Why Your AI Strategy Needs a Preliminary Phase Before the Pilot
Most enterprises approach AI with a familiar sequence.
Identify an AI use case.
Build a proof of concept.
Run a pilot.
Measure the result.
Then ask why the solution never reached production.
The problem often starts before the pilot.
The enterprise has not established the conditions required for AI to scale.
This is where Enterprise Architecture has an important role.
The TOGAF ADM includes a Preliminary Phase that establishes and prepares the enterprise architecture capability before the architecture development phases are undertaken. The purpose is to establish the architecture capability, principles, governance approach, organizational model, and supporting foundations needed for architecture work within the enterprise. The Open Group frames the Preliminary Phase around establishing the organizational context for architecture, including where, what, why, who, and how architecture work will be conducted within the enterprise.
AI strategy needs the same discipline.
Your first AI initiative should not begin with a pilot.
In this article, I use the term "AI Preliminary Phase" to describe an AI-specific extension of the principles and objectives of the TOGAF Preliminary Phase. It is not a named phase of the TOGAF ADM.
The term "AI Preliminary Phase" used in this article is a proposed AI-specific extension of the principles and objectives of the TOGAF Preliminary Phase. It is not a formal phase or named construct defined by The Open Group's TOGAF Standard.
The purpose is to apply enterprise architecture preparation, governance, principles, capability, organizational, and repository concepts to the specific challenges of enterprise AI.
The real question is not:
What AI pilot should we build?
The better question is:
Is our enterprise prepared to build, govern, operate, and scale AI?
The Pilot-First Problem
A pilot optimizes for learning about a specific use case.
An AI strategy must address the enterprise.
Those are different objectives.
A pilot asks:
- Can the model produce useful results?
- Does the user like the experience?
- Does the use case show business value?
An enterprise AI strategy asks:
- Where should AI create business value?
- Which AI capabilities should become strategic capabilities?
- Which data and knowledge assets support those capabilities?
- Which architecture principles should govern AI?
- Who owns AI decisions?
- How should AI risk be governed?
- Which AI workloads belong on shared platforms?
- Which capabilities should teams reuse?
- What operational model supports AI in production?
- What changes when AI moves from one use case to hundreds?
A pilot typically answers questions about a specific use case, user group, capability, or technical hypothesis. A scalable AI strategy answers an enterprise question.
The Missing Step Before AI Pilots
I propose introducing an AI Preliminary Phase before serious AI experimentation.
The concept extends the thinking behind the TOGAF Preliminary Phase into Enterprise AI.
The TOGAF Preliminary Phase establishes the architecture capability and organizational foundations needed before architecture development proceeds.
The Preliminary Phase establishes the architecture principles, governance approach, support strategy, organizational structures, architecture repository approach, and other foundations needed to conduct architecture work effectively.
AI introduces additional architecture and governance concerns that are particularly important when applying this enterprise architecture discipline to AI.
AI systems depend on data, knowledge, models, prompts, retrieval, agents, evaluation, security, identity, human oversight, runtime controls, observability, and economics.
Without enterprise-level decisions across these concerns, individual pilots tend to make their own architectural, technology, governance, and operational decisions.
The result is predictable.
Different models.
Different security patterns.
Different retrieval approaches.
Different evaluation methods.
Different governance processes.
Different data pipelines.
Different observability tools.
Different cost structures.
Different approaches to human oversight.
The enterprise can then accumulate multiple AI implementations without a coherent, intentional enterprise AI architecture.
The AI Preliminary Phase
Think of the AI Preliminary Phase as an enterprise readiness, architecture capability, and foundation-setting stage.
Its purpose is simple: Prepare the enterprise before committing to AI initiatives at scale. [Note: It does not imply that TOGAF requires an enterprise to complete a Preliminary Phase specifically before any AI initiative.]
I would structure it around eight decisions.
- AI Strategy
- AI Principles
- AI Governance
- AI Capability Readiness
- AI Architecture Foundation
- AI Data and Knowledge Foundation
- AI Operating Model
- AI Roadmap
TOGAF does not prescribe these eight AI decisions as a formal Preliminary Phase framework. They are an AI-specific structure that I propose, based on the objectives and concerns addressed by the TOGAF Preliminary Phase.
These decisions create the context for subsequent AI initiatives.
The pilot then becomes an architectural experiment within an enterprise direction.
Decision 1: Define the AI Strategic Intent
Start with business strategy.
Do not start with an LLM.
Ask:
- What business outcomes require AI?
- Which value streams contain meaningful intelligence opportunities?
- Which business capabilities should become AI-enabled?
- Which decisions require prediction, synthesis, reasoning, or autonomous execution?
- Which AI opportunities deserve enterprise investment?
- Which opportunities should remain local experiments?
This creates an AI opportunity portfolio.
A useful classification is:
- AI for operational efficiency
- AI for employee productivity
- AI for customer experience
- AI for decision intelligence
- AI for revenue generation
- AI for risk management
- AI for new products and services
The objective is not to place AI everywhere.
The objective is to establish where AI has strategic relevance.
Decision 2: Establish AI Architecture Principles
Every enterprise AI program needs architectural principles.
Without explicit principles, teams create local architecture decisions.
Your AI principles should answer questions such as:
- Will AI architecture remain cloud agnostic where strategic flexibility requires it?
- Will the enterprise use an approved AI gateway?
- Will applications access foundation models directly?
- Which AI workloads require human approval?
- How will sensitive data enter AI systems?
- How will identity propagate across agents and tools?
- How will AI outputs be evaluated?
- How will model changes be governed?
- How will prompts and policies be versioned?
- How will AI costs be measured?
- How will enterprises separate experimentation from production?
These principles become constraints for later architecture decisions.
TOGAF treats architecture principles as general rules and guidelines intended to inform and constrain architecture decisions across the enterprise. They provide a basis for making consistent architecture decisions over time.
The same logic applies to AI.
Decision 3: Establish AI Governance Before AI Scale
AI governance should not begin after production incidents.
Establish governance before the enterprise creates dozens of AI systems.
Define ownership for:
- AI strategy
- AI architecture
- AI risk
- AI data
- AI models
- AI security
- AI evaluation
- AI operations
- AI compliance
- AI vendor management
- AI economics
TOGAF's architecture governance provides a foundation for governing architecture-related decisions; enterprise AI governance extends beyond architecture to include AI risk, data, models, evaluation, compliance, operations, and economics.
Then define decision rights.
- Who approves a foundation model?
- Who approves an AI agent with transactional authority?
- Who determines acceptable hallucination rates?
- Who approves sensitive data access?
- Who owns an AI system after production deployment?
- Who decides whether an AI system requires human approval?
- Who approves model changes?
Governance needs explicit accountability.
A governance committee without decision rights creates meetings.
A governance model with explicit decision rights creates accountability and control mechanisms.
Decision 4: Assess Enterprise AI Readiness
Your organization needs more than AI enthusiasm.
It needs capability readiness.
As an AI-specific extension of the TOGAF Preliminary Phase, assess enterprise AI readiness across:
- Strategy
- Leadership
- Architecture
- Data
- Knowledge
- Security
- Identity
- Engineering
- AI skills
- Platform
- Governance
- Risk
- Operations
- FinOps
- Change management
For each dimension, determine:
- Current state
- Target state
- Capability gap
- Business impact
- Required investment
- Owner
- Target timeline
This can be represented as an AI capability heatmap.
The resulting capability gaps can then inform the roadmap.
This step helps expose a common failure pattern::The enterprise selects an AI use case first and discovers infrastructure, data, security, governance, and skills constraints during implementation.
Readiness assessment surfaces those questions before implementation commitments are made.
Decision 5: Define the AI Architecture Foundation
The enterprise needs an architectural baseline before multiple AI initiatives begin.
A practical AI architecture foundation includes:
- AI application architecture
- Model access architecture
- LLM gateway
- Prompt management
- Context engineering
- Knowledge architecture
- RAG architecture
- Agent architecture
- Tool integration
- Identity and access control
- AI security controls
- Evaluation architecture
- Observability
- AI governance controls
- Cost management
- Model lifecycle management
- Human oversight
This does not mean building every platform component before the first AI initiative.
The Preliminary Phase should establish the architecture capability, principles, governance, standards, and decision boundaries. Detailed AI solution and platform architectures can then be developed through the appropriate architecture development and implementation activities.
It means deciding which capabilities require enterprise standards and which capabilities belong inside individual solutions.
This distinction matters.
Enterprise capability does not mean enterprise implementation for every component.
Decision 6: Treat Data and Knowledge as AI Architecture
Many AI pilots fail because teams treat data as an implementation detail.
Enterprise AI requires a deliberate knowledge architecture.
An AI-specific architecture foundation should establish the principles, standards, capabilities, and decision boundaries for enterprise data and knowledge architecture. Detailed implementation choices should be addressed during subsequent architecture and solution design.
Assess:
- Source systems
- Data ownership
- Data quality
- Metadata
- Access controls
- Sensitive data
- Document structure
- Knowledge freshness
- Change data capture
- Batch ingestion
- Chunking
- Embeddings
- Vector storage
- Keyword retrieval
- Graph relationships
- Metadata filtering
- Knowledge evaluation
- Retrieval evaluation
- Knowledge lineage
An enterprise RAG architecture is not simply:
Documents → embeddings → vector database → LLM
The architecture needs to address how knowledge enters the system, how users gain access, how retrieval works, how knowledge changes, how results are evaluated, and how evidence reaches the model.
The principles, standards, capabilities, and decision boundaries governing these concerns belong in the architecture foundation; detailed implementation choices belong in subsequent architecture and solution design.
Decision 7: Define the AI Operating Model
Technology architecture does not operate itself.
As an AI-specific extension, define the operating model required to establish and sustain the architecture capability.
A mature AI operating model addresses:
- Centralized AI capabilities
- Federated domain ownership
- AI platform teams
- AI architecture
- AI engineering
- Data engineering
- AI governanc
- Security
- Model operations
- Evaluation
- Production support
- FinOps
- Vendor management
A useful question is: “What happens at 2 AM when an AI system produces unacceptable results?”
- Who detects it?
- Who investigates it?
- Who disables it?
- Who rolls back a model?
- Who changes a prompt?
- Who changes retrieval?
- Who communicates with the business?
- Who owns the incident?
AI production requires operational ownership.
Decision 8: Create the AI Roadmap
The AI roadmap should be developed from the strategic intent, capability gaps, architecture direction, governance requirements, and implementation priorities established through this preparation.
The roadmap should have multiple horizons.
Horizon 1: Foundation
Establish governance.
Define principles.
Assess readiness.
Establish architecture standards.
Create reusable platform capabilities.
Horizon 2: Strategic AI Products
Select high-value AI initiatives.
Build production-oriented solutions.
Measure business outcomes.
Validate architecture patterns.
Horizon 3: Enterprise Scale
Expand reusable capabilities.
Standardize operating practices.
Increase AI portfolio coverage.
Institutionalize evaluation and governance.
The roadmap should connect business outcomes to capabilities, architecture, initiatives, investment, and measurable results.
Where the Pilot Belongs
The argument is not “never run pilots.”
Pilots remain useful.
The issue is sequencing.
The pilot should sit inside the strategy and architecture.
A proposed AI architecture lifecycle, inspired by the TOGAF Preliminary Phase and adapted for enterprise AI:
- Business Strategy
- AI Strategic Intent
- AI Preliminary Phase
- AI Principles
- AI Governance
- AI Readiness Assessment
- AI Architecture Foundation
- AI Roadmap
- AI Initiative Selection
- AI Pilot
- Production Architecture
- Production Deployment
- Implementation Governance
- Continuous Improvement
This changes the role of the pilot.
The pilot should not be responsible for discovering the enterprise's entire AI strategy. It should validate defined business, technical, architectural, and operational hypotheses within that strategic direction.
It becomes a controlled validation of an already established strategic direction.
The Pilot Should Answer Five Questions
Once the AI Preliminary Phase exists, each pilot should answer specific questions.
Business Value: Does the solution produce the expected business outcome?
AI Capability: Does the selected AI capability solve the target problem?
Architecture: Does the architecture satisfy enterprise principles?
Operational Readiness: Does the solution meet production requirements?
Scalability: Does the pattern provide a credible path to multiple users, workloads, and business domains?
A pilot that validates only business value and AI capability leaves important architectural, operational, and scalability questions unresolved.
From PoC Architecture to Production Architecture
Consider an enterprise building an internal knowledge assistant.
The PoC might use:
- One document repository
- One embedding model
- One vector database
- One foundation model
- One prompt
- One application
- One developer
- One evaluation dataset
This setup may be sufficient for experimentation.
Production introduces different requirements.
- Identity
- Authorization
- Data isolation
- Knowledge freshness
- Retrieval quality
- Model governance
- Prompt versioning
- Evaluation
- Observability
- Auditability
- Cost controls
- Availability
- Incident management
- Human escalation
The architecture needs to evolve.
The Preliminary Phase prepares the organization to make those decisions deliberately.
The AI Preliminary Phase Creates Architecture Guardrails
Think about what happens when ten AI teams work independently.
Team A selects one model.
Team B selects another.
Team C creates its own RAG pipeline.
Team D builds direct model integrations.
Team E creates an agent framework.
Team F implements its own evaluation methodology.
Team G creates its own AI security controls.
Each decision might appear reasonable in isolation.
The enterprise outcome becomes harder to govern.
An AI Preliminary Phase establishes architectural guardrails before those decisions proliferate.
The objective is not to eliminate team autonomy.
The objective is to establish architectural boundaries within which teams operate.
The Architecture Repository Needs AI Assets
TOGAF includes the Architecture Repository as part of the enterprise architecture capability.
An enterprise applying TOGAF's architecture repository concept to AI should consider maintaining reusable AI architecture assets such as:
- AI architecture principles
- Reference architectures
- AI patterns
- RAG patterns
- Agent patterns
- Evaluation patterns
- Security patterns
- Knowledge architecture patterns
- Model selection criteria
- AI governance controls
- Approved technology standards
- AI threat models
- AI quality standards
- FinOps patterns
- Operating procedures
- Architecture decision records
- AI maturity assessments
These assets reduce repeated architecture decisions.
They also improve consistency across AI initiatives.
The AI Preliminary Phase Connects Strategy to Architecture
This is where Enterprise Architecture becomes useful for AI strategy.
A practical AI strategy-to-architecture chain can look like this:
- Business Strategy
- Business Capabilities
- Value Streams
- AI Strategic Intent
- AI Opportunities
- AI Capabilities
- AI Architecture Principles
- AI Governance
- AI Architecture
- AI Roadmap
- AI Products
- Business Outcomes
Each layer answers a different question.
Business Strategy asks where the enterprise wants to go.
AI Strategic Intent asks where AI contributes.
AI Opportunities identify where AI has business relevance.
AI Capabilities define what the enterprise needs.
AI Architecture defines how those capabilities are structured, realized, integrated, governed, and evolved as enterprise systems.
AI Roadmap defines how the enterprise moves from current state to target state.
AI Products create measurable outcomes.
A Practical AI Preliminary Phase Canvas
You could run the Preliminary Phase as a structured architecture workshop.
The following canvas is a proposed AI-specific workshop structure, not a TOGAF-defined Preliminary Phase deliverable.
Section 1: Strategic Intent
Business priorities
AI ambitions
Target outcomes
Investment themes
Strategic constraints
Section 2: Opportunity Portfolio
Business capability
Value stream
AI opportunity
Expected value
Complexity
Risk
Strategic relevance
Section 3: Principles
AI architecture principles
Data principles
Security principles
Model principles
Knowledge principles
Agent principles
Governance principles
Economic principles
Section 4: Readiness
People
Process
Technology
Data
Knowledge
Security
Governance
Operations
Section 5: Architecture
Baseline architecture
Target architecture
Reference architecture
Shared AI capabilities
Domain capabilities
Technology standards
Section 6: Governance
Decision rights
Risk classification
Approval gates
Evaluation standards
Human oversight
Production ownership
Section 7: Operating Model
Central capabilities
Domain capabilities
Platform teams
Architecture team
Governance team
Operations
FinOps
Section 8: Roadmap
Foundation
Strategic initiatives
Platform capabilities
Migration
Scale
Measurement
The Architecture Gate Before the Pilot
Before approving an AI pilot, ask seven questions.
-
What business capability does this initiative support?
-
What measurable business outcome does it target?
-
Which AI architecture principles apply?
-
Which enterprise capabilities does it require?
-
Which governance controls apply?
-
What is the production architecture path?
-
What reusable capability will the enterprise gain from this initiative?
If the team cannot answer these questions, the enterprise should first determine which architectural or governance decisions must be resolved before the pilot proceeds.
The right response is not necessarily to cancel the initiative.
The appropriate response may be to resolve the missing architecture or governance decisions before the pilot proceeds.
What the Preliminary Phase Changes
Without an AI Preliminary Phase:
- Use case first.
- Technology selection follows.
- Architecture emerges during implementation.
- Governance follows incidents.
- Production requirements appear late.
- Teams duplicate capabilities.
- Costs become difficult to control.
With an AI Preliminary Phase:
- Strategy comes first.
- Capabilities become explicit.
- Principles guide decisions.
- Governance starts early.
- Architecture establishes reusable patterns and decision boundaries.
- Pilots validate defined hypotheses.
- Production requirements enter early.
- The roadmap connects initiatives.
- Enterprise capabilities compound over time.
The Bigger Principle
AI transformation is an architecture problem as much as a technology problem.
The foundation determines how individual AI systems evolve.
A pilot validates a use case or defined hypotheses.
An architecture establishes a repeatable way to build many use cases.
A strategy establishes why the enterprise should build them.
Governance establishes who decides, who owns, and who controls them.
An operating model establishes how the capability survives beyond the project.
This is why AI strategy needs a Preliminary Phase.
Before asking:
What should we pilot?
Ask:
What must be true for AI to become an enterprise capability?
That question changes the architecture discussion.
It moves the enterprise from isolated AI experiments toward an intentional AI capability.
The AI Architecture Thinking Framework
Use this sequence when evaluating an enterprise AI strategy:
The following is a proposed AI architecture thinking framework, inspired by the objectives of the TOGAF Preliminary Phase and adapted for enterprise AI.
Why does AI matter to the business?
Where should AI create measurable value?
What AI capabilities does the enterprise need?
What architectural rules should govern those capabilities?
What organizational and technical gaps exist?
Who decides, approves, owns, monitors, and intervenes?
What target architecture supports the strategy?
Who builds, governs, operates, and evolves the capability?
What should the enterprise prioritize first?
Which hypothesis deserves controlled validation?
What needs to become reusable enterprise capability?
What measurable business result justifies continued investment?
This sequence puts the pilot in its proper place.
The pilot becomes one step inside an architecture-led AI strategy.
It no longer carries the responsibility of defining the strategy.
Conclusion
Enterprise AI does not fail only because a model produces poor results.
It also fails when the enterprise starts building before deciding how AI should work across the organization.
The TOGAF Preliminary Phase emphasizes establishing the architecture capability and organizational foundations needed to conduct architecture work effectively.
For AI, extend this thinking into an AI Preliminary Phase.
Define strategic intent.
Establish principles.
Assess readiness.
Create governance.
Define architecture foundations.
Establish the operating model.
Build the roadmap.
Then select the pilot.
Your first AI initiative should validate defined business, technical, architectural, and operational hypotheses.
It should not become the place where your enterprise discovers its AI strategy.
The strategic question is therefore not:
**What AI pilot should we run?*8
It is:
What enterprise conditions must we establish before AI pilots begin?
That is where an AI strategy starts becoming an enterprise capability.
✍️ About the Author
Sanjoy Kumar Malik — Principal AI Architect, Enterprise AI Strategist, and Senior Engineering & Technology Leader with 20+ years of corporate IT experience and a broader 27+ year professional journey, spanning Enterprise Architecture, software architecture, cloud-native systems, engineering leadership, and AI architecture. He is a TOGAF 10 Certified Enterprise Architecture Practitioner and AWS Certified Solutions Architect – Professional.
Sanjoy focuses on translating business strategy and AI opportunity into coherent enterprise architecture and scalable engineering execution. He works at the intersection of business, technology, architecture, and AI, helping organizations establish the architectural foundations, technology capabilities, and engineering systems required to turn AI initiatives into production-grade, scalable, governed, and economically sustainable enterprise capabilities.
He is the creator of The 28-Category AI Architecture Decision Framework (28-CAADF), a systematic approach to making AI architecture decisions in an era where intelligence itself is becoming an architectural capability.