Sovereign Substrates

AI adoption in the enterprise has been slow, expensive and inconsistent.

Despite significant financial commitment, several organizations have reported exhausting their AI budgets earlier than anticipated, and having to place caps on expenses mid-year without concrete results.

The reasons are many and they point to the organization and not the technology. Enterprise finance needs predictability in costs, and a way to understand ROI.
There are operational mismatches as organizations tend to work in vertical silos whereas AI operates best with lateral access to understand and access cross-functional workflows and data. Metrics for measuring success vary, internal expertise is unevenly distributed, vendor lock-in and data protection risks are real, and with geopolitical entanglement endorsed at the governmental level, there are policy and access risks.

At a higher level there are governance, ownership and budgetary concerns, and this is before model selection, routing and varying API costs.

These bottlenecks are pre-inference; before an AI model is called.

Cheaper models are not the solution

If financial costs were the only obstacle, the availability of excellent and cheaper OSS models from Chinese labs Z.ai, Tencent, Deepseek, Minimax, and Kimi would have circumvented the problem and the current landscape would look very different.
Lower costs, though desirable and advantageous, haven't proven to be the key that unlocks scale in the enterprise.

The Missing Foundation

AI, despite being intelligent, multimodal technology capable of reasoning isn't plug-and-play because it requires inputs and context, permissions and provenance, and clear security and governance controls.

The ecosystem required to make AI work inside the enterprise requires a significant architectural change; a sovereign, pre-inference, model-agnostic layer inside the organizational boundary which augments existing infrastructure, assets and workflows, while protecting proprietary data and processes.

Context is continuously produced by legacy applications, which means either updating those apps, or adding a reliable data conversion layer to their output, so that a machine-readable substrate becomes the organic byproduct of everyday use.

Beyond this, the core can be extensible but requires at least three additional layers:

• A canonical substrate to establish organizational state (what is true now), and memory (what has been learned) including context (what is relevant to the task) and provenance (where it came from)
• A governance layer for policy, identity and permissions
• An execution layer for model routers, agent runtime and tools etc.

Why This Design

The criterion outlined a year ago in Agents of Chaos is more obvious today.

"Maintaining stateful interactions and memory adds complexity and most companies aren’t ready to invest in the infrastructure needed for real agent-like systems when the technology arc hasn’t reached maturity. Even in the B2B context, it is a reenactment of the Cloud-based vs on-prem debate as organisations require personalisation - knowing their structures, workflows, repositories - for the technology to be truly effective, and this has security and dependency implications."

This encapsulates the struggle enterprises have with integrating AI today: vendor dependency, governance and policy, memory and context. Coding agents have continued to improve as expected, arguably more than expected. That conclusion still holds: true agency requires personalization, in consumer and enterprise contexts alike. In the enterprise, personalization is a durable, governed understanding of owned and protected structures, workflows and repositories.

Separating knowledge from governance, policy and execution gives organizations control over context, identity and proprietary data prior to inference. Additionally, prefixing data conversion pipelines ensures that used data is current, and integration efforts minimize friction.

Sovereign Does Not Mean Self-Built

This is likely to operate best as a composable substrate with a mixture of internal and external tools. The effort to build everything in-house would be cost-prohibitive even if it were possible. No organization, even the most sophisticated tools is capable of creating, maintaining and operating all of its software without diluting its focus.

Sovereignty here means absolute control with interchangeable components making up the less important parts of the stack, while its core is owned and doesn't require the organization to be beholden to any specific vendor for its operations and growth. Most importantly perhaps, the substrate illustrated above lets most human capital continue doing what it does without having to relearn or upskill, while simultaneously infusing the organization with the intelligence and efficiencies of AI.

A composable sovereign substrate is the architectural requirement for seamless AI integration in the enterprise.

Anomaly's Stance: Composable Proof-of-Concept

We believe AI fundamentally changes the operation, value and capability of software, and this requires a new architecture for human-software interaction.

As inference costs collapse, capability standardizes and intelligence converges at comparable levels, the catalyst differentiating performance for the enterprise class will be the substrate and not the model. Building this infrastructure inside the organizational boundary is strategically more sound and sustainable, than chasing models and re-platforming annually while being vulnerable to variable costs and data leakage.

To illustrate this principle, we’ve built a narrow, composable component that can slot into any organization's data conversion layer. Skygate is a candidate intelligence data platform to help organizations create structured, machine-readable candidate objects.

Skygate dissolves talent data to make it polymorphic: information imbued with meaning and structure, separated from its original form, machine-operable, and recomposable into whatever output that context demands.

Recruitment firms and large employers can turn resumes and hiring documents into structured, reusable data that can power existing systems and the next generation of AI-ready and operated applications, while keeping data and workflows private, remaining model-agnostic and increasing productivity.

It is upon a foundation with similar traits that hybrid human/machine corporate systems and structures will likely be built.