AI is a stack. Most organizations only see the top layer.
Kevin Hein · March 2026
Founder, CRYTCL Inc. · Tirias Research Senior Analyst
The common mental model
Most organizations think of AI as a model accessed through an interface. You prompt it, it responds, you evaluate the output. That mental model is useful for early exploration. It becomes a liability when you're building something that needs to work reliably inside a real organization.
The six layers
Applications
Top layerInternal copilots, operational workflows, and domain-specific interfaces. This is what users see and interact with — but it depends entirely on everything below it.
Agents
Layer 5Orchestrated reasoning systems with tools, controls, and bounded autonomy. Agents are only as reliable as the knowledge and data they have access to.
Knowledge Infrastructure
Layer 4RAG architecture, semantic search, knowledge graphs, and memory systems. This is the layer that connects AI to your organizational context — and the one most often skipped.
Data Platform
Layer 3Pipelines, governance, source connectivity, lineage, and enterprise data shape. Clean, connected, well-governed inputs determine the ceiling on everything above.
Model Layer
Layer 2Private models, secure integration, tuning strategy, routing, and evaluation. Model selection is a real variable, but it is constrained by what the layers below provide.
Compute & Security
Bottom layerDeployment posture, access controls, isolation, observability, and risk management. The foundation everything else rests on — designed first, not retrofitted.
Why the layers matter
Each layer has dependencies on the layer below it. An agent can only reason effectively if the knowledge infrastructure gives it accurate, scoped, traceable information. Knowledge infrastructure can only retrieve accurately if the data platform gives it clean, connected, well-governed inputs. Weakness at any layer propagates upward. A sophisticated model deployed on a weak knowledge layer will produce sophisticated-sounding answers that aren't grounded in reality.
Where most organizations get stuck
Most organizations enter at the Application layer — they deploy a copilot or an AI interface — without having designed the layers beneath it. The application works in demos. It fails in production. The instinct is to replace the application or the model. The actual problem is one or two layers down.
The architectural principle
Treat AI like any other enterprise system: design the stack, establish the operating model, and let application decisions follow from infrastructure decisions rather than the other way around. The organizations that do this produce AI programs that compound over time. The organizations that skip it run pilots indefinitely.