Should an enterprise use RAG, agents, fine-tuning, or automation?
The choice depends on knowledge volatility, required determinism, integration depth, data sensitivity, latency, cost, and the amount of human oversight needed.
Capability briefing
AI Architecture
Answer-engine summary
AI architecture connects business goals with data, retrieval, models, orchestration, evaluation, security, governance, and operations so AI systems can run safely in production.
Definition
AI architecture is the design of data, models, retrieval, orchestration, evaluation, security, and operational controls around AI capabilities.
Why it matters
The right architecture keeps prototypes from becoming fragile production liabilities.
Where this matters in enterprise decisions
AI architecture matters when an organization must choose between RAG, agents, fine-tuning, workflow automation, vendor platforms, or internal services while protecting data, cost, quality, and operational resilience.
Q&A for leaders
Common business questions
These answers are visible on the page and mirrored in structured data so search engines and answer engines can parse the same information human readers see.
Where should sensitive data live?
Sensitive data should remain inside governed enterprise boundaries with explicit access control, logging, retention rules, and vendor-risk review before model interaction.
How should AI quality be measured?
Measure task success, factuality, retrieval quality, safety, latency, cost, user feedback, and exception rates using repeatable evaluation sets and production telemetry.
How modular should the architecture be?
Core interfaces should isolate models, retrieval stores, prompts, tools, policies, and observability so the enterprise can change vendors or patterns without rewriting the whole system.
Common failure modes
- A prototype directly becomes production without evaluation, access control, or operational ownership.
- Model choice is treated as the architecture while data, workflow, and governance are ignored.
- Agentic systems receive broad permissions without bounded tools, logs, and escalation paths.
- Cloud and model costs are discovered after adoption instead of being designed into the architecture.
Architecture and governance implications
- AI architecture should be reviewed by architecture, security, data governance, risk, and delivery stakeholders.
- Production systems need evidence for data sources, model behavior, prompts, access, and incidents.
- Architecture standards should define reusable patterns rather than forcing every team to invent its own approach.
Related capabilities
Connected expertise areas
Related canonical writing
AI SDLC··9 min read
AI Changes Software Engineering, Part 3: The New Operating Model
Part 3 of the AI software engineering series: why AI adoption becomes an operating-model problem, what the emerging AI engineering stack looks like, and what leaders should ask Monday morning.
AI SDLC··10 min read
AI Changes Software Engineering, Part 2: AI Exposes Enterprise Fault Lines
Part 2 of the AI software engineering series: AI exposes existing enterprise weaknesses, changes requirements, raises the cost of architecture mistakes, turns testing into evaluation, and moves governance into runtime.
AI SDLC··8 min read
AI Changes Software Engineering, Part 1: The Bottleneck Has Moved
Part 1 of a three-part field guide on how AI changes software engineering: code generation accelerates, but the bottleneck moves toward architecture, evaluation, governance, ownership, and judgment.
AI Governance··10 min read
Enterprise AI Adoption: Why Most Programs Stall Between Pilot and Production
Why enterprise AI adoption often stalls between promising pilots and durable production: the bottleneck is usually data, governance, ownership, workflow integration, and operating model maturity rather than model quality alone.
Regulator-Defensible Architecture··12 min read
Three Pains Every Tier-1 CDO Is Failing Right Now
Why Tier-1 enterprises are simultaneously struggling with enterprise AI, regulator-defensible data architecture, and continuous modernization, and why current architectures solve at most one of the three.