Subscribe

Sign In

Created For Sponsor

AIArtificial IntelligenceCognizant

AIArtificial IntelligenceCognizant

sponsored

The Sovereignty Imperative: Why Institutional Intelligence Must Outlive Any Model

The Sovereignty Imperative: Why Institutional Intelligence Must Outlive Any Model

By Vijay Narayan and Samih Fadli

Share

Newsweek is a Trust Project member

As enterprise AI adoption rises, it is creating a new dependency. When businesses connect AI models to their operating context, rules and workflows, important new capabilities begin to depend on how a specific model behaves. If that model is restricted or changed, the business may not lose its data, but the AI capability built on it can become operationally unusable.

The exposure is sharpest with externally hosted foundation models, where enterprises have limited control over access, model evolution and operating terms. These models are some of the most valuable tools a business can adopt, however, so the answer is not to rely on them less but to build around them more deliberately.

To preserve continuity and allow enterprise intelligence to be reused across models and workflows, organizations need an enterprise-controlled, portable layer of memory, ontology and governance that is independent of any model.

The new reality of enterprise AI risk

The structural dependency that’s now becoming visible runs deeper than the vendor relationships enterprises are used to from the past. When a traditional software provider changed its terms, what was at stake was access, cost or functionality. When a foundation model changes, however, the business must do significant work to rebuild, test and govern how the AI-enabled capability produces decisions and takes actions. In this new reality, organizations are building deep operational capability on a ground they do not own and cannot fully control.

The risk has become concrete in recent months as across the industry, enterprises have experienced disruptions to AI capabilities from regulatory directives, the retirement of commercial models, and provider-side changes that occurred on timelines outside of their control.

In each case, despite retaining their data, the affected organizations lost something significantly more difficult to replace: the customized reasoning configurations, behavioral adaptations and automated workflows built on top of specific model architectures.

Current events are early signals of a structural pattern that will likely intensify as regulatory environments tighten and foundation model providers continue their rapid competitive cycles.

How enterprise knowledge and AI models become fused

A foundation model brings reasoning: the general ability to recognize patterns and respond fluently across modalities and domains. That is powerful, but also generic. Most of what makes a model useful to a specific enterprise comes from somewhere else: the enterprise’s own knowledge, its operational decisions, the judgement of its people, the definitions that keep its data meaning the same thing across the company.

The risk lies in how the two come together.

To make a model useful, an enterprise has to teach it how the business works. For example, a health care organization fine-tunes a model on years of prior-authorization decisions, capturing how policies are interpreted in borderline cases. A manufacturing business builds prompt libraries that teach an AI assistant its maintenance alerts, equipment histories and safety rules.

Some of the assets like data and rules are portable and sit in systems the enterprise controls. Yet the resulting capability is none of these in isolation, it is the validated coupling between those portable assets and one model’s particular behavior: the fine-tuned weights, prompts shaped to one model’s idiosyncrasies, the integration wired to one provider’s API and the months of testing that proved the system decides correctly in the cases that matter.

Building for continuity

Because that coupling can’t easily be undone once it forms, the answer is to create a sovereign layer where rules and decisions can be applied and recorded. That sovereign layer sits between the enterprise’s operational data and the external reasoning engines it employs, and serves three functions:

First, it acts as the independent repository for the enterprise ontology: the shared definitions and semantic structures that give consistent meaning to enterprise data across every AI interaction. This ensures that when the underlying model changes, the ontology remains intact.

Second, it maintains institutional memory. Every significant AI-assisted decision along with its context and reasoning chain is recorded in an enterprise-controlled ledger independent of the model that produced it; it compounds over time and persists through any model transition or provider change.

Third, it enforces governance. Safety constraints, compliance and privacy rules, risk thresholds and the point at which decisions must defer to a human are all defined and enforced by the enterprise itself, not delegated to the built-in safety mechanisms of external models. When a new model is introduced, the enterprise’s governance framework is applied to its outputs consistently, ensuring standards are upheld regardless of which reasoning engine is in use.

By maintaining these three capabilities in its own hands, an organization gains the flexibility to adopt new models as they emerge and operate across multiple providers at once. The added benefit is an ability to deploy sovereign environments for sensitive operations, without losing the institutional knowledge that defines their competitive advantage.

This is not a hypothetical exposure. Rules for advanced AI are tightening and are not always aligned across geographies. Providers release new versions and retire old ones in an intense competitive cycle, and geopolitical dynamics are increasingly affecting the cross-border availability of advanced technology capabilities.

For decades, software was treated as the primary asset and data as an input. AI shifts that balance. Models may change, but the institutional intelligence that makes them useful must endure: its reasoning patterns, feedback loops and accumulated judgement.

Most often, leaders ask which model to build on. An equally important question is how to build so that when the model changes, the enterprise’s intelligence carries forward intact.

Vijay Narayan is EVP and Head of Manufacturing, Logistics, Energy and Utilities at Cognizant, where he also leads the company’s global Physical AI business.

Samih Fadli, PhD, is CTO for AI for Physical AI at Cognizant.

sections

company

editions

contact

terms of use

sections

company

editions

contact

terms of use

© 2026 Newsweek Digital LLC