The AI Dependency Risk Boards Can’t Ignore

For the last two years, most executive conversations about artificial intelligence have followed a familiar pattern.

Which model is best? Which vendor is safest? Which platform should we standardize on? Should we use Anthropic, Google, Microsoft, OpenAI, etc., or one of the emerging Chinese models? Should we let employees use consumer tools? Should we build internally? Should we wait?

These are reasonable questions. They are also too narrow.

The more important question is no longer simply which AI model is best today. The more important question is whether your organization should build its future AI operating model around one provider, or whether it should design an architecture that can use many models, route work intelligently, and adapt when technology, cost, regulation, or geopolitics changes.

That is the strategic importance of the model-agnostic AI harness.

The phrase sounds technical. It is not. At its simplest, a model-agnostic AI harness is the organizational layer that sits between people, workflows, systems, and LLMs (large language models). It lets a company decide which model should handle which task, under which policy, in which jurisdiction, with which data, at what cost, and with what level of human review.

Think of it as the AI equivalent of choosing your modern work platform. In the last generation, boards and executives had to decide whether the enterprise would standardize on Microsoft 365, Google Workspace, or another collaboration suite. That decision shaped identity, email, document management, meetings, security, compliance, data retention, employee habits, and vendor dependency for a decade.

AI will be bigger.

The decision facing leaders now is not simply whether employees can use chatbots. It is whether the organization will allow one AI provider to become the default brain, interface, workflow layer, and governance system for the enterprise, or whether it will build enough internal control to avoid becoming captive to one model, one vendor, one jurisdiction, or one regulatory regime.

This is not a theoretical concern. It became real when the U.S. government directed Anthropic to suspend access to its advanced Fable 5 and Mythos 5 models for foreign nationals. Anthropic, facing compliance obligations, disabled access to both models. Whatever one thinks of the national security rationale, the signal to the market was unmistakable.

Access to frontier AI is no longer just a product decision. It is a geopolitical risk.

For a C-suite team or board, this should change the conversation.

The next AI strategy discussion should not begin with, “Which model do we like best?” It should begin with, “How exposed are we if the model we choose becomes unavailable, restricted, repriced, politically constrained, legally challenged, or strategically misaligned with our operating footprint?”

That is the journey leaders now must take.

The first lesson: AI is not one thing

Before discussing the strategic implications, executives need a simple vocabulary.

These are all companies: Anthropic, OpenAI, Alphabet, Mistral.
These are all harnesses: Claude, ChatGPT, Google Gemini, Mistral Vibe.
These are all models (LLMs): Fabel 5, ChatGPT 5.5, Gemini 3.1 Pro, Mistral Medium 3.5

Claude is Anthropic’s product environment and brand. Many people call it a chatbot, an assistant, or an autonomous agent interface. For business purposes, it is helpful to think of Claude as a harness: the interface, workflow, safety, memory, tool-use, and orchestration layer that lets users interact with an underlying model.

Fable 5 is the model. It is the brain in this analogy. Claude is the body, cockpit, control surface, and safety system around it.

The distinction matters.

A large language model, or LLM, is the statistical reasoning engine trained on vast amounts of text, code, images, and other data. It predicts, reasons, writes, summarizes, classifies, translates, codes, and increasingly operates software tools.

A harness is the environment that lets people or systems use that model productively. It may include a chat interface, enterprise permissions, document retrieval, audit logs, connectors to business systems, prompt libraries, workflow automation, memory, agent tools, cost controls, compliance policies, human approval steps, etc.

A routing platform decides which model should handle a given request. A simple request might go to a cheaper local model, possibly even in-house. A complex legal, engineering, or strategic analysis might go to a frontier model. Sensitive requests involving European customer data might be routed to an EU-hosted model. Coding tasks might go to a model that is especially strong at software development. You get the picture, different LLMs are better suited and compliant with certain tasks and those requests can be routed to them.

A closed model is a model whose weights, training details, and operating environment are controlled by the provider. Most leading frontier models today are closed. Customers access them through an application or API, but they do not possess or control the model itself.

An open-weight model is different. The model weights are made available for others to download, inspect, modify, host, or run in their own environments, subject to licensing terms. Open-weight does not always mean fully open source in the traditional software sense, but it gives customers more control than a closed model delivered only through a vendor-hosted service.

These distinctions are not academic. They define control.

If your company uses Claude directly, you are relying on Anthropic’s product environment, Anthropic’s models, their safety architecture, commercial terms, and regulatory exposure.

If your company uses a model-agnostic harness, you may still use Anthropic, OpenAI, or Google. But you are not structurally dependent on one of them. The harness becomes your control layer. The models become suppliers.

That distinction will define the next phase of enterprise AI.

The second lesson: the AI race is a repeated game

The Fable 5 incident is best understood through game theory.

In a simple game, one side makes a move and the other side responds. In a repeated game, each move changes future incentives. Every action teaches the other players what kind of game they are in.

The United States government has a legitimate concern. Frontier AI systems may increase risks in cyber operations, biological research, military planning, intelligence gathering, and autonomous tool use. No serious board should dismiss those concerns. Powerful AI can be economically productive and strategically dangerous at the same time.

But when the government restricts access to a leading model, it does more than reduce near-term misuse. It also changes the expectations of every other player.

China sees that U.S. model access can be restricted. Europe sees that U.S. AI companies may be subject to political control. Multinational companies see that access to a critical technology can be interrupted. Developers see that closed models can be removed from the toolchain. Customers see that the best model may not be the most reliable model.

This is where game theory becomes useful.

The U.S. government may believe it is playing a security game: restrict the most dangerous capability and reduce risk. But the global market may interpret the move as a reliability game: do not build critical infrastructure on a platform that can be switched off by another government.

China may interpret the move as an acceleration signal: dependence on American AI is strategically unacceptable, so domestic alternatives must improve faster.

Europe may interpret it as another confirmation of digital sovereignty: cloud, SaaS, chips, data, and AI cannot be permanently dependent on American firms and American political decisions.

Organizations may interpret it as an architectural warning: do not make one model the single point of failure for the company’s future operating model.

The result is a different game from the one policymakers may think they are playing.

The U.S. may gain control over one model, but lose trust in the platform layer. It may reduce short-term access to a frontier system, but accelerate long-term investment in alternatives. It may protect a capability edge, but train the global market to route around American providers.

That is the strategic risk.

The third lesson: bans can create competitors

The semiconductor example is instructive.

The U.S. export controls on advanced Nvidia chips were designed to slow China’s ability to train and run frontier AI systems. In some ways, they did. Access to the most advanced compute still matters. But the controls also encouraged China to invest more aggressively in domestic chips, model efficiency, distributed training, software optimization, and open-model ecosystems.

This is the paradox of technological containment. It can slow an adversary. It can also teach the adversary to become more self-reliant.

The same logic now applies to AI models.

If American frontier models remain broadly available, global developers, enterprises, and governments have a strong incentive to build around them. The U.S. benefits from user feedback, enterprise integration, developer mindshare, safety testing, revenue, and platform dependency. The models improve because they are used. The vendors deepen their position because they become embedded in workflows.

If access becomes unreliable, the incentive changes.

Customers begin asking, “What is our fallback?”
Developers ask, “Can we route around this?”
Governments ask, “Can we build our own?”
Organizations ask, “Can we avoid making this vendor a critical dependency?”

This is how open models gain momentum.

A Chinese open-weight model such as Moonshot AI’s Kimi K2.7 does not need to be better than every closed American model to become strategically important. It only needs to be good enough, available enough, controllable enough, and cheap enough to be useful in a diversified architecture.

That is the key. The winner in enterprise AI may not always be the best model. It may be the model that fits the governance, cost, sovereignty, and control requirements of the organization.

For many tasks, the best frontier model is unnecessary. A local model may be sufficient to summarize internal documents, classify maintenance tickets, draft standard operating procedures, translate routine communications, or answer common HR questions. A mid-tier model may be enough for procurement analysis, sales enablement, or customer support. A frontier model may be reserved for complex engineering, legal, strategic, scientific, or software development work.

This is why routing matters.

A company that sends every AI prompt to the most expensive frontier model is likely overpaying, but a company that sends every prompt to the cheapest model is likely underperforming. A company that sends sensitive data to the wrong jurisdiction is creating risk, but a company that uses only one provider is creating dependency.

The model-agnostic harness solves for this by treating models as a portfolio.

The fourth lesson: the platform may move above the model

In the early cloud era, companies chose infrastructure providers. In the SaaS era, they chose application suites. In the AI era, many assume they are choosing models.

That may be wrong.

The more durable platform may be the harness and routing layer above the models.

If the harness owns the user experience, enterprise identity, policy enforcement, workflow integration, document retrieval, prompt management, audit trail, data controls, budget controls, and model routing, then the model becomes replaceable.

This is a profound shift.

Today, many executives experience AI through branded interfaces: ChatGPT, Claude, Gemini, Perplexity, and others. But enterprises do not run on consumer experiences. Enterprises need governance, integration, reliability, compliance, cost discipline, and change management.

That creates room for a new category of AI service provider.

These providers will not simply resell model access. They will design, host, maintain, and govern the AI harness for the enterprise. It may be installed locally or hosted in a private cloud, connected to multiple frontier models and provide EU-hosted routing for European data. They may build industry-specific prompt libraries, workflow agents, compliance controls, and performance dashboards. They may advise executives on where AI should be used, where it should be prohibited, where humans must approve outputs, and where automation creates measurable value.

For example, for large manufacturers, this will be especially important.

Manufacturing companies are not simple knowledge-work environments. They have plants, production lines, quality systems, engineering drawings, supply chains, maintenance logs, safety procedures, ERP systems, MES platforms, industrial control environments, union considerations, customer requirements, export controls, and strict operational discipline. A generic chatbot is not an AI operating model.

A model-agnostic harness in manufacturing must answer practical questions.

Can the AI access work instructions, understand maintenance history, retrieve the latest quality documentation, distinguish between engineering guidance and approved procedure, prevent sensitive customer or defense-related data from leaving the correct environment, route routine tasks to a low-cost models and complex engineering analysis to a frontier model? Can it log decisions, explain which model was used, support regulated workflows, integrate with Microsoft 365, SAP, ServiceNow, PLM, CAD, and plant-level systems? Can it work across languages and regions? Can it keep operating if one vendor is restricted?

For a large global manufacturer, building this internally is possible but expensive. It requires architecture, security, data engineering, AI governance, vendor management, legal review, workflow redesign, and change leadership. Most companies will not want to do all of this alone.

The old systems integrator implemented ERP, CRM, cloud, and workplace platforms. The new AI Systems Integrator will implement the operating layer through which employees, agents, data, and models interact.

The fifth lesson: Europe will have its own AI sovereignty moment

I have seen this movie before.

Years ago, inside the Microsoft ecosystem, European and German cloud sovereignty was not an abstract policy discussion. It was a boardroom, customer, and regulatory issue. Customers wanted the functionality of American cloud platforms, but they also wanted assurances about jurisdiction, data access, control, trust, and regulatory alignment.

That tension helped create sovereign cloud models, German cloud discussions, EU data boundary commitments, and a broader recognition that cloud architecture is never just technical. It is legal, political, and strategic.

AI will have the same moment.

Europe has already been moving toward digital sovereignty. The drivers are familiar: GDPR, data residency, the Cloud Act, platform dependency, industrial competitiveness, cybersecurity, and concern about the dominance of American technology firms. The recent push for EU-native cloud, SaaS, and AI infrastructure is part of a broader attempt to ensure that European companies and governments are not permanently dependent on systems they do not control.

The Fable 5 incident adds fuel to that argument.

For Europe, the issue is not simply whether Anthropic, OpenAI, Google, or Microsoft are good companies. Many European organizations trust and depend on them. The issue is whether European institutions and enterprises can rely on U.S. AI infrastructure if access to the most advanced capabilities can be shaped by U.S. national security policy, tariffs, export controls, trade disputes, or domestic political changes.

Over the past 18 months, U.S. behavior around tariffs, export controls, industrial policy, and geopolitical bargaining has reinforced a difficult lesson for allies: partnership with the United States remains essential, but dependency on the United States carries risk.

That is the game-theory signal.

If one player becomes less predictable, the rational response of other players is to reduce dependency. Not necessarily to defect, not to abandon cooperation. But to build options.

For Europe, this means more investment in EU-native SaaS, sovereign cloud, European AI infrastructure, local model hosting, regulated AI providers, and ethical AI services aligned with European law and values.

This creates a strategic opening for companies like Anthropic as well. If U.S. AI companies want to remain trusted in Europe, they may need more than sales offices and data centers. They may need European headquarters, European governance structures, European research & development, European model-hosting options, and credible legal separation for regulated markets.

In other words, American AI companies may need to become more, European, in Europe.

This is not only a defensive move. It could be a commercial advantage. An AI provider that can offer frontier capability, strong safety practices, EU-hosted infrastructure, transparent governance, and alignment with European regulation may become the preferred partner for regulated industries.

The winners will not be the companies who argue that sovereignty concerns are irrational. The winners will be the companies that design for them.

The sixth lesson: closed models are powerful but politically fragile

Closed frontier models offer major advantages.

They are often the most capable. They are easier for enterprises to consume. They come with professional support, safety systems, enterprise contracts, uptime commitments, security reviews, and vendor accountability. For most organizations, using a hosted closed model through a trusted provider will be the fastest path to value.

There is nothing inherently wrong with this. In fact, for many companies, it will be the right starting point.

A hosted solution with Anthropic and Claude, OpenAI and ChatGPT Enterprise, Google and Gemini, or Microsoft and Copilot can reduce complexity. It allows the organization to lean on the provider’s safeguards, product development, security posture, and model improvements. For companies early in their AI journey, this may be far more realistic than attempting to build and operate their own harness.

But the board must understand the trade-off.

A closed model is not only a technology choice. It is a dependency choice.

You depend on the provider’s pricing, their roadmap, terms of service, safety decisions, the provider’s jurisdiction, provider’s regulatory exposure, willingness and ability to continue serving your markets and much more.

That may be acceptable. But it should be explicit.

The wrong approach is accidental dependency. The right approach is conscious dependency.

Executives should ask: where are we comfortable depending on a single provider, and where do we need optionality? Which workflows are low risk? Which are strategic? Which involve regulated data? Which involve export-controlled information? Which involve customers in multiple jurisdictions? Which would create operational disruption if the model became unavailable?

Not every use case needs a model-agnostic architecture. But every AI strategy needs to know where model dependency creates business risk.

The seventh lesson: the U.S. needs a better strategy

This section was a question posed to me by someone close to this U.S. administration.

The most advantageous U.S. strategy is not unrestricted openness and not a broad panic ban. It is a segmented access regimen.

The United States should aim to keep American models as the default global operating layer while controlling the highest risk use cases.

That means keeping global access to most capabilities. The U.S. should not train the world to route around American platforms. If allied companies, global enterprises, universities, developers, and legitimate users conclude that U.S. model access is unreliable, they will build alternatives. Once those alternatives mature, the U.S. may not easily regain platform dominance.

It also means restricting specific dangerous capability bands rather than whole models. These capabilities could include, autonomous cyber exploitation, bio-design workflows, weapons optimization, advanced vulnerability chaining, and high-scale agentic tool execution. This may require special controls, such as an LLM government regulatory control framework, where automated repetition of the above mentioned bands are reduced in the LLMs capabilities to a “kitchen knife” as opposed to an “autonomous drone swarm”.

For the LLMs specifically designed for the above mentioned bands, registration and regulation may be the safer option. But broad restrictions on entire models can be blunt instruments. They may block harmless and productive uses while pushing customers toward less governed alternatives.

A better system would use identity, telemetry, rate limits, audit trails, and behavior based monitoring rather than simple nationality based bans. A nationality based ban is easy to understand but strategically crude. It can penalize legitimate users, foreign national employees, allied researchers, multinational teams, and enterprises with complex operating footprints. Behavior based controls are more difficult to implement, but they preserve more legitimate use while targeting the actual risk.

The U.S. should also create an allied AI access zone. Trusted users in allied markets should have reliable access to U.S. frontier models under clear rules. Cutting off allies does not simply reduce risk. It hands market share, legitimacy, and urgency to Chinese and European alternatives, endangering the U.S. default position.

The U.S. should also subsidize domestic open weight alternatives. It should not rely only on closed labs. If developers and enterprises are moving toward open harnesses and routing platforms, the U.S. should make sure the best open models, evaluation systems, inference stacks, safety tools, and orchestration frameworks are American or allied.

This is a fundamental gap in U.S. frontier technology policy. AI leadership will not be secured by closed labs alone, just as it will not be secured by chips alone, if the country lacks the electrical capacity to power data centers. Current open-model leaderboards show Chinese labs holding many of the strongest positions among open weight models, with American alternatives less prominent than one would expect from the country leading the broader AI race. If that pattern persists, the U.S. risks conceding influence over one of the most important control points of the next AI stack: the models which developers, enterprises, and governments can run, adapt, route, and trust on their own terms.

Finally, the U.S. should make its models the safest and most reliable enterprise default. The strategic asset is not only model quality. It is trust, uptime, legal predictability, API stability, governance, and integration depth.

The U.S. wins if the world says, “I can safely build on American AI infrastructure because it is powerful, reliable, governed, and not arbitrarily revoked.”

The U.S. loses if the world says, “Never build critical infrastructure on an American frontier model because access can be politically switched off.”

That is the strategic difference.

The eighth lesson: cost control will come through model routing

There is another reason the model-agnostic harness will matter to executives: cost.

Most organizations are still in the experimentation phase of AI. A few teams use ChatGPT, Claude, Gemini, Copilot, or other tools. A few employees become power users. The monthly cost looks manageable because usage is still fragmented.

That will change. And in many organizations, is consuming budget by the second quarter.

As AI moves from experimentation to operations, usage will multiply. Employees will not only ask occasional questions. They will summarize meetings, draft proposals, review contracts, analyze spreadsheets, write code, create training materials, search internal knowledge bases, generate reports, inspect data, translate documents, and operate workflow agents. Its only growing.

At small scale, the cost of using a frontier model may seem trivial. At enterprise scale, it becomes a budget line. As seen in recent earnings calls.

This is where many companies will make a predictable mistake. They will route too much work to the most capable and most expensive model because it is easier. The result will be the AI equivalent of flying every employee by private jet, because the destination matters.

Not every task needs a frontier model.

A request to summarize a routine meeting does not need the same model as a request to analyze a complex acquisition. A frontline worker asking for the latest approved cleaning procedure does not need the same model as an engineer diagnosing a recurring failure across multiple plants. A marketing draft, an HR policy summary, a maintenance ticket classification, and a board-level risk analysis are not the same class of work.

A model-agnostic harness allows the organization to route work based on value, complexity, risk, and cost.

Low-risk, repetitive, high-volume tasks can be routed to smaller or local models. Mid-level analytical tasks can go to lower-cost commercial models. Sensitive tasks can stay within a private or regional environment. Complex, ambiguous, high-value work can be escalated to a frontier model. The company can reserve the most expensive intelligence for the moments where it actually changes the outcome.

This turns AI from an uncontrolled consumption model into a managed operating model.

The analogy is familiar. Companies do not run every computing workload on the most expensive server. They do not store every file in the most expensive storage tier. They do not assign senior partners to every-day routine tasks. They classify work and allocate resources accordingly.

AI will need the same discipline.

For large organizations, especially manufacturers with thousands of employees and many recurring operational workflows, routing will become one of the most important levers of AI economics. A plant-level question about a standard operating procedure may be answered by a local model connected to approved documents. A supplier-risk question may go to a commercial model with access to procurement data. A strategic scenario about reshoring production, tariff exposure, and customer demand may go to a frontier model. The value of the answer should determine the cost of the intelligence used to produce it.

This will also change how executives evaluate AI vendors.

The question will not be, “What is the price per user?” It will be, “Can we control the cost per workflow?” Can we see which models are being used? Can we set policies by function, geography, data sensitivity, and task type? Can we cap spend? Can we route lower-value work to lower-cost models? Can we compare answer quality, against cost? Can we prove that a frontier model is worth the premium for a specific use case?

Without routing, AI spend may become another form of SaaS sprawl. With routing, it becomes a managed portfolio.

This is why the harness matters. It is not only a sovereignty layer or a risk-management layer. It is also a financial control layer.

Boards should understand the implication. The AI budget will not be controlled only through procurement negotiations. It will be controlled through architecture. The companies that build routing discipline early will have a structural cost advantage over those that simply push every prompt through the same expensive model.

In the next phase of enterprise AI, the question will not be whether the organization uses powerful models. It will be whether it uses powerful models only when power is required.

What this means for the boardroom

For boards and C-suite teams, the practical implication is simple: AI architecture is now enterprise risk management.

The AI strategy should answer six questions.

First, what is our current model dependency? Which providers are employees already using? Which are embedded in workflows? Which are connected to company data? Which are being used without governance?

Second, which AI use cases require frontier capability and which do not? Many organizations are likely to overuse expensive models because they have not classified work by complexity, sensitivity, and business value.

Third, where does data need to stay? A global company may need different routing decisions for the U.S., EU, UK, ME, and APAC operations.

Fourth, what is our fallback plan? If a provider becomes unavailable, restricted, degraded, or too expensive, can the business continue operating?

Fifth, who owns the harness? Is the company relying entirely on vendor-provided interfaces, or does it have its own control layer for identity, policy, logging, routing, evaluation, and cost management?

Sixth, who is the right partner? AI implementation is not only a software deployment. It is operating-model transformation. It requires governance, process redesign, leadership alignment, employee adoption, risk management, and measurable value creation.

This is especially important in manufacturing.

Manufacturers are under pressure to improve productivity, reduce downtime, retain knowledge, accelerate engineering, improve quality, support frontline workers, and manage complex supply chains. AI can help with all of this. But the wrong architecture can create new risk.

A plant manager does not need a philosophical debate about frontier models. They need an AI system that helps technicians diagnose problems safely, retrieve approved procedures, summarize shift handovers, identify recurring defects, and escalate uncertainty to humans.

A quality leader does not need a chatbot experiment. They need traceability, auditability, approved sources, and clear boundaries between suggestions and decisions.

A CIO does not need another uncontrolled SaaS sprawl. They need architecture, vendor discipline, identity integration, data controls, and cost governance.

A CEO does not need an AI theater project. They need productivity, resilience, and strategic flexibility.

A board does not need hype. It needs assurance that the company is not locking itself into a fragile AI dependency at the same moment AI is becoming central to operations.

The model-agnostic future

The likely future is not one model to rule them all.

The likely future is a portfolio.

Companies will use frontier closed models for the most difficult work. They will use cheaper commercial models for routine work, then use local or open-weight models for sensitive, repetitive, or cost-sensitive tasks. They will use specialized models for engineering, legal, customer service, procurement, or manufacturing workflows. They will route by cost, risk, quality, jurisdiction, latency, and business criticality.

The harness will become the control plane.

In large enterprises, this may be built internally. In mid-market companies, it may be provided by a managed AI service provider. In regulated industries, it may be delivered through sovereign infrastructure. In Europe, it may be designed around EU-hosted models and European regulatory expectations. In manufacturing, it may be wrapped around operational systems and frontline workflows.

This does not mean every company should rush to build its own harness tomorrow. That would be the wrong lesson. The right lesson is that every company needs an intentional AI architecture.

For some, the right answer will be a trusted hosted platform from a major provider. For others, it will be a hybrid model. For the largest and most complex organizations, it will be a model-agnostic harness with routing, governance, and local deployment options.

The strategic question is not whether to use Anthropic, OpenAI, Google, Microsoft, Mistral, Moonshot, or another provider.

The strategic question is how to avoid making today’s provider choice, tomorrow’s strategic constraint.

Existing offers already exist, so start the framework discussion

The good news for executives is that the model-agnostic future is not a theoretical architecture waiting for the market to invent it. Enterprise-grade offerings already exist across the AI stack, from model-routing layers and agent orchestration frameworks to managed platforms from major cloud providers, specialist AI infrastructure firms, and systems integrators.

The boardroom question is therefore not whether this capability can be built, but how deliberately the organization should approach it. Before committing to one model, one vendor, or one AI workplace ecosystem, leadership should establish the decision framework: which use cases require frontier intelligence, which can be handled by lower-cost or local models, where data must reside, what risks are acceptable, and which partners can help operate the environment responsibly.

Getting that discussion started now is more important than chasing the latest model announcement, because the architecture chosen today may define the company’s AI flexibility, cost structure, and geopolitical exposure for years.

The choice in front of leaders

The AI transformation now facing organizations looks deceptively familiar.

It resembles the choice between Microsoft 365 and Google Workspace. It resembles the move to cloud, ERP decision, the cybersecurity platform decision. It resembles every large technology choice that later became an operating model.

But AI is different because it will not sit quietly in the background. It will write, reason, recommend, summarize, search, code, negotiate, design, instruct, and eventually act. It will shape how employees think, how managers decide, how customers interact, how plants operate, and how companies learn.

That means the architecture matters.

A single-provider strategy may be simpler and faster. A model-agnostic strategy may be safer and more flexible. A fully self-operated harness may offer the greatest control but require the most resources. A hosted managed harness may offer a practical middle path. None of these choices is universally correct.

What matters is that the decision is made consciously, not accidentally.

The companies that win with AI will not simply be the ones that pick the smartest model in 2026. They will be the ones that build the capacity to change & govern models, route work, control cost, protect data, and adapt as the geopolitical environment changes. And it will change, with the rise of China.

A closing note for executives

If your organization is beginning to implement AI, this is the moment to step back before the architecture hardens. The right question is not, “Which chatbot should we buy?” The right question is, “What AI operating model will let us create value, control risk, and preserve strategic flexibility?”

That requires more than a technology assessment. It requires leadership alignment, use-case prioritization, governance design, partner selection, operating-model change, and a practical roadmap that employees can actually adopt.

If you are a CEO, board member, CIO, transformation leader, or manufacturing executive trying to make these choices, I help leadership teams clarify the strategy, align the organization, and build a practical path from AI ambition to operational value. To discuss an AI alignment sprint or advisory engagement, send me a message on LinkedIn.