There is a clarity problem at the centre of most enterprise AI strategy discussions. The vocabulary is inconsistent. Vendors use "context layer," "control plane," "agent framework," and "AI platform" to mean different things in different conversations. CIOs and CTOs are making architecture decisions without a shared map of what the enterprise AI stack actually contains - which layers do what, where the boundaries are, and critically, which layers belong to the enterprise versus which belong to vendors.

The lack of a clear stack taxonomy is not just conceptually frustrating. It produces real architectural errors: governance responsibilities that fall between layers nobody owns, vendor lock-in that was not anticipated because the layer boundaries were not understood at the time of the decision, and AI deployments that perform well on one dimension while being completely ungoverned on another.

This piece is a reference map. It defines each layer of the enterprise AI OS stack, describes what it does, identifies who typically owns it, and names the strategic decision each layer requires from CIO, CTO, and CDO leadership.

The complete stack - six layers

The enterprise AI OS stack has six distinct layers. They are not always cleanly separated in practice - vendors bundle them, enterprises conflate them - but the functional distinctions are real and the governance implications of each are independent.

Layer 6 - Application layer

AI-powered products and workflows

• Layer 5 - Agent framework layer

  • Orchestration, memory, tool use, multi-agent coordination

• Layer 4 - Control plane layer

  • Identity, permissions, policy enforcement, audit trail, containment

• Layer 3 - Context layer

  • Knowledge graph, semantic definitions, real-time operational data, provenance

• Layer 2 - Data infrastructure layer

  • Data warehouses, lakehouses, operational databases, API connectivity

• Layer 1 - Model layer

  • Foundation models, fine-tuned models, inference infrastructure

Layer 1: The model layer

The model layer is where foundation models live - the GPTs, Claudes, Geminis, and Llamas that provide the reasoning capability underlying enterprise AI applications. It includes inference infrastructure, fine-tuned model variants, and the compute that runs them.

What it does: Provides reasoning capability. Takes context and instructions as input, produces outputs.

Who owns it: Almost nobody, in any meaningful sense. Foundation models are vendor-supplied. The enterprise rents access to them through APIs. Fine-tuned variants add some differentiation, but the underlying capability is external.

The strategic truth: This is the commodity layer. Every competitor in your industry has access to the same foundation models. Your competitive advantage does not live here. Your build-vs-buy decision is essentially made for you - you buy this layer. The strategic question is which vendor's model you use for which use case, and how you avoid being locked to a single provider at a layer that will change significantly over the next twenty-four months.

The CTO decision: Model-agnostic architecture above this layer. The context and control layers above should be designed to swap models without rebuilding everything that runs on top of them.

Layer 2: The data infrastructure layer

The data infrastructure layer is the existing enterprise data stack - data warehouses, lakehouses, operational databases, CRMs, ERPs, HRMS platforms, and the API connectivity that links them. It is the layer that holds the raw data the context layer draws from and the operational systems the agent frameworks interact with.

What it does: Stores and serves enterprise data. Provides the source systems that the context layer reads from and the operational systems that AI agents act on.

Who owns it: The data engineering and IT infrastructure teams. This layer exists independently of enterprise AI - it was built to serve analytics, operations, and business applications. AI is a new consumer of it, not a reason to rebuild it.

The strategic truth: Most enterprises have this layer. The AI-relevant investment is in API connectivity - specifically, the ability to expose operational data from HRMS, CRM, and ERP systems to the context layer in real time, with appropriate access controls, rather than through batch exports that arrive stale.

The CDO decision: Which data sources need real-time API exposure versus batch sufficiency for AI use cases. For operational AI agents making consequential decisions - credit, insurance, compliance - the answer is real-time for the data that drives those decisions.

Layer 3: The context layer

The context layer is the intelligence layer - it transforms raw enterprise data into governed, structured, semantically meaningful context that AI agents can reliably reason from. It is the most actively invested layer in enterprise AI architecture in 2026, and the one where the definition of "complete" is still forming.

What it does: Provides knowledge graph representation of business entities and relationships, semantic definitions that disambiguate enterprise terminology, real-time operational data feeds that reflect current system state, metadata and lineage tracking for context quality assurance, and certified provenance that makes agent outputs auditable.

Who owns it: This is the layer the enterprise should most aggressively own. The knowledge graph encodes how your enterprise understands itself - your entities, your definitions, your relationships. 

This is proprietary. Nobody else has it. The enterprise that builds a complete, well-maintained context layer has built something that cannot be replicated by a competitor regardless of budget, because it is built from the enterprise's own accumulated knowledge and data.

The strategic truth: The context layer is the moat. It is the layer that makes your AI fundamentally different from a competitor's AI running the same model. Invest here disproportionately relative to the model layer, which is a commodity.

The CDO decision: Knowledge graph ownership and governance. Who maintains the semantic definitions? Who certifies new context sources? Who resolves conflicts when two data sources disagree? These are data governance questions that the CDO owns, not IT operations questions.

Layer 4: The control plane layer

The control plane is the governance layer - it governs what AI agents are allowed to do with the context they have received. It enforces access policies, manages agent identity and permissions, generates the tamper-evident audit trail, and provides containment when agent behaviour deviates from defined parameters.

What it does: Enforces least-privilege access at the action layer, maintains agent identity registry with revocable credentials, generates immutable audit logs of every agent action, provides real-time anomaly detection for agent behaviour, and enables instant containment when needed.

Who owns it: Shared ownership between the CISO and CIO - the CISO owns the security and access control dimensions, the CIO owns the governance and compliance dimensions. In organisations with a Chief AI Officer, this layer typically falls within their remit with CISO oversight.

The strategic truth: This is the layer that makes enterprise AI safe to scale. Without it, agent sprawl - 150,000 agents by 2028 per Gartner - becomes ungovernable. The control plane is the Kubernetes moment for enterprise AI: the layer that brings order to the chaos of distributed autonomous systems. It does not need to be built before the first agent is deployed. It needs to be built before the agent portfolio is large enough that retrofitting it becomes more expensive than starting over.

The CIO/CISO decision: Build the control plane as shared infrastructure - a governance layer that every agent in the enterprise runs through - rather than embedding governance controls individually in each agent. The shared approach is cheaper, more consistent, and significantly more auditable.

Layer 5: The agent framework layer

The agent framework layer provides the orchestration infrastructure that AI agents use to complete multi-step tasks - memory management, tool use, planning, multi-agent coordination, and workflow integration. This is the layer where LangChain, LlamaIndex, CrewAI, AutoGen, and similar frameworks operate.

What it does: Manages agent state and memory across a task, coordinates tool calls and API interactions, orchestrates multi-agent workflows, and handles the planning and replanning that autonomous task execution requires.

Who owns it: AI and MLOps teams, with strong involvement from the software engineering function. This is a technical infrastructure layer - it requires engineering investment to configure, maintain, and evolve as agent use cases grow.

The strategic truth: This layer is actively commoditising. LangChain, CrewAI, and AutoGen are converging on similar capability sets. The differentiating investment is not in which framework you use but in how well your context and control layers are integrated with the framework you choose. 

An agent framework running on a weak context layer and no control plane produces capable-sounding but ungoverned agents. The same framework running on a complete context and control layer produces reliable, auditable, production-grade AI.

The CTO decision: Framework selection, integration standards for context and control layer access, and the protocols - MCP, A2A - through which agents communicate with the context and data layers beneath them.

Layer 6: The application layer

The application layer is where AI-powered products and workflows live - the underwriting agent, the claims processing agent, the customer service agent, the compliance monitoring agent. This is the layer that business units see, use, and measure.

What it does: Delivers AI-powered business outcomes - faster decisions, automated workflows, intelligent assistance, and autonomous task execution across business functions.

Who owns it: Business units and product teams, with engineering support. The application layer is closest to the business problem and furthest from the infrastructure.

The strategic truth: This is where the AI investment is most visible - and where the temptation to shortcut the lower layers is strongest. Business units want applications. They do not want to hear that the context layer needs to be built before the application will perform reliably. 

The CIO and CTO's job is to ensure that application layer pressure does not result in deploying Layer 6 on an incomplete Layer 3 and a missing Layer 4 - which produces the failure pattern that is most common in enterprise AI in 2026: capable applications on inadequate infrastructure.

The ownership question - and why it matters strategically

The stack map is useful for investment decisions. It is most useful for ownership decisions - specifically, identifying which layers the enterprise should own deeply versus which layers it can safely rent.

The two layers an enterprise should own most deeply are the context layer and the control plane governance policies. 

These are the layers that encode how the enterprise understands its own business and what it will and will not allow its AI to do. They are also the layers that compound most significantly over time - a richer knowledge graph and a more refined access policy framework make every agent running on top of them better, without requiring changes to the agents themselves.

The layer that most enterprises are missing

The research from 2026 is consistent on which layer is most commonly absent. It is not the model layer - everybody has a model. It is not the application layer - business units have built or bought applications. It is not even the agent framework layer - LangChain and similar tools are widely deployed.

The missing layer is the real-time operational data component of the context layer - and the control plane.

Most enterprises have static context: knowledge bases, document repositories, indexed policy documents. They do not have the live connection to operational systems - HRMS, CRM, ERP - that provides agents with current facts rather than historical snapshots. And they do not have the control plane that governs what agents are allowed to do with the context they receive.

The result is a six-layer stack with strong Layer 1, Layer 2, Layer 5, and Layer 6 investment - and significant gaps at Layer 3 and Layer 4. The agents are capable. They are not reliable. They are not governed. And they are not producing the outcomes the investment in Layers 1, 5, and 6 promised, because the infrastructure those applications run on is incomplete.

Tartan's Context and Control OS sits at Layers 3 and 4 - providing the real-time operational data feed that completes the context layer and the control plane infrastructure that governs what agents are permitted to do with it. Not a replacement for the knowledge graph vendors or the agent framework vendors. The missing piece that makes the rest of the stack work.

The stack is the map. The gaps are the strategy. And in 2026, the gaps are consistent enough across enterprise AI deployments that naming them is not a diagnosis - it is a description of the default state. The default state is not good enough for production-grade agentic AI at scale. Closing it is the infrastructure decision of the year.