AI’s Context Interoperability Imperative
For two years the boardroom AI debate has centred on one question: which model? That question is fading. Frontier models are converging in capability and are increasingly interchangeable behind standard interfaces. The question replacing it is context interoperability: can an enterprise supply reliable business context to any AI agent, authorise what that agent does with it, and retain control when it acts, regardless of which model or vendor sits behind it?
AIMG Research’s latest analysis concludes that context interoperability is not a single protocol market. It is a layered architecture and governance priority, and treating it as a connector problem is the most expensive mistake leaders can make.
Four layers, not one
By enterprise context we mean everything needed to interpret a business situation and act appropriately: definitions, relationships, policies, permissions, workflow state and history. Making that context interoperable requires four distinct things.
Technical interoperability asks whether components can discover capabilities and exchange structured requests. Semantic interoperability asks whether they interpret business terms such as revenue, customer and risk consistently. Operational interoperability asks whether a multi-step task can proceed with valid prerequisites, approvals, error handling and verification. Governance interoperability asks whether the organisation can identify who is acting, authorise each operation, explain what happened and stop further execution.
Today’s standards address the first two layers well. The last two remain firmly the enterprise’s responsibility.
The standards: access, meaning, execution
The Model Context Protocol (MCP) has become the common layer for technical access. It originated at Anthropic in November 2024, moved into neutral stewardship under the Agentic AI Foundation in December 2025, and sits in a portfolio that grew to ten projects last week. Apache Ossie, formerly Open Semantic Interchange and now incubating at the Apache Software Foundation, targets the semantic layer, allowing definitions of metrics, dimensions and relationships to move between analytics and AI tools. The Model Hardware Standard (MHS), in research preview since August, extends context interoperability to programmable laboratory and industrial equipment.
These are complementary layers, not rival standards. Enterprises will combine them rather than pick a winner. But a protocol-compliant connector only proves an agent can reach your ERP. It does not prove the agent is using your definition of revenue, respecting your authorisation boundaries, or able to recover from a half-completed transaction. Access without agreed meaning produces confident answers to the wrong question.
Where the value moves
As access standardises, connectors commoditise. The defensible control points move up the stack to business context, policy enforcement, workflow execution and assurance. The market already reflects this. The Salesforce–Anthropic partnership separates the conversational interface from the system that owns business rules. Snowflake’s agreement to acquire Natoma is a bet on identity, policy and audit at the tool-call level. Neo4j and DataHub position ontologies and context management as products in their own right. And the announced Schneider Electric–PTC deal, around $22.6 billion in equity value, shows the strategic weight now placed on joining engineering and operational context.
For product leaders at enterprise software vendors, “supports MCP” will soon be table stakes. Differentiation will come from which layer of context interoperability you own, how you prove it works, and whether customers can leave without losing their context and audit history. Proprietary connectors are a waning moat; demonstrable semantic fidelity and governance are the durable one.
Read the evidence carefully
Early results are encouraging but easily misattributed. Waters reports that in a pharmaceutical pilot its AI-powered peak integration cut manual interventions across 400 chromatograms from 90 to four. That is a company-reported pilot of a specific AI workflow, not proof of what an interoperability standard contributes. Business cases should separate model gains, process redesign, integration savings and platform effects, and count the full operating cost, including human review.
Governing interoperable context
Every interoperable interface is also a trust boundary. Official MCP security guidance documents confused-deputy attacks, token passthrough, server-side request forgery and excessive privilege. The minimum controls: an accountable owner and mandate for every production agent; least-privilege authorisation at each tool boundary; approvals bound to exact parameters so an agent cannot alter the action after sign-off; reviewed server admission; end-to-end observability; and tested revocation and shutdown. Agent output is never an authorisation decision. Nor does adopting a standard settle regulatory status under the EU AI Act. Auditability means reconstructing what happened and why it was permitted.
What leaders should do now
Keep authoritative definitions and permissions outside the model, supply only the context a task needs, and verify every action at the execution boundary. Then start with one bounded workflow. In the first 30 days, document its baseline, data owners, semantic definitions, permitted actions and stop conditions. In the next 30, implement approved servers, scoped credentials, gateway policies and approval binding, and test against unauthorised requests, poisoned content and partial failures. In the final 30, pilot against a held-out task set and expand only when business and risk owners sign off on the evidence.
Ask every vendor to demonstrate, on your own tasks and failure scenarios, semantic fidelity, authorisation depth, observability, resilience, portability and full operating cost.
The winners of the next phase of enterprise AI will not be those with the most agents connected. They will be those whose context is interoperable by design: shared in meaning, governed in action, and stoppable at will.