Architecting AI in the Real Enterprise

A book for enterprise architects who need to understand AI well enough to make real decisions about it — not to build models from scratch, but to integrate them into the living, breathing, politically complex ecosystem of an enterprise.

Why most AI initiatives fail — and the three-pillar framework that changes the outcome.

AI does not add a new box to your architecture diagrams. It changes what the boxes do — introducing non-deterministic components, inverting the relationship between code and data, and expanding the build-vs-buy decision into a four-way tradeoff. Here is what that means for how you design.

Not all AI is the same, and treating it as if it were is how architecture reviews go sideways. This chapter maps the AI stack onto the layered architecture thinking you already know — so you can evaluate any AI technology that crosses your desk with clear eyes.

Six new components belong in your architecture toolkit: LLMs, embeddings, vector databases, prompt templates, RAG, and agents. Each has specific latency, cost, and determinism properties. Here is how to reason about them the same way you reason about the components you already know.

AI amplifies your data quality in both directions — good data becomes a force multiplier, bad data becomes an active liability. This chapter covers the three pipeline types AI requires, how to extend your existing data governance framework, and why the data work consistently takes longer than the model work.

Seven patterns cover the vast majority of enterprise AI integration needs. This chapter walks through each one — from the simplest microservice wrapper to the AI Gateway that governs them all — so you can fit AI components into the architecture you already have without rebuilding it from scratch.

Deploying AI models by hand, with no equivalent discipline to CI/CD, is how organizations accumulate expensive surprises. MLOps and AI governance are the practices that close that gap — from automated evaluation pipelines to the AI Register that tells you what is actually running in your enterprise.

Responsible AI is not a compliance checkbox. It is a set of architectural decisions — about fairness testing, explainability layers, PII routing, safety guardrails, and accountability structures — that you as the architect own. This chapter makes each pillar concrete.

Google Cloud, AWS, and Azure have each made distinct bets on how AI should be delivered. Those bets create genuine trade-offs. This chapter gives you an honest side-by-side comparison and a decision framework you can bring into your next architecture review.

Nobody gets to build an AI-native enterprise from a blank slate. This chapter covers the three migration approaches — sidecar, augmentation, and rebuild — and a concrete playbook for sequencing them through the messy, imperfect systems that actually run your business today.

Nearly every GenAI system in production today is a variation on one of ten core patterns. This chapter is your field guide — walking through each one from simplest to most sophisticated, with clear guidance on when to reach for one over another.

Agents are the new integration layer — connecting systems, data, and human intent through dynamic reasoning rather than hardcoded flowcharts. This chapter covers how to design, govern, and operate agent systems at enterprise scale, including the five orchestration patterns every architect needs to know.

A fintech company's GenAI feature went viral. The next month's AI infrastructure bill was $900,000 against $200,000 in revenue. This chapter covers how AI costs actually work, where the hidden expenses lurk, and the engineering strategies that can reduce your AI bill by 50 to 90 percent without sacrificing the user experience.

The architecture is the easy part. The hard part is the eighty percent — the teams that don't trust the system, the executives who expected magic, the managers who see AI as a headcount threat. This chapter covers what most architecture books skip entirely.

Four enterprise AI transformations — a global bank drowning in regulatory documents, a healthcare network burning out its physicians on paperwork, a retailer bleeding $280M a year in forecasting errors, and an insurer promising five-day claims but delivering frustration. What the architecture actually looked like, what worked, and what everyone got wrong.

Fourteen chapters about transforming your enterprise. This last one is about transforming yourself. A concrete 90-day plan, an honest assessment of the skills gap, and the five pitfalls that claim most experienced architects who make this transition.

Ten reference architectures for common enterprise AI use cases — each with the key components, data flows, integration points, and design decisions that actually matter. Starting points, not blueprints. Adapt them.