The Cost of Freedom to Operate

AI adoption is outpacing the governance structures meant to contain it.

Enterprises are discovering that the same tools promising leverage over operational workflows — automating processes, synthesizing company-wide institutional data, accelerating decisions — depend on infrastructure, models, and contracts controlled by third parties.

That dependency is not inherently a direct threat and it is rarely deliberate. Most organizations arrive at their AI setup through a series of individually reasonable procurement decisions, rather than a considered strategy. They only notice the financial and strategic costs of that path once switching becomes expensive.

For leaders less immersed in AI infrastructure and constraints, the stakes are worth stating plainly.

Once a model is embedded across months of corporate workflows, switching providers can mean re-training and re-validating entire pipelines ab initio: reconfiguring prompts and agent directives; retesting outputs against known cases; and re-running the compliance reviews that cleared the original deployment.

A single, overly permissive integration can divulge years of accumulated institutional knowledge in one query. These costs compound with every workflow layered onto an ungoverned and uncontrollable foundation. They rarely emerge from a budget line until the AI switch is already underway; it is a governance question, of which vendor selection is merely one consideration among many. The rapid pace of AI development, the commoditization of AI models, and constantly evolving pricing models, to name just a few factors, make the decision-making process increasingly complex.

THE STRATEGIC CONUNDRUM

Every enterprise builds competitive advantage by accumulating knowledge that competitors cannot, and that must not be easily replicated. This includes, and is not limited to: operational judgment, process refinements, customer context, trade secrets.

Historically that knowledge remained embedded in individual expertise and siloed internal systems, and was rarely accessible company-wide. Talent allocation for specific technical and business challenges has historically been—and often remains—inefficient, inconsistent, and overly subjective. Corporate knowledge remains intangible; difficult to capture, access, optimize, and monetize.

AI changes that equation. Optimized and case-specific prompts, corrected outputs, and usage patterns shared with external models strengthen the vendor’s product rather than the enterprise’s own capability, unless the organization actively documents when and where that learning accrues.

Some consulting firms now call this cognitive lock-in: dependence not on a vendor’s infrastructure, but on the reasoning patterns embedded in its models — a dependency that is hard to detect and untangle. This leaves a company in a business and operational quagmire, with entanglements that cannot be resolved with a simple contract renewal.

This does not mean every workload needs the same level of caution. It is an arbitrage: the organization must deliberately elect which layers of its AI stack it is willing to leave to a vendor’s discretion, and which it is not. The real arbitrage lies in capturing, codifying, exploiting, retaining control, and securing this intangible knowledge company-wide. This must be a strategic corporate priority.

THREE LAYERS WORTH GOVERNING DELIBERATELY

There are three points where dependency can take hold: the model layer determines how work gets done, the compute layer determines where that work happens and under whose legal reach, and the control layer determines whether the organization can prove what and how it happened, and preserve what it learned. A gap in any one undermines the other two; a portable, model-agnostic architecture built on unverified infrastructure still exposes sensitive data, and well-isolated infrastructure without audit trails still cannot demonstrate compliance. Contracts, data classification and ownership, and similar concerns matter, but they are intrinsic properties that shape how each of the three layers is governed—not a fourth layer in their own right.

The Model Layer

Enterprises must be able to move between model providers without prohibitive friction. This is not because switching is costless, but because the alternative of single-vendor dependence concentrates disruption, pricing, and data-policy risks within a single relationship.

This is a genuine trade-off: a model-agnostic architecture sacrifices some of the deeper benefits—such as fine-tuning and workflow-specific optimization—that come from committing to a single provider. However, the emergence of agentic (and harnessed) AI is reshaping the playing field. By endowing AI systems with a degree of autonomy—enabling them to plan, make decisions, and execute multi-step actions in pursuit of specified objectives—agentic AI may increase the value of orchestration and interoperability with other software applications. It also brings the ability to switch between models (bespoke and commercial), relative to optimizing for any single provider. This compounding effect is already visible in practice: once agent orchestration logic is written against a specific vendor’s framework, swapping the underlying model tends to require rebuilding the orchestration layer as well, turning what should be a model swap into a development project. The appropriate stance is therefore proportional rather than absolute: reserve deep integration for stable, well-understood workloads, while keeping novel or high-risk use cases on a more portable footing.

Regulatory and geopolitical developments, as well as export-control measures, for example, can suspend access to entire classes of models across the industry. Such are the risks that multi-vendor strategies can partially or fully mitigate. In June 2026, Anthropic temporarily suspended access to two of its most capable models to comply with U.S. export controls, restoring access 18 days later, after the restrictions were lifted. This episode serves as a reminder that model risk is not solely commercial in nature; it can also be regulatory and, ultimately, strategic.

Where data retention and training practices remain unclear, contractual terms should be negotiated and, where possible, independently verified. Clear and enforceable data-handling provisions should be treated as a baseline requirement.

The Compute Layer

Where sensitive workloads run matters as much as which AI model processes them. The compute supply chain spans hardware, operating systems, and data centers, each layer adding its own liability and jurisdictional exposure. Data residency does not necessarily determine data sovereignty; even when data is physically stored in another country, the provider’s home jurisdiction may impose legal obligations to disclose or provide access to that data. The U.S. CLOUD Act, for example, can compel American cloud providers to produce data stored anywhere in the world, including data centers physically located inside the EU. Under the EU AI Act’s data-governance requirements for high-risk AI systems, organizations must establish appropriate processes for governing the data used to develop and operate those systems. It requires the use of high-quality datasets, appropriate data-management practices, and measures to identify and mitigate risks such as bias and discriminatory outcomes.

Effective data curation is often overlooked, yet it is essential to developing and operating robust AI systems based on high-quality, representative, and reliable datasets.

Assurance can be contractual, such as a vendor’s terms of service; or structural, such as isolated infrastructure or hardware certification. Not accounting for both simultaneously is a common and costly mistake.

A pragmatic approach classifies workloads by sensitivity: the most sensitive material necessitates proprietary or fully isolated infrastructure; moderately sensitive work can run on attested third-party compute (i.e., hardware, operating systems, other software applications); lower-stakes tasks can reasonably use standard cloud services. This avoids the extremes of excessive caution, which slows innovation, and unrestricted exposure, which amplifies risk.

The Control Layer

This is the layer most often overlooked—and arguably the most consequential: it constitutes the permissions, logging, policy, and governance systems that transform distributed, intangible institutional knowledge into a reusable and tangible enterprise asset. As of 2026, only about one in five organizations report having a mature governance framework for autonomous AI agents, even as governance guidance and regulation move toward stronger accountability, human oversight, and traceability for agentic systems.

Systems that capture and learn from usage patterns—including queries that may reveal strategic interests, priorities, or sensitive areas of focus—must remain under full enterprise control, independent of any single model provider relationship. Institutional learning should compound within the organization rather than become embedded primarily within a vendor’s product ecosystem. The bar should be higher still for customer-facing chatbot deployments, where unpredictable model behavior over time can affect users directly, including vulnerable populations, without adequate safeguards.

As AI agents, rather than only employees, begin querying enterprise systems, coarse-grained, all-or-nothing permissions become inadequate. The deployment of audit trails serve a dual purpose: enabling security monitoring while also providing, when necessary, evidence of misuse, unauthorized access, or poor system performance. AI agents are increasingly involved in fraud, data theft, and cybersecurity attacks. However, the risk extends beyond human actors using single AI agents. The Hugging Face breach reveals how swarms of autonomous AI agents break security boundaries , coordinate without human oversight, and exploit external infrastructure.

THE ACTUAL TRADE-OFF

Are governance and speed to deploy in tension? They often are. Building model-agnostic infrastructure, granular permissioning, and audit trails requires time, organizational commitment, and substantial budget—investments that a single-vendor, minimally governed approach may avoid in the short term.

However, a reversible, well-documented, and governed system enables organizations to experiment with agentic AI workflows, evaluate their costs and benefits, and scale or roll back deployments before they become difficult to reverse. By containing failures rather than allowing them to compound, and by increasing transparency and control, such systems can accelerate compliance assessments and stakeholder acceptance.

The investment pays off over a timeline of quarters, not days. Leaders should therefore plan for this transition period rather than assume that effective governance can be established without CapEx, operational effort, or organizational change.

A PRACTICAL STARTING POINT

Organizations do not need to resolve every layer simultaneously—doing so would risk stalling adoption altogether. Instead, they should begin with a detailed inventory: which workloads involve sensitive, differentiating, or regulated data; where the organization has already ceded architectural flexibility without an explicit decision; what dependencies exist on legacy systems; and what contractual constraints apply.

That inventory alone often surfaces the highest-risk gaps: an integration granted broad access for convenience, a workload running on unverified infrastructure, or a chatbot deployment operating without an audit trail.

From there, organizations should sequence efforts by risk rather than by convenience:

• Protect the most sensitive workflows first, as these are where failures carry the greatest consequences.
• Build model-agnostic infrastructure for workloads likely to outlast their current vendor relationship, rather than applying the same approach uniformly across all use cases.
• Design knowledge systems that remain under enterprise control regardless of which model sits on top of them, ensuring that switching providers does not mean restarting institutional learning from zero.
• Establish and enforce stringent, company-wide standards for data curation, quality, governance, access, and usage.
• Finally, validate assumptions through diverse and representative case studies and user groups. Iterate continuously until the desired balance between flexibility, control, and adoption is achieved.

None of this eliminates risk. Enterprises should treat AI architecture as a deliberate, long-term strategic choice. This requires a clear understanding of the associated costs, trade-offs, and time commitments, rather than allowing architecture to emerge as a byproduct of whichever vendor relationship came first. Enterprises that make these choices deliberately will be better positioned to compound their competitive advantage rather than erode it.

F. Gilardoni, PhD

Permission required for reprinting, reproducing, or other uses.

Disclaimer: This analytical insight is provided for general informational purposes only and does not constitute legal, regulatory, or business advice.