Forward-deployed / Learning zone
Generative AIa Generative AI module
Lesson 04

The generative AI product stack

TL;DR

A generative AI feature is never just a model. The model is one layer inside a stack that runs from raw request to shipped answer, and most of a product team's engineering work happens in the layers above and below the model, not inside it. Four layers make up the stack: the model that generates content, the grounding and memory that give it real, current facts to work from, the action layer that lets it call tools or take steps in a loop, and the operations layer that measures, secures, and prices the whole system in production. A team that only budgets for "the model" — one API call, one line item — has priced roughly a quarter of the actual system. This lesson is the map. Each layer is a full module elsewhere in this curriculum, and this lesson tells you which one to open next.

🎯 For the product leader

Why it matters — "We're using GPT" or "we're using Claude" describes one layer of a four-layer system. Every layer above and below it — where facts come from, what the system can do, how you know it is working — is a separate set of decisions, costs, and risks.

What it changes in your decisions — You scope and budget by layer, not by model. A project plan that lists "integrate the model" as one line item is missing three-quarters of the actual build.

Ask yourself — "For this feature, what grounds the model in real facts, what lets it act, and how will we know when it's wrong — and have we budgeted each of those separately?"

Risk if ignored — A team ships "the model part" on time, then discovers grounding, action, and operations were never scoped, and the feature cannot go live without months of work nobody planned for.

The mental model: an engine is not a car

A model is an engine. An engine alone cannot take you anywhere — it needs fuel lines, a chassis, a steering system, and a dashboard before it becomes a car. The model in a generative AI product is the same. It needs a way to get real facts (grounding), a way to act on the world (the action layer), and a way to be watched and controlled once it is running (operations). Skip any of these, and you have an engine on a workbench, not a product.

The stack

Four layers · every real product wires all four

The model is Layer 1. Skipping any of the other three ships a demo, not a product.

Layer 1
The model

Generates text · image · audio · video · code

Layer 2
Grounding & memory

Real facts from your own data · what the system remembers

RAGMemory
Layer 3
Action

Trigger real effects · plan and act in steps

Tool callingAgent loop
Layer 4
Operations

Prove it's working · keep it safe · know what it costs

EvaluationSecurityCost

The four layers, and where to go deeper

Layer What it does The module that covers it
The model Turns a request into new content, in one of the five modalities This module names the shift; the mechanics of the model itself live in inference internals.
Grounding & memory Supplies real, current, private facts, and remembers what happened earlier in a session RAG & vector databases for grounding; context engineering for memory.
Action Lets the model trigger real steps — calling an API, running code, taking an action in a loop Function calling for single actions; Agentic AI for the full loop.
Operations Measures quality, blocks unsafe output, and tracks what the system costs to run Evals, safety engineering, and cost attribution.

Not every feature needs every layer at full strength. A simple text-summarization tool may need almost no grounding and no action layer at all — just a model and a thin operations layer to catch bad output. A customer-facing support agent needs all four layers, built with real care, because it must know real facts, take real actions, and be watched closely once it is live. Scoping a feature means deciding how much of each layer it actually needs, not assuming every feature needs the same four-layer weight.

Why teams under-price this stack

Vendor demos show the model layer, because that is the layer a vendor sells and controls. Grounding, action, and operations are almost always work your own team does, on your own data, inside your own systems — which makes them invisible in a demo and easy to leave out of an early estimate. The fix is structural: price a generative AI feature as four line items, not one, before you commit to a date.

Failure modes

Practitioner checklist