For two years, the enterprise AI conversation was a conversation about models. Which foundation model performs best. Whose context window stretches furthest. Whose pricing makes the unit economics work. Every major technology vendor competed on the model dimension. Enterprises evaluated models, chose models, and built their AI strategies around model selection.
By mid-2026, that conversation has been replaced by a different one - one that is happening loudly at the vendor level and quietly, often without full awareness, at the enterprise level.
The new contest is not about models. It is about the control plane. Specifically: who controls the connective context layer - the agentic control plane - that decides what enterprise AI agents know, what they are allowed to do, and who is accountable when something goes wrong. Microsoft,
Google, Databricks, Snowflake, Salesforce, Atlan, and a growing list of infrastructure vendors are all making explicit moves on this layer. CRN's Hot Agentic AI 2026 list named the agent control plane as the new enterprise battleground in June 2026. The battle is live.
Most enterprises are not consciously participating in it. They are making individual product decisions - adopting Copilot here, building on Databricks there, deploying agents through a SaaS vendor's built-in framework - without recognising that each decision is a position statement on who governs their enterprise AI OS. And taken together, those decisions may be ceding that governance to vendors in ways that will prove structurally difficult to reverse.
What each major vendor is actually competing for
Understanding the battle requires understanding what each major player is trying to own - because the strategies are not identical, and the implications for enterprise independence differ significantly.
Microsoft is competing through integration depth. Copilot is woven into Windows, Edge, Office 365, and Azure AI Studio. The enterprise that runs Microsoft's stack runs agents that operate within
Microsoft's context surface - Microsoft IQ, defined at Build 2026 as the context layer that grounds agents in work patterns, business data, and enterprise knowledge. The strategic play is gravity: the more enterprise workflows operate within Microsoft's surface, the more the context layer is defined by Microsoft's semantic and governance choices. Exiting this stack is not impossible - but it is significantly harder than it appears at the point of adoption.
Google is competing through the data platform. The real story from Google Cloud Next 2026 was not the model updates. It was Gemini reframed as an orchestration layer, an agent runtime, a governance system, and a connective fabric. BigQuery and the data fabric are being positioned as the reasoning surface, not just the storage surface. The implication: the enterprise that runs its data on Google's lakehouse is increasingly running its AI context on Google's semantic layer. The sticky layer is now the semantics and the orchestration, not the data.
Databricks is competing through the lakehouse and the ontology. Genie One and Genie Ontology - announced at Data + AI Summit 2026 in front of 30,000 attendees - are positioned as a live context layer that grounds agents in governed operational data. Ali Ghodsi's keynote framing was unambiguous: AI has a context problem, not an intelligence problem. Databricks is positioning Unity
Catalog and Genie as the answer to that context problem - meaning the enterprise that runs on Databricks is increasingly running its AI context within Databricks' governance model.
Salesforce and ServiceNow are competing from the application layer down. Both are embedding agent control into their platforms - Agentforce for Salesforce, Now Assist for ServiceNow - so that enterprise agents operating in their domains operate under their governance frameworks. The control plane is built into the SaaS product, not separate from it.
"Exiting Google-managed semantics, Gemini agents, or BigQuery abstractions may prove harder than migrating the data itself. The semantics and the orchestration are now the sticky layer." - CIO.com, June 2026
Why this matters more than vendor selection usually does
Enterprise technology vendor decisions have always involved lock-in considerations. The difference with the AI control plane is the nature of what gets locked in.
When an enterprise commits to a cloud provider, it locks in infrastructure. Migration is expensive and disruptive - but infrastructure is infrastructure. The workloads can be described, the costs can be calculated, and the migration can be planned if the strategic calculus changes.
When an enterprise commits to a specific vendor's context and control layer, it locks in something more fundamental: how its AI understands the business. The knowledge graph, the semantic definitions, the ontological structure that tells AI agents what entities mean in this enterprise's context - these are not infrastructure. They are institutional knowledge, encoded in a vendor's proprietary format, governed by a vendor's proprietary policies, and increasingly difficult to migrate precisely because the value compounds with every agent interaction that builds on them.
The CIO who adopted Copilot for productivity two years ago did not make a decision about the enterprise AI OS. But they may have made a structural commitment to Microsoft's context layer that will shape enterprise AI governance for the next decade - because the context that has been built, the integrations that have been configured, and the agents that have been deployed on Microsoft's surface are all compounding within Microsoft's governance model, not the enterprise's own.
The three questions that reveal whether the enterprise is at risk
The exposure to unintended control plane lock-in is not binary - it exists on a spectrum. Three questions reveal where on that spectrum the enterprise sits.
Can you swap the foundation model without rebuilding your context layer? If changing from GPT-4 to Claude requires significant rework of the knowledge graph, semantic definitions, or retrieval infrastructure that agents depend on, the context layer is model-specific.
That is not a context layer - it is a model-specific integration that will need to be rebuilt with every significant model change.
Who owns the semantic definitions your AI agents reason from? If the answer is "Databricks Unity Catalog," "Google Knowledge Graph," or "Microsoft Purview," the enterprise has ceded the most strategically important layer of its AI OS to a vendor.
The semantic definitions are the layer that encodes how the enterprise understands itself. They should be owned by the enterprise, hosted in a vendor-neutral format, and portable.
Can the control plane be audited independently of the vendor that provides it? If the audit trail for AI agent actions lives exclusively within a vendor's infrastructure - accessible only through their tools, governed only by their retention policies - the enterprise's ability to demonstrate regulatory compliance is dependent on that vendor's cooperation. In regulated industries, that is not an acceptable dependency.
What conscious participation in this battle looks like
The enterprise that recognises it is in this battle - and chooses to participate consciously rather than by default - makes different decisions than the enterprise that treats each AI vendor selection as an isolated procurement choice.
It adopts open protocols - MCP, A2A - that allow agents to communicate with context and data sources without binding the agent to a specific vendor's runtime. It builds the knowledge graph and semantic layer in a format that is portable across model providers and agent frameworks. It treats the audit trail and governance policies as enterprise-owned assets, even when the enforcement infrastructure is vendor-provided. And it evaluates each vendor's AI platform offering not just on capability but on the governance independence it preserves or forfeits.
None of this requires rejecting Microsoft, Google, or Databricks. These platforms provide genuine value and will continue to be part of most enterprise AI stacks. What it requires is choosing which layers of the AI OS the enterprise builds inside those platforms versus which layers it builds to be portable across them.
The model is a commodity. The application is a workflow. The context and control layer is the operating system. The enterprises that build their AI OS consciously - owning the semantic layer, governing the control plane, preserving portability across models and frameworks - are the ones that will still own their AI strategy in five years. The ones that let it accumulate inside vendor stacks by default are already in a battle they have not yet recognised they are losing.
