synthetic data
secure enclaves
local inference
federated learning
confidential computing

Appropriate use depends on the use case and threat model.

Responsible AIprotected

Responsible AI

Responsible AI should assess:

Validity
reliability
safety
security
resilience
transparency
explainability
privacy
fairness
accountability
human oversight
contestability

Fairness Assessment

Where relevant, evaluate:

Protected groups
representation
outcome differences
error differences
proxy variables
historical bias
feedback loops
accessibility
remediation

Fairness requirements should be defined with legal, business, domain, and community context.

Explainability

Determine what explanation is required for:

Developers
operators
reviewers
users
customers
regulators
auditors
impacted individuals

Different audiences need different explanations.

Transparency

Transparency may include:

Disclosure that AI is used
Purpose
data use
limitations
human oversight
appeal process
contact
content labeling
source citations
confidence or uncertainty

Contestability and Appeal

High-impact systems should define:

How an individual challenges an output
Human review
correction
escalation
response time
evidence
recordkeeping
remediation

AI Observabilityprotected

AI Observability

Observe:

User request
application
model
provider
model version
prompt version
retrieved sources
tool calls
agent steps
output
latency
token use
cost
errors
safety events
policy decisions
human approvals
feedback

Telemetry Privacy

AI telemetry may contain:

Personal information
confidential business information
source documents
credentials
regulated data
intellectual property

Define:

Redaction
retention
access
encryption
sampling
legal hold
deletion
monitoring use

AI Reliability, Resilience, and Continuityprotected

AI Reliability

Assess:

Availability
latency
timeout
provider outage
rate limits
quota
context overflow
retrieval failure
tool failure
malformed output
model refusal
model degradation
regional disruption

AI Resilience Patterns

Patterns may include:

Provider failover
model fallback
smaller-model fallback
cached responses
graceful degradation
non-AI fallback
human escalation
queueing
circuit breakers
retry
bulkheads
regional redundancy
local inference

Fallback models should be evaluated for the same use case before activation.

AI Business Continuity

Define:

Critical AI functions
maximum tolerated outage
manual fallback
alternate process
provider dependency
data dependency
recovery procedures
model restoration
prompt restoration
vector-index restoration
credential recovery
vendor communication

AI Incident Responseprotected

AI Incident Response

Incident categories may include:

Data exposure
harmful output
unauthorized action
hallucinated material decision
model compromise
prompt injection
tool abuse
bias event
privacy complaint
vendor outage
model degradation
runaway cost
intellectual-property concern
regulatory issue

AI Incident Process

Include:

Detect
Contain
Disable or limit
Preserve evidence
Assess impact
Notify owners
Notify legal, privacy, security, or regulators where required
Correct
Reevaluate
Restore
Review
Update controls

Kill Switches

Material AI systems should provide the ability to:

Disable a model
disable a prompt
disable retrieval
disable tools
disable an agent
reduce permissions
switch to read-only
require all-human approval
route to fallback
suspend a provider

AI Cost Architectureprotected

AI Cost Architecture

Cost components may include:

Model tokens
image generation
speech
embeddings
reranking
vector storage
retrieval
data processing
GPU infrastructure
network
observability
evaluation
human review
engineering
support
security
vendor platform fees

AI Unit Economics

Possible measures include:

Cost per user
cost per active user
cost per request
cost per successful outcome
cost per document
cost per case
cost per transaction
cost per customer interaction
cost per agent action
cost per hour saved
cost per dollar of benefit

Cost per successful outcome is often more meaningful than cost per token.

AI Cost Controls

Use:

Budgets
quotas
rate limits
model routing
caching
context limits
prompt optimization
retrieval optimization
smaller models
batch processing
reserved capacity
usage alerts
chargeback
showback
anomaly detection
agent spending limits

Denial-of-Wallet Protection

Controls may include:

Request limits
user quotas
tool limits
recursion limits
context limits
output limits
budget thresholds
anomaly alerts
authentication
abuse detection
kill switches

Vendor, Provider, and Procurementprotected

Vendor and Provider Strategy

Assess providers based on:

Model capability
reliability
security
privacy
data terms
training use
region
compliance
latency
cost
model roadmap
support
observability
indemnification
portability
exit
concentration risk

Provider Concentration

Identify dependency on:

Single model provider
single cloud provider
proprietary vector store
proprietary orchestration
proprietary evaluation tooling
unique model behavior
provider-specific prompts
provider-specific tools

Mitigation may include:

Abstraction
multimodel evaluation
portable prompts
standard APIs
export
fallback providers
contractual protection

AI Procurement

Procurement should assess:

AI functionality
embedded models
provider chain
data use
training use
retention
subprocessors
security
privacy
evaluation
model updates
explainability
human oversight
audit rights
incident notification
service levels
cost
export
termination
data deletion

Embedded AI in SaaS

Evaluate:

Whether AI can be disabled
What data it accesses
Whether customer data trains models
Which provider is used
Where processing occurs
How outputs are logged
Whether prompts are retained
Whether administrators can control use
Whether users are notified
Whether the feature changes contract scope

AI Lifecycle and Change Managementprotected

AI Lifecycle States

AI Lifecycle Gates

AI Change Management

Material changes include:

New model
new model version
new provider
new prompt
new knowledge source
new embedding model
new tool
increased permission
increased autonomy
new user group
new geography
new regulated data
changed decision impact

Standards, Patterns, and Anti-Patternsprotected

AI Standards

Standards should address:

Use-case registration
Risk classification
Model approval
Provider approval
Data access
Retrieval
prompts
agents
tools
human oversight
evaluation
security
privacy
fairness
transparency
observability
cost
incident response
change management
retirement

AI Architecture Patterns

Recommended patterns may include:

Enterprise AI gateway
Retrieval-augmented assistant
Secure internal copilot
Customer-service drafting assistant
Human-approved agent action
Bounded workflow agent
Model routing
Sensitive-data local inference
Multimodel review
AI evaluation pipeline
AI content moderation
Human escalation
Non-AI fallback

AI Anti-Patterns

Examples include:

Direct unrestricted model access
Shared model API keys
AI with production administrator rights
Retrieval without source authorization
Prompt-only security
Model output treated as fact
Autonomous action without limits
No model version tracking
No evaluation dataset
No cost controls
Pilot used indefinitely in production
Sensitive data sent to unapproved tools
AI decisions without appeal
Silent provider model changes
Logging all prompts without privacy review

Multimodel Architecture

Multiple models may be used for:

Routing
fallback
verification
specialization
cost optimization
modality
regional requirements
provider resilience

Multimodel design should avoid unnecessary complexity.

Multimodel Review

Independent model review may improve:

Error detection
perspective diversity
factual checking
risk detection
confidence

It does not guarantee correctness and should be evaluated against the specific task.

AI and Enterprise Alignmentprotected

AI and Enterprise Integration

AI systems should integrate through governed:

APIs
events
messaging
workflow engines
data products
tool registries
identity services
approval services

Avoid embedding uncontrolled business logic solely in prompts.

AI and Enterprise Data Architecture

AI architecture should align with:

Data domains
data ownership
systems of record
metadata
data quality
lineage
classification
retention
data products
master data

AI and Application Portfolio Management

The application portfolio should identify:

AI-enabled applications
embedded AI
model dependencies
AI cost
AI risk
duplicate copilots
provider concentration
unsupported AI components
AI retirement candidates

AI and Security Architecture

Align with:

Zero Trust
identity governance
secrets management
data protection
logging
threat detection
vulnerability management
incident response
third-party risk

AI and Business Architecture

Map AI to:

Business capabilities
value streams
business processes
customer journeys
employee journeys
decisions
controls
outcomes

AI should not be mapped only to technical systems.

AI Portfolio and Business Caseprotected

AI Portfolio Prioritization

Evaluate use cases based on:

Business value
feasibility
data readiness
risk
user adoption
integration complexity
time to value
cost
scalability
strategic differentiation
regulatory impact
change readiness

AI Opportunity Categories

Productivity: drafting, summarization, search, meeting support, coding, document processing.

Decision Support: forecasting, anomaly detection, recommendations, risk analysis, scenario analysis.

Customer Experience: service assistance, personalization, self-service, translation, accessibility.

Operational Automation: case handling, workflow routing, document classification, monitoring, scheduling.

Product Innovation: AI-enabled products, adaptive interfaces, new data products, intelligent services.

Business Case

The business case should include:

Baseline
target outcome
volume
time saved
quality improvement
revenue
cost reduction
risk reduction
implementation cost
operating cost
human review cost
adoption
uncertainty
dependency
measurement method

Benefits Categories

Hard Savings: Actual budget removed.

Cost Avoidance: Future spend prevented.

Productivity: Time or throughput improvement.

Revenue: New or protected revenue.

Quality: Reduced error or improved consistency.

Risk Reduction: Reduced probability or impact.

Benefits should not be double-counted.

Metrics and Dashboardsprotected

AI Metrics

Track:

AI systems inventoried
Use cases by risk tier
Approved models
Approved providers
Active AI users
Adoption
Task success
Groundedness
Hallucination rate
Human override
Escalation
Safety events
Privacy events
Security incidents
Agent actions
Unauthorized actions
Cost
Cost per outcome
Latency
Availability
Model changes
Evaluation coverage
Monitoring coverage
Human oversight coverage
Benefits realized
Retired AI systems
Open exceptions

Executive AI Dashboard

Include:

AI portfolio size
Use cases by business capability
Use cases by risk tier
AI investment
AI operating cost
Benefits realized
High-risk systems
Systems without current evaluation
Systems without owners
Provider concentration
Agent autonomy
AI incidents
Open exceptions
Major regulatory exposure
Required executive decisions

Operational AI Dashboard

Include:

Model availability
Latency
Error rate
Token use
Cost
Budget variance
Retrieval quality
Hallucination indicators
Safety-filter events
Agent tool calls
Failed actions
Human approvals
Model drift
Evaluation failures
Prompt changes
Data-source changes
Incidents
Exceptions nearing expiration

Enterprise AI Architecture Maturity Modelprotected

Level 1 — Experimental

AI use is ad hoc.
Public tools are used without visibility.
Ownership is unclear.
Evaluation is informal.
Costs are fragmented.
Governance is reactive.

Level 2 — Developing

Initial AI policies exist.
Some approved platforms are available.
High-profile use cases receive review.
Basic inventory exists.
Evaluations vary by team.

Level 3 — Defined

AI architecture standards exist.
Use cases are inventoried.
Risk tiers are defined.
Approved models and providers are governed.
Data access, evaluation, and human oversight are required.
Production gates exist.

Level 4 — Managed

AI systems are continuously monitored.
Model and prompt changes are governed.
Cost and quality are measured.
Agents use controlled tool registries.
AI governance is integrated with enterprise architecture, security, data, procurement, and risk.
Benefits are tracked.

Level 5 — Adaptive

AI platforms dynamically route models based on quality, risk, latency, and cost.
Evidence collection is automated.
Evaluations continuously reflect production behavior.
Guardrails adapt to risk.
Provider concentration is actively managed.
AI architecture evolves through measurable business outcomes and operational learning.
Human accountability remains explicit.

Roles and Responsibilitiesprotected

Executive AI Sponsor — Accountable for AI strategy, enterprise priorities, funding, risk appetite, executive oversight.

Chief AI Officer or AI Leader — Responsible for AI operating model, capability development, portfolio, governance coordination, value realization.

Chief Enterprise Architect — Accountable for enterprise alignment, reference architecture, platforms, integration, lifecycle, standards.

AI Architect — Responsible for AI system architecture, models, prompts, retrieval, agents, evaluation, observability, resilience.

Business Owner — Accountable for business outcome, funding, users, human oversight, risk acceptance where authorized, retirement approval.

Product Owner — Responsible for product roadmap, requirements, adoption, feedback, user experience, outcome measurement.

Model Owner — Responsible for model fitness, evaluation, versioning, performance, limitations, change recommendations.

Data Owner — Accountable for data access, quality, purpose, retention, privacy, source authority.

Security — Responsible for threat modeling, access controls, testing, monitoring, incident response, exceptions.

Privacy and Legal — Responsible for privacy requirements, legal obligations, automated decision requirements, contracts, disclosure, rights.

Model Risk or Responsible AI — Responsible for risk assessment, fairness, explainability, oversight, validation, assurance.

Platform Owner — Responsible for AI gateway, model access, reliability, cost controls, observability, lifecycle.

Finance and FinOps — Responsible for cost baseline, budget, allocation, unit economics, savings validation.

Procurement and Vendor Management — Responsible for provider review, contract terms, vendor risk, renewal, exit, concentration.

Example Environmentprotected

Organization: A 30,000-person multinational financial and professional services organization with:

More than 1,100 applications
Microsoft Azure
AWS
Private cloud
Multiple data platforms
Central identity
Federated product teams
More than 60 known AI use cases
Public generative AI usage
Several vendor copilots
Custom RAG systems
Early AI agents
Strict privacy and financial-control obligations
Formal model risk management for traditional models
Limited governance for generative AI

Current State:

The AI inventory is incomplete.
Different business units contract directly with model providers.
AI use cases are classified inconsistently.
Some internal copilots retrieve documents without preserving source permissions.
Prompt versions are not centrally recorded.
Model evaluations focus on demonstration quality.
Agents have broad service-account permissions.
Cost is tracked by cloud subscription rather than use case.
AI incidents are handled through general security processes.
Vendor contracts vary in data-use protections.
Model updates may occur without formal change review.

Example Executive Findingsprotected

ARC-010-001 — Enterprise AI Inventory Is Incomplete. Severity: Critical. Confidence: High. Evidence: Procurement, identity, API, cloud, and SaaS records identify materially more AI services than the formal AI inventory. Business Impact: Leadership cannot reliably assess AI risk, cost, data exposure, provider concentration, or regulatory obligations. Recommendation: Establish a reconciled AI inventory covering custom systems, public tools, embedded SaaS AI, models, prompts, retrieval, agents, and providers.

ARC-010-002 — Retrieval Does Not Consistently Preserve Source Permissions. Severity: Critical. Confidence: High. Evidence: Several internal assistants index shared repositories into common vector collections without enforcing document-level user authorization at retrieval time. Risk: Users may receive information they are not authorized to access. Recommendation: Implement identity-aware retrieval, document-level access filtering, tenant isolation, access revocation, and retrieval-security testing.

ARC-010-003 — AI Agents Have Excessive Permissions. Severity: Critical. Confidence: High. Evidence: Early agents use broadly privileged service accounts and can modify records across multiple systems without transaction limits or human approval. Risk: Prompt injection, model error, or compromised workflow could cause material unauthorized action. Recommendation: Create unique agent identities, least-privilege tool scopes, action limits, transaction thresholds, approval gates, monitoring, and kill switches.

ARC-010-004 — Model Evaluation Is Not Business-Specific. Severity: High. Confidence: High. Evidence: Approval decisions rely primarily on vendor benchmarks and demonstrations rather than representative enterprise datasets and defined task thresholds. Risk: Systems may appear capable while failing on material business cases, edge cases, or sensitive scenarios. Recommendation: Implement use-case-specific evaluation suites, human review, adversarial testing, regression testing, and production monitoring.

ARC-010-005 — Prompt and Model Changes Are Not Governed. Severity: High. Confidence: High. Evidence: Production prompts and model versions can be changed without documented review, regression testing, or rollback criteria. Risk: System behavior may change materially without stakeholder awareness. Recommendation: Version prompts and models, define material-change criteria, require regression testing, and maintain rollback capability.

ARC-010-006 — AI Costs Are Not Attributed to Business Outcomes. Severity: High. Confidence: Medium. Evidence: Model, retrieval, vector, and evaluation costs are aggregated by cloud subscription, preventing use-case unit economics. Financial Impact: The organization cannot compare models, optimize routing, validate benefits, or assign accountability. Recommendation: Allocate AI cost by application, use case, business owner, model, provider, and successful outcome.

ARC-010-007 — Provider Concentration Is Material. Severity: High. Confidence: Medium. Evidence: Most generative AI workloads depend on one provider, use provider-specific prompts, and lack tested fallback models. Risk: Price changes, service disruption, policy changes, or model retirement could affect multiple critical systems. Recommendation: Implement model abstraction, portable prompt design, fallback evaluation, provider-risk monitoring, and tested continuity plans.

ARC-010-008 — Embedded SaaS AI Is Not Governed Consistently. Severity: High. Confidence: High. Evidence: Multiple SaaS platforms introduced AI features through product updates without a consistent review of data access, model providers, retention, or administrative controls. Recommendation: Integrate embedded AI review into SaaS procurement, renewal, change management, privacy, and architecture governance.

ARC-010-009 — Human Oversight Is Nominal. Severity: High. Confidence: Medium. Evidence: Some workflows require users to approve model recommendations but do not provide sources, uncertainty, alternatives, or adequate review time. Risk: Users may rubber-stamp outputs rather than exercise meaningful judgment. Recommendation: Redesign oversight to provide evidence, limitations, override capability, escalation, and reviewer training.

ARC-010-010 — AI Incident Response Is Incomplete. Severity: High. Confidence: High. Evidence: Existing incident plans do not explicitly address prompt injection, harmful output, unauthorized agent action, model degradation, retrieval leakage, or runaway cost. Recommendation: Create AI-specific incident playbooks integrated with security, privacy, legal, model risk, communications, and business operations.

Automation Opportunitiesprotected

AI discovery
Use-case registration
Risk-tier suggestions
Model inventory
Provider inventory
Prompt versioning
Agent inventory
Tool registry
Data-source mapping
Retrieval access testing
Model evaluation
Regression testing
Red-team testing
Prompt-injection testing
Output validation
Model routing
Provider failover
Cost allocation
Budget alerts
Agent transaction controls
Human approval routing
AI observability
Incident detection
Kill-switch execution
Lifecycle alerts
Model-change alerts
Exception expiration
Executive dashboards
Operational dashboards
Benefits tracking
A mature automated AI governance workflow could discover a use case, create an inventory record, suggest a risk tier, identify applicable standards, validate approved providers and models, check data-source authorization, generate an evaluation plan, run automated tests, route material findings for human review, validate agent permissions, establish budgets and limits, approve production through accountable governance, monitor quality/risk/cost, detect material changes, trigger reevaluation, support incident containment, track benefits, and retire the system when no longer justified.

Pro Tipsprotected

Start with business outcomes.
Inventory embedded SaaS AI.
Use risk-based governance.
Keep human accountability explicit.
Treat AI as a complete system.
Preserve source permissions in retrieval.
Use unique identities for agents.
Apply least privilege to tools.
Set transaction and spending limits.
Use the smallest model that reliably works.
Evaluate models on enterprise tasks.
Version prompts and models.
Treat model changes as production changes.
Monitor quality after deployment.
Include non-AI fallback.
Test provider failover.
Measure cost per successful outcome.
Protect AI telemetry.
Design kill switches early.
Include privacy and records management.
Validate human oversight.
Assess provider concentration.
Integrate AI governance with procurement.
Integrate AI architecture with data and integration architecture.
Retire failed pilots.
Do not confuse model confidence with truth.
Do not confuse model consensus with proof.
Make the governed path the easiest path.
Scale controls with impact, autonomy, and reversibility.
Keep final authority with accountable people.

Common Mistakesprotected

Treating AI as Only a Model — The complete system includes data, prompts, retrieval, agents, tools, users, workflows, monitoring, and human authority.
Starting With a Model Instead of a Business Problem — Model selection should follow use-case and outcome definition.
Assuming the Largest Model Is Best — Larger models may increase cost and latency without improving the business outcome.
Using Vendor Benchmarks as Production Evidence — Enterprise use cases require representative evaluation.
Treating a Demonstration as an Evaluation — A successful demo does not establish reliability, safety, or scalability.
Using Prompt Instructions as the Main Security Control — Prompts cannot replace identity, authorization, isolation, validation, and monitoring.
Indexing Data Without Preserving Permissions — Retrieval must enforce source access.
Giving Agents Broad Service Accounts — Agents require unique identities and least privilege.
Adding Human Approval Without Meaningful Review — Oversight must provide evidence, time, authority, and escalation.
Logging Everything Without Privacy Review — AI telemetry may contain highly sensitive information.
Failing to Version Prompts — Prompt changes can materially alter system behavior.
Ignoring Embedded SaaS AI — AI functionality may enter through ordinary product updates.
Treating All AI as High Risk — Excessive governance can block beneficial low-risk use cases.
Treating All AI as Low Risk — Risk depends on impact, data, autonomy, scale, and reversibility.
Allowing Pilots to Become Permanent Production — Pilots need expiration and production gates.
Ignoring Provider Model Changes — Managed models may change behavior even when endpoint names remain stable.
Relying on One Provider Without Fallback — Provider concentration can become enterprise operational risk.
Measuring Cost Per Token Only — Cost per successful business outcome is more meaningful.
Ignoring Human Review Cost — Human oversight may be a material part of total cost.
Assuming Multiple Models Guarantee Correctness — Consensus can repeat the same error or source bias.
Allowing Autonomous Action Without Limits — Permissions, transaction limits, approval, monitoring, and kill switches are required.
Forgetting Retirement — Retirement includes model access, prompts, retrieval, vector data, tools, agents, contracts, and records.

Security Considerationsprotected

AI architecture analysis may contain sensitive information such as model credentials, API keys, system prompts, proprietary prompts, agent tool definitions, privileged actions, data-source locations, security controls, vulnerabilities, customer data, employee data, regulated data, retrieval indexes, vendor contract terms, model evaluation failures, incident records, merger plans, and intellectual property.
Before using AI: remove passwords, API keys, access tokens, and private keys; redact secrets; mask personal information; avoid exposing unrestricted system prompts; restrict architecture diagrams; use approved enterprise AI tools; review retention and training settings; apply contractual confidentiality; limit access to generated reports; and preserve evidence securely.
AI-generated analysis should not independently approve AI use cases, assign legal risk, accept model risk, approve production, approve autonomous permissions, select a provider, approve sensitive data access, determine regulatory compliance, approve high-impact decisions, terminate a vendor, delete data, or make personnel decisions. These require accountable human review.

Brian Diamond

Founder, BrianOnAI

Twenty-five years designing, operating, and governing enterprise infrastructure — from MSP operations across dozens of client environments to enterprise infrastructure leadership. This blueprint codifies the operating model he's implemented in production, not theory.

Related Blueprints

Design an Enterprise Architecture Strategy and Operating Model Develop an Enterprise Technology Strategy Build an Enterprise Business Capability and Value Stream Model Assess and Rationalize an Enterprise Application Portfolio Design an Enterprise Integration Architecture Design an Enterprise Data Architecture Create Enterprise Architecture Principles and Standards Design an Enterprise Architecture Governance and Review Process Develop a Target-State Enterprise Architecture and Transformation Roadmap Assess Cloud Readiness and Select a Migration Strategy Prepare a Cloud Workload for Production Perform an Enterprise Security Assessment

⚠ Normalization Warnings — 12 for review

SCALE: This is an ARC-scale document with ~90 H1 sections. Domain body sections were grouped into thematic body/group parents (Scope & Definition, Principles, Operating Model & Governance, Reference Architecture, Model Strategy, Retrieval & Knowledge, Prompt & Output, Agent Architecture, Human Oversight, Evaluation, Security, Privacy, Responsible AI, Observability, Resilience, Incident, Cost, Vendor, Lifecycle, Standards/Patterns, Enterprise Alignment, Portfolio, Metrics/Dashboards) to avoid a flat 50+ section render. Confirm grouping boundaries.
CLASSIFICATION TO CONFIRM: 'Prerequisites' classified as a checklist TOOL (gather-before-start items are completable). Alternative: body/prose.
RESTRUCTURE: 'Primary AI Prompt' and 'Follow-Up Prompts' kept as TWO separate prompt_pack tools (primary-ai-prompt with 1 prompt; followup-prompts with 32 prompts) rather than one combined pack, because the follow-ups are 32 distinct standalone assessments. 'when' guidance lines are editorial additions; prompt text is verbatim from source. Confirm split.
MATRIX FROM PROSE: 'AI Use-Case Inventory', 'AI System Inventory', 'Agent Tool Registry', and 'AI Model Registry' were prose 'Capture:' lists describing record fields. Converted to matrix tools with a subset of fields as columns and the full field list preserved in rubric. skeleton_rows:0, example_rows empty (no doc rows). Confirm column selection — full field lists retained in rubric so nothing dropped.
CLASSIFICATION TO CONFIRM: 'Responsibility Matrix' is a real source table rendered as a matrix tool with all 15 activity rows verbatim in example_rows. rubric empty (source defines no scoring scheme).
CLASSIFICATION: 'Example Risk Tiers', 'Autonomy Levels', 'AI Lifecycle States', 'AI Lifecycle Gates', and the 'Enterprise AI Architecture Maturity Model' classified as body/reference (consulted taxonomies/tiered models), not tools. AI Risk-Tier Model factor list kept as body/prose introducing the tiers.
CLASSIFICATION: 'Example Environment' and 'Example Executive Findings' classified as body/example. Findings condensed from multi-paragraph Severity/Confidence/Evidence blocks into structured paragraphs — wording preserved, not paraphrased. Confirm condensation acceptable.
VALIDATION CHECKLIST: source 'Validation Checklist' has 11 H2 subgroups; preserved as checklist groups with headings verbatim. This is the acceptance gate for the whole architecture (phase 3).
PROMPT PICKED PHASE: primary-ai-prompt and followup-prompts assigned phase 2, though many follow-ups (readiness review, retirement, roadmap) span phase 3 activity. Phase assignment is best-fit.
ROADMAP: both the Twelve-Month Roadmap (4 quarters) and Three-Year Evolution Roadmap (3 years) were merged into playbook.roadmap as 7 horizon entries. Quick Wins mapped to playbook.quick_wins with 0–90 day horizon.
STATS: deliverables count (11) = number of tools. quick_wins (18) counted from source Quick Wins list. Confirm counts.
HOURS: source 'Estimated Time Saved: 60–240 Hours' mapped to hours_saved '60–240'.

SEO Block

Title tag: Design an Enterprise AI Architecture | ABME (43 chars)
Meta: Design an enterprise AI architecture that governs models, agents, retrieval, and cost — with risk tiers, human oversight, and a twelve-month roadmap. (149 chars)
Schema: HowTo · noindex: false
Related: arc-001, arc-002, arc-003, arc-004, arc-005, arc-006, arc-007, arc-008, arc-009, sec-001, sec-002, sec-005, sec-007, sec-008, bc-001, bc-002, cl-001, cl-010
Keywords: enterprise ai architecture, ai governance framework, ai risk tiers, ai gateway, model routing strategy, rag security, ai agent guardrails, ai evaluation framework, responsible ai, ai reference architecture, provider concentration risk, ai lifecycle management

Copied

Stop AI From Scaling Faster Than Your Controls — Design an Enterprise AI Architecture That Governs Every Model, Agent, and Retrieval Path

Executive Brief

Your Challenge

Common Obstacles

The ABME Approach

Insight Summary

The Journey

Establish Visibility and Control

Build the Enterprise AI Foundation

Evaluate, Operationalize, and Scale

What's Inside the Execution Layer

Prerequisites Checklist

Primary AI Architecture Prompt

Primary AI Prompt

AI Architecture Follow-Up Prompts

Discover the Enterprise AI Inventory

Create an AI Use-Case Inventory

Create an AI System Inventory

Classify AI Risk

Assess Whether AI Is Appropriate

Select an AI Model

Assess Open-Weight Model Use

Design an AI Gateway

Design a RAG Architecture

Assess Retrieval Security

Create a Prompt Governance Standard

Design an AI Agent

Review Agent Permissions

Design Human Oversight

Create an AI Evaluation Plan

Evaluate a RAG System

Red-Team an AI System

Create an AI Threat Model

Assess AI Privacy

Assess AI Fairness

Design AI Observability

Design AI Resilience

Create an AI Incident-Response Plan

Calculate AI Unit Economics

Control AI Costs

Assess AI Vendor Risk

Review Embedded SaaS AI

Create an AI Production Readiness Review

Assess a Material AI Change

Create an AI Retirement Plan

Build an Enterprise AI Roadmap

AI Use-Case Inventory

AI System Inventory

Agent Tool Registry

AI Model Registry

Prompt Governance Record

Prompt ID

Purpose

Owner

Version

Model

Risk Tier

Approved Use

Test Results

Change History

Effective Date

Review Date

Related Tools

Related Data Sources

AI Responsibility Matrix

Validation Checklist

Strategy and Governance

Inventory and Ownership

Architecture

Data and Retrieval

Models and Providers

Agents

Evaluation

Security and Privacy

Responsible AI

Operations and Cost

Lifecycle

Full Playbook

Overviewpublic

Business Problempublic