l'architecture d'entreprise, un levier essentiel pour réussir l'intégration de l’IA

une perspective alignée sur TOGAF©

L’IA a quitté le domaine de la marge expérimentale pour devenir une composante fondamentale de la stratégie d’entreprise moderne. D’un côté elle joue un rôle catalyseur déterminant pour l’automatisation de processus, la génération de contenu, l’aide à la décision et l’efficacité opérationnelle. De l’autre elle introduit des changements et des défis inédits : systèmes non déterministes, dépendance critique aux données, agents autonomes, cycles d’évolution rapides, contraintes réglementaires, modèle de coûts variables et enjeux éthiques.

Répondre à ces défis nécessite une approche holistique et maîtrisée. Sans cette approche les entreprises s’exposent à des risques systémiques qui se propagent à tous les niveaux : stratégie, opérations, conformité, finances et équipes. Pour les maîtriser, l’Architecture d’Entreprise (AE) constitue un levier particulièrement efficace. Voici quatre types de risques majeurs et la manière dont l’architecture d’entreprise permet d’y répondre.

Risques stratégiques

• Désalignement entre l'IA et les objectifs métier : des initiatives IA se déploient en silo, sans lien avec la stratégie globale de l'entreprise.

• Prolifération incontrôlée (Shadow AI) : les équipes adoptent des outils IA sans supervision, créant des angles morts pour la direction.

• Perte de cohérence du portefeuille applicatif : l'IA s'intègre de façon anarchique dans l'écosystème existant, générant des redondances et des incohérences.

L'AE aligne les projets d'IA sur la stratégie

• L'architecture d'entreprise part toujours des objectifs métier. Avant de déployer un modèle IA, elle pose la question : ce projet sert-il réellement la stratégie de l'organisation ?

• Elle permet ainsi d'éviter la prolifération de projets IA opportunistes, non coordonnés, et de s'assurer que chaque initiative s'inscrit dans une vision cohérente et partagée.

operational risks

Lack of standards and reusable patterns: every AI project starts from scratch, increasing costs and delays. Unmanaged vendor lock in due to lack of global visibility. Fragile integrations: AI models connected to critical systems without mapping data flows or assessing failure impacts.

EA structures and standardizes

By mapping existing systems, EA identifies where and how AI can be integrated without weakening the landscape. It defines reusable patterns—standardized ways to integrate, deploy, and connect AI models—preventing teams from reinventing the wheel or building incompatible solutions.

governance & compliance risks

Regulatory non compliance (EU AI Act, GDPR…): without a reference architecture, it becomes difficult to demonstrate traceability, explainability, or personal data management.

Lack of clear accountability: no one knows who owns a model in production, who updates it, or who monitors its drift.

Inability to audit decisions made by or with AI.

EA creates traceability and accountability

It defines who is responsible for what. For each AI component, EA clarifies: who authorized its deployment, what data it uses and where it comes from, which business processes it influences, and how it is monitored and updated

financial & human risks

Hidden costs and technical debt: unguided architectural choices create debt that becomes expensive to resolve later.

Duplicate investments: multiple teams fund similar solutions due to lack of shared visibility.

Loss of trust from stakeholders (employees, customers, regulators) in case of incidents or undetected bias.

EA optimizes investments

By providing a global view of the application portfolio and ongoing projects, EA prevents duplication and unnecessary spending.

It also helps anticipate technical debt before it becomes unmanageable.

applying best practices

The good news: organizations don’t need to invent a new architectural framework to govern AI. Methods like TOGAF©’s ADM (Architecture Development Method) provide a structured, iterative process to create, manage, and evolve enterprise architectures.

Regardless of the framework, we recommend the following approach:

Frame your vision and prepare the ground for adoption
Rethink your Business, Information System, and Technology architectures to harness AI
Build a concrete roadmap targeting the best opportunities
Govern uncertainty and adapt projects with agility

preliminary phase

Creating the right conditions for AI

Organisations must prepare their culture, governance, and architectural capabilities before introducing disruptive technologies like AI. This phase defines principles, governance, and readiness.

Main objective: Align AI initiatives with business strategy, ethics, and compliance.
Actions:
- Identify regulatory obligations (AI Act, nLPD, etc.)
- Define AI principles (“Ethics and transparency”, “Human in the loop”, etc.)
- Form an AI architecture team and governance committee
- Assess current AI maturity and capability gaps
Deliverables:
- AI Architecture Principles
- AI Governance Framework (agent autonomy, human responsibility)
- Skills and capability development plan

Example:

A healthcare provider sets up an AI governance committee to oversee projects and ensure all AI tools protect patient data.

architecture vision

Setting the direction for AI transformation

This phase answers the fundamental question: Why this change, and why now?
It establishes the strategic justification for AI, ensures it addresses real business needs—not technological hype—and secures executive buy in through dialogue with stakeholders and security teams.

Main objective: Create a shared vision of AI and define its scope.
Actions:
- Identify stakeholders and their challenges
- Define AI objectives (efficiency gains, cost reduction, etc.)
- Build business cases including ROI and risk assessment
Deliverables:
- AI Adoption Charter
- Vision and Scope Document
- Stakeholder Engagement Plan

Analogy:

Like conceptual sketches of a building, this phase ensures everyone agrees on the overall vision before construction begins.

business architecture

Rethinking processes to fully leverage AI (Human–AI collaboration)

Once the vision is set, architectural design can begin. The goal is to define how AI becomes a shared business capability rather than a collection of isolated solutions.

Main objective: Align AI use cases with business strategy.
Actions:
- Map current business processes
- Identify AI opportunities (fraud detection, predictive maintenance, etc.)
- Define future processes enhanced by AI
Deliverables:
- Prioritized list of AI use cases
- AI enhanced business processes
- Gap analysis report

Examples:

In healthcare, AI assisted triage and diagnostics integrate into patient admission workflows.
In retail, companies like Walmart and Carrefour use AI to dynamically adjust prices and promotions based on local demand, competitor actions, or even weather data.

information systems architecture

Structuring data and applications to support AI

Once business capabilities and processes are defined, they translate into:

Data the business must manage
Semantic context
Application interactions enabling agents to reason, coordinate, and act
Event flows and responsibilities
…

At this stage, architecture takes shape—not as a monolithic diagram, but as a living ecosystem of services, agents, and data flows.

Data Architecture

AI effectiveness depends on clean, unified, well governed data. Before choosing an LLM, data quality must come first.
Main objective: Model AI data flows and structures.

Actions:

Identify reliable data sources
Define training, inference, and feedback pipelines
Define ingestion strategies (chunking, embeddings, etc.)
Plan storage (vector DBs, graph DBs, relational DBs, etc.)

Application Architecture

This defines the AI “brain” (LLM), “skeleton” (agent framework), and “hands” (tools), all coordinated by its “nervous system”: orchestration.

Main objective: Define and structure AI software components.
Actions:
- Define components: LLM engines, agents, reasoning modules, APIs, AI gateway, input/output guardrails
- Describe interactions: vector DB queries, knowledge graph traversal, external API calls (CRM, ERP, KB…)
- Choose an application architecture (agentic microservices, event driven, streaming, etc.)

Example:
A secure ingestion pipeline (FHIR/HL7 compliant) continuously extracts medical records into a centralized datalake. Downstream, an AI inference engine analyzes the data and distributes predictions via an API Gateway to front end applications, enriching clinicians’ dashboards with real time diagnostics and recommendations.

technology architecture

Designing the infrastructure and foundation

AI requires significant compute power and specialized platforms. This phase defines the hardware, cloud services, and networks that will bring AI applications to life.

Main objective: Define the target infrastructure supporting AI.
Actions:
- Choose database types (vector, graph, relational, multimodal, time series)
- Define platforms: laptops/workstations for prototyping, cloud for production
- Size compute resources: CPUs for small models, GPUs/TPUs for larger LLMs
- Choose between public cloud, sovereign cloud, hybrid, or on premise
- Select observability tools to monitor performance and detect drift
- Define the MLOps/LLMOps foundation
- Ensure regulatory compliance by design (AI Act, GDPR)

Outcome:
A robust, scalable, secure technology foundation integrated with existing systems.

opportunities & solutions

Selecting the best deployment paths for AI

This is where architecture becomes reality. The phase moves from the “what” (ideal vision) to the “how” (practical implementation). It involves major decisions: Build vs. Buy, ROI based prioritization, and grouping ideas into actionable workstreams.

Main objective: Identify projects, vendors, and solutions to realize the vision.
Actions:
- Build vs. Buy analysis
- Structure delivery into work packages and prioritize high impact areas
- Prioritize by ROI: start with high value, low risk initiatives
- Use iterative design: validate feasibility with lightweight prototypes before scaling
- Rationalize technology: identify reusable AI components and select interoperable partners

Example:
Recruitment & onboarding with AI

Buy: For CV screening, the company purchases the AI module of its existing HR system (Workday/SAP).
Build: For onboarding, it develops a secure internal AI chatbot trained exclusively on confidential internal documents.

migration planning

Building the AI roadmap

After identifying opportunities and defining what to do, this phase determines when and at what cost the transformation will be executed.

Main objective: Finalize the transition plan and secure funding.
Actions:
- Maximize ROI through an incremental, controlled approach
- Define transition steps (AI assists humans before becoming autonomous)
- Estimate resources: budget, timelines, teams
- Publish the roadmap: show when each department will receive its AI tools

Example:
A city launches a unified citizen portal with an AI assistant.
Phase 1: AI handles high volume, low risk requests (trash schedules, birth certificate requests).
Phase 2: AI acts as a copilot for complex cases (building permits), pre filling forms for human validation.
Phase 3: End to end automation for 80% of procedures, with human oversight for exceptions.

implementation governance

Ensuring AI projects stay aligned with the target architecture

Main objective: Ensure deployed AI solutions comply with architectural and governance requirements.
Actions:
- Establish review checkpoints
- Monitor compliance with governance and regulatory standards
- Approve or adjust deployments as needed

architecture change management

Adapting to evolving AI needs

AI evolves rapidly, technologies change, regulations shift, and models degrade over time. This phase ensures continuous adaptation.

Main objective: Continuously update AI architecture to maintain value and reliability.
Actions:
- Monitor model drift
- Conduct technology watch to identify faster/cheaper AI options
- Manage updates without disrupting operations
- Update architecture based on new risks, opportunities, or regulations

Example:
After months in production, monitoring tools reveal that customer service AI models struggle with new incident types. The architecture committee approves incremental retraining and infrastructure upgrades, restoring performance without rebuilding the entire architecture.

requirements management

Navigating uncertainty and change

This is a continuous process feeding into all other phases. Between investment decisions and full deployment, business vision, data regulations, and market conditions may evolve.

Main objective: Continuously track and validate business and AI requirements.
Actions:
- Elicit needs from end users
- Integrate new regulatory requirements
- Validate alignment between technical teams and business needs

Successful AI integration is not a technological choice alone, it is the outcome of a structured enterprise architecture approach. Our role is to support you using recognized standards adapted to your context. Only a structured approach can turn ambition into a clear, coherent AI architecture built on key capabilities: reliable data, security, compliance, ethics, and a prioritized roadmap.
A method like ADM enables organizations to move fast without losing control, steering their AI transformation end to end.
Without architecture, AI creates chaos. With a structured approach, it creates value.

Tags:

plus d'insights

tous les insights

rejoignez itecor au congrès du management de projet 2026

company newsgovernance & service management

March 17, 2026

[IT Governance Breakfast] L’intégration de l’IA dans le SI et dans la culture d’entreprise