The AI Stack: A Layer Cake
The Five Layers
(1) Foundation models — the base LLMs/multimodal models. (2) Inference and serving — how models are hosted and called. (3) Orchestration — frameworks that chain calls, manage memory, route between models. (4) Tools and integrations — how AI reaches out to systems (databases, APIs, files). (5) Agents and agentic systems — autonomous goal-pursuing entities built on the stack below.
Why SEs Must Know This
Customers buy at different layers. Some want a model API. Some want a managed agent platform. Some want a finished agentic application. Misreading which layer they need is the #1 cause of misaligned discovery.
Whiteboard in 60 Seconds
Draw a vertical stack of five labeled boxes. Then ask the customer: "Where do you live today?" and mark the layer. Most enterprise security buyers are operating at layers 3-5 — they are not building foundation models, they are integrating them. The whiteboard moment immediately shifts the conversation from "which AI model is best" to "what is your orchestration, tool, and agent strategy?" — which is the right conversation.
Mapping Customer Scenarios to Layers
Scenario: "We want to use Gemini to answer analyst questions about our SIEM data." That is layers 3-4: orchestration (RAG + tool calls) plus integrations (SIEM API). Scenario: "We want agents that triage alerts autonomously." That is layer 5, with dependencies all the way down. Scenario: "We want to fine-tune a model on our threat data." That is layer 1. Getting clarity on which layer helps you scope the conversation, avoid over-engineering, and position the right products.
Prefer slides, quizzes, and saved progress? Read this lesson in the library — free, no sign-up.