Skip to main content

Architecture

Why this function belongs in the AI Operating Model

The AI Operating Model is built around a portfolio of reusable AI products, not standalone solutions for every single case. Holding to this principle is impossible without architecture: it is architecture that ensures a new initiative relies on existing AI products and platform services wherever possible, instead of building a tenth integration with the same LLM in its own way.

Architecture connects three layers: the initiative's business scenario, the AI product portfolio, and the company's overall technology landscape. Without this touchpoint, the AI function quickly accumulates technical debt and loses control over costs (what Bain calls tool sprawl and opaque economics).

Where it engages

AI Operating Model stageRole of architecture
Assessment / product selectionProposes reusing an existing AI product or justifying a new one
DeliveryDesigns the target solution architecture, integrations, standards
Before productionConducts architecture review: standards compliance, scalability, maintainability
Portfolio developmentMaintains technology guidelines and reference patterns

What the function receives as input

  • The business scenario and non-functional requirements (load, availability, latency).
  • A map of integrations with corporate systems and data sources.
  • Solution options from the delivery track or the AI product owner.

What the function delivers as output

  • The target solution architecture and a justification of the choice (build / reuse / buy).
  • Technology constraints and standards (stack, integration contracts, patterns).
  • The architecture review verdict as a condition for passing the stage gate before production.

Key touchpoint artifacts

  • Architecture review — the solution architecture review document (available in the AI Operating Model artifact library).
  • AI landscape map — which AI products, models, and integrations already exist and where they are reused.
  • Technology guidelines — reference patterns for integrating and deploying AI products.

Anti-patterns

  • A custom solution for every initiative. Duplicated integrations and models, rising costs and tech debt instead of reusing products.
  • Architecture review as a box-ticking exercise. The review is purely formal and doesn't influence the decision — standards degrade.
  • Architecture in an ivory tower. Guidelines are written but disconnected from real delivery, so they get bypassed.