Agentic AI will not scale until organisations stop treating context as a collection of metadata tables and start treating it for what it truly is: a control plane for meaning, policy, identity, and truth.

The change is simple, but profound: Context is not a dataset. It is the layer that determines what the system is allowed to believe and what agents are permitted to do.

In practice, this control plane will not exist as a single system. It will be implemented through a combination of identity services, policy engines, semantic layers, and governance platforms. What matters is not where it technically resides, but that it behaves as a coherent layer governing how autonomous systems interpret reality and act within it.

This fundamentally reframes the challenge. It is no longer about adding more data, but about building the architectural layer that governs autonomy itself.

The Era of Operational Context

Organisations already manage multiple fragments of context: identities, glossaries, lineage, policies, MDM hubs, taxonomies, metric layers, and consent rules. However, these elements exist in isolation and share two critical weaknesses: they do not agree with each other, and, more importantly, they do not execute anything.

This misalignment becomes a structural problem when autonomous agents are introduced. These systems cannot operate responsibly in an environment where definitions conflict, policies remain trapped in static PDFs, and lineage exists solely for audit purposes, with no impact on real-time decisions.

The real bottleneck is no longer the absence of context. It is the absence of governed, authoritative, real-time context.

In a mature operational setting, the system must be able to answer, instantly, fundamental questions: what is the authoritative definition right now? Which policy applies to this specific purpose, in this jurisdiction, for this user? What was considered true at the exact moment the agent acted? And when something goes wrong, who is accountable for decisions made autonomously?

Historically, humans have absorbed inconsistency through judgement and implicit context. Autonomous systems cannot, and when they fail, they do not fail slowly. They scale errors at machine speed.

Any organisation seeking to scale AI responsibly will inevitably converge on five foundational capabilities. These are not optional features; they are architectural requirements that define a true system of context. They form a dependency graph, and treating them as a flat checklist is one of the most common ways mature-sounding programmes quietly stall.

The first is the ability to resolve identity: a stable, verifiable, time-aware view of entities and their relationships across domains and systems. This is what anchors agents and prevents decisions based on inconsistent representations of the world. It is also the foundation on which everything else rests. Semantics cannot stabilise until identity is resolved.

The second is representing meaning, not as static documentation, but as semantic models that execute. Definitions, relationships, events, and metrics must be machine-interpretable and capable of influencing behaviour in real time. Without stable identity beneath them, semantic models simply inherit the conflicts they were meant to resolve.

The third capability is enforcing policy. Access, consent, retention, and jurisdiction rules must no longer be passive guidelines; they must be evaluated dynamically at the point of each decision, based on context and purpose. But policy cannot be made truly executable until the semantic layer beneath it is coherent and authoritative. This is where many organisations discover that their governance programmes have been building on unstable ground.

The fourth is maintaining current context. Truth must exist in near real time, because stale context breaks autonomous workflows faster than incorrect data ever did. In autonomous systems, context latency is operational risk.

The fifth capability is providing proof. Every decision, action, and policy evaluation must be explainable, traceable, and defensible, not just for audit, but as an intrinsic part of the system.

These five capabilities will form the foundation of a new context operating model: the layer where autonomy and accountability no longer conflict but coexist by design. But reaching that state is not a matter of architectural ambition alone. It is, for most organisations, a matter of years and hard organisational work.

By 2028, these five capabilities will form the foundation of a new context operating model — the layer where autonomy and accountability no longer conflict but coexist by design.

It is important to emphasise that this is not an incremental evolution of traditional governance. That model was designed for humans, at design time, assuming slow decisions, manual review, and human judgement as the final safeguard. Context systems, by contrast, exist to govern machines at runtime. When AI systems can reason and act, policies must be executable, semantics must influence behaviour, and controls must operate continuously. This is not maturity. It is a structural break.

From Data Management to Context Stewardship

The shift underway is profound: we are moving from governing data as a static asset to orchestrating decisions as a continuous, governed capability.

Emerging platforms will embed agents capable of automatically resolving conflicts, enriching entities, detecting policy violations, assessing risk, and correcting context drift, all continuously.

But this transformation is not purely technical. In practice, context fails more often due to organisational misalignment than technical limitations. Each of the five capabilities described above crosses deeply contested organisational seams. Identity resolution alone typically spans IAM teams, MDM, data engineering, and legal. Executable policy enforcement touches InfoSec, compliance, and data governance, groups that often hold conflicting definitions of what a policy

even is. The coherent control plane this model requires presupposes either a genuinely empowered central function or a level of federated trust that takes years to build deliberately. Neither arrives automatically.

This is why context stewardship becomes central to the operating model. Governance stops being a slow, manual process and becomes a competitive advantage. But the organisations that will realise that advantage are the ones that treat the authority, accountability, and sequencing questions as seriously as the architectural ones. The value proposition evolves from “we centralise and clean your data” to “we automate decisions safely, transparently, and at scale, because context is governed and executable.”

Making Semantics and Governance Executable

As AI responsibility rises to board level, two structural failures can no longer persist.

The first is “semantic theatre”: glossaries and ontologies that exist but never influence real system behaviour. The second is “governance by backlog”: manual approval processes that simply do not scale in a world of autonomous decision-making.

The only viable path forward requires a clear shift: semantics that shape behaviour, policies expressed as code, continuously validated and versioned context, and governance enforced at machine speed. This is the line between organisations that talk about AI governance and those that truly operationalise it.

The Question That Now Matters

For years, organisations have asked: “Do we have AI-ready data?”

That is no longer the right question.

The question that defines the present and the future is this: do we have decision-ready context? Do we have the control plane that governs autonomous behaviour? And just as importantly: do we have the organisational structure to build and sustain it?

Context is not something you finish. It is something you continuously define, in relation to the decisions that need to be made, and the authority structures that make those decisions trustworthy.

That race has already begun, and it will define the next decade of enterprise architecture.

Within this decade, competitive advantage will not be determined by model size, but by the ability of platforms to close the full loop: decide, act, measure, improve, and govern. The organisations that get there first will not be those with the most sophisticated architectures on paper. They will be the ones that resolved the harder question of who owns meaning, who enforces policy at runtime, and how those two functions stay coherent as the system scales.

The control plane may be technically federated, but it must behave as a coherent layer. Without that coherence, autonomy collapses. Platforms that successfully close this loop accumulate advantage over time. Others remain stuck integrating tools, never truly improving outcomes.