The AI Industry Has Become Obsessed With Models
Over the last several years, the AI industry has become increasingly focused on models. Every major announcement, benchmark, funding round, and product launch seems to revolve around which model is smartest, fastest, cheapest, or most capable. Organizations evaluating AI initiatives often begin by comparing GPT, Claude, Gemini, Llama, and a growing list of open-source alternatives. This focus is understandable. The advances in model capabilities have been extraordinary, and the pace of innovation continues to accelerate.
At the same time, I believe this focus on models is causing many organizations to overlook a more important question. Access to intelligence is rapidly becoming a solved problem. Every year models become more capable, more accessible, and more affordable. The question many enterprises are now facing is no longer whether a model can perform a task. The question is whether that intelligence can be deployed safely, reliably, and effectively within a real operating environment.
This distinction becomes increasingly important as organizations move beyond experimentation. During the early stages of AI adoption, the model is naturally the center of attention because it is the most visible component of the system. However, as organizations attempt to integrate AI into daily operations, they begin encountering challenges that have very little to do with model performance. The conversation gradually shifts from intelligence to execution, from capability to accountability, and from models to operations.
I believe this transition marks the beginning of the next phase of enterprise AI. The first phase was about proving that machines could perform increasingly sophisticated cognitive tasks. The next phase will be about determining how organizations operationalize that intelligence at scale.
The Enterprise AI Reality Gap
One of the most common patterns I see in enterprise AI initiatives is the gap between a successful demonstration and a successful deployment. AI demonstrations are often remarkable. Models summarize complex information, generate reports, write code, answer questions, and complete tasks that previously required highly trained professionals. During a proof of concept, it is easy to conclude that the primary challenge has been solved.
The reality inside an enterprise environment is often different. Once organizations begin deploying AI within real workflows, entirely new requirements emerge. Security teams require governance. Compliance teams require auditability. Business leaders require accountability. Employees require predictable and repeatable outcomes. The organization needs confidence that actions taken by AI systems align with policy, business objectives, and regulatory requirements.
These requirements expose an important truth. Most enterprise AI challenges are not intelligence challenges. They are operational challenges. The model may be fully capable of performing the task, but the surrounding systems required to manage, govern, monitor, and coordinate that work often do not exist.
This reality gap explains why many organizations can demonstrate AI successfully yet struggle to scale it. The problem is rarely that the model is not smart enough. The problem is that intelligence alone does not create an operational system.
The Model Is Only the Engine
One of the reasons I believe the industry has become overly focused on models is that they are the most visible part of the system. When a model demonstrates a new capability, the improvement is immediately obvious. We can see it write better code, answer more complex questions, reason more effectively, or generate higher quality content. It is easy to conclude that the model itself is where most of the value resides.
I am increasingly convinced that this view is incomplete.
A useful way to think about the challenge is through the lens of an aircraft. A modern jet engine is an extraordinary piece of technology. It generates enormous power and enables capabilities that would otherwise be impossible. Yet no one would describe a jet engine as an airplane. The engine is essential, but it is only one component within a much larger system.
An airplane becomes useful because of everything that surrounds the engine. Navigation systems guide it to its destination. Communications systems allow it to interact with air traffic control and other aircraft. Flight controls enable operators to direct its behavior. Instrumentation provides visibility into performance and safety. Operating procedures ensure consistency and predictability. Pilots and crew provide judgment, oversight, and accountability. None of these systems replace the engine, but without them the engine alone cannot fulfill its purpose.
I believe AI models should be viewed in much the same way. Models provide intelligence in the same way an engine provides power. They are a critical component, but they are not the complete system. Intelligence alone does not create a business capability. It does not create governance, accountability, coordination, oversight, or operational consistency. It does not determine who can take an action, what systems can be accessed, how work is reviewed, or how decisions are audited. Those responsibilities belong to the systems that surround the model.
This distinction becomes increasingly important as organizations move beyond experimentation and begin deploying AI into operational environments. During a proof of concept, a model may appear to be the primary determinant of success. Once AI is integrated into real workflows, organizations quickly discover that outcomes depend on much more than model quality. Access to the right information, integration with business systems, workflow design, governance controls, memory, and coordination mechanisms often have as much influence on the final result as the model itself.
What is becoming increasingly clear is that an agent is not simply a model. An agent is the combination of intelligence and the operational systems that allow that intelligence to function effectively. The model provides reasoning, but context, memory, tool access, workflow design, governance, and coordination often determine whether that reasoning produces a useful outcome.
This reality is becoming increasingly visible in both research and production environments. Organizations are finding that improvements in memory, retrieval, tool access, workflow design, and coordination can dramatically improve outcomes without changing the underlying model. In some cases, a well-designed operational framework can deliver greater gains than upgrading to the next generation of model. The same intelligence operating within a better environment frequently produces better results.
That observation has important implications. If organizations can achieve materially different outcomes using the same model, then the source of value is not solely the intelligence itself. The systems that provide context, coordination, and control play an equally important role in determining success.
The implication is simple. Intelligence matters, but intelligence alone is rarely enough. The systems that surround intelligence often determine whether it becomes a useful operational capability or remains an impressive demonstration.
As organizations deploy AI at greater scale, these surrounding systems become increasingly important. Models provide intelligence, but enterprises ultimately succeed or fail based on how that intelligence is connected to organizational context, coordinated across people and systems, and governed through appropriate controls. The challenge is no longer simply building smarter AI. The challenge is building the operational framework that allows AI to work.
Introducing the AI Harness
As organizations move AI from experimentation into production, they consistently encounter the same challenge. The model itself is only one component of the overall system. To operate effectively within an enterprise, AI requires context, memory, identity, permissions, governance, oversight, and coordination. These capabilities sit outside the model, yet they are often what determine whether an AI initiative succeeds or fails.
Different parts of the industry have begun describing this layer in different ways. Some refer to it as an agent runtime, an AI control plane, or an AI operating system. While the terminology varies, they are all attempting to describe the same emerging reality: enterprise AI requires an operational layer that sits between business processes and AI intelligence.
I have increasingly found myself referring to this layer as an AI Harness.
An AI Harness is the operational system that surrounds AI models and agents and enables them to operate safely, reliably, and effectively within real-world environments. While the model provides intelligence, the harness provides the operational framework required to transform that intelligence into outcomes.
In my experience, nearly every enterprise AI challenge ultimately falls into one of three categories: context, coordination, or control.
Context is the information required for intelligence to become useful. AI systems need access to organizational knowledge, historical decisions, operational data, procedures, and institutional memory. Without context, even the most capable model is forced to operate with an incomplete understanding of the environment in which it is expected to perform.
Coordination is the ability to organize work across people, systems, workflows, and increasingly multiple AI agents. As organizations move beyond individual assistants and begin deploying teams of specialized agents, coordination becomes essential. Work must be delegated, handed off, reviewed, and tracked across multiple participants while maintaining continuity and shared understanding.
Control is the mechanism that ensures AI operates within acceptable boundaries. Enterprises require governance, permissions, auditability, oversight, policy enforcement, and human approval workflows. Control provides the accountability necessary for organizations to trust AI with increasingly important responsibilities.
Most of the capabilities commonly associated with enterprise AI—memory, tool access, identity, permissions, workflows, governance, human oversight, and agent orchestration—are ultimately implementations of context, coordination, or control.
Collectively, these capabilities form an operational layer around intelligence. The model provides reasoning. The harness enables execution.
Why Harnesses Become Critical as Organizations Scale
The importance of the harness becomes more apparent as organizations increase their use of AI. A single chatbot can operate with relatively little supporting infrastructure. The risks are limited, the workflows are simple, and the consequences of failure are often manageable.
Enterprise environments are fundamentally different. Large organizations quickly move beyond isolated conversations and begin deploying AI across multiple teams, workflows, and business processes. What begins as a single assistant evolves into dozens or hundreds of operational use cases. AI systems become responsible for coordinating information, performing analysis, executing tasks, and supporting decisions across increasingly complex environments.
As autonomy increases, complexity increases alongside it. As complexity increases, governance requirements increase as well. Organizations need confidence that actions are performed consistently, that approvals occur when necessary, and that decisions remain transparent and auditable. The operational challenge grows much faster than the intelligence challenge.
This creates an interesting shift in priorities. Early AI adoption tends to focus on model selection. Mature AI adoption tends to focus on operational architecture. Over time, the harness becomes the foundation that allows intelligence to scale safely across the organization.
The Rise of Multi-Agent Systems
Another reason the harness is becoming increasingly important is the emergence of multi-agent systems. Most real-world work is performed by teams rather than individuals. Enterprises divide responsibilities among specialists, managers, reviewers, and coordinators because complex outcomes require multiple forms of expertise working together.
AI is following a similar trajectory. Organizations are increasingly deploying specialized agents designed to perform specific functions. Some agents gather information. Others perform analysis. Others review outputs, enforce policy, or coordinate work across multiple systems. The future of enterprise AI is unlikely to consist of a single model performing every task. It is far more likely to consist of teams of specialized agents collaborating to achieve outcomes.
As soon as multiple agents enter the picture, coordination becomes essential. Agents must share context, transfer work, communicate progress, and operate under consistent governance. Without an operational layer, context becomes fragmented, oversight becomes difficult, and accountability becomes unclear.
The harness provides the environment in which these agentic workforces operate. It becomes the mechanism through which intelligence is coordinated rather than merely generated.
Why Cybersecurity Is the Perfect Example
Cybersecurity provides one of the clearest examples of why this operational layer matters. Security organizations operate in environments where accountability, evidence, and governance are not optional. Every action must be explainable. Every investigation must be auditable. Every decision must align with established policies and procedures.
For this reason, cybersecurity teams quickly discover that intelligence alone is insufficient. A security analyst does not simply produce an answer. The analyst follows a process, gathers evidence, documents findings, coordinates with other teams, and operates within established controls. The same expectations apply to AI systems operating in security environments.
The challenge therefore is not simply building smarter AI. The challenge is creating an environment where AI can operate responsibly. Security teams require separation of duties, least-privilege access, approval workflows, evidence collection, audit trails, and policy enforcement. These are not model capabilities. They are operational capabilities.
What cybersecurity organizations are learning today will likely become relevant across every industry that adopts AI at scale.
Models Will Become Commodities
One of the most common assumptions in enterprise AI today is that long-term advantage will come from selecting the right model. While model capabilities remain an important consideration, I believe many organizations are overestimating the durability of that advantage.
History suggests that competition eventually drives capabilities up and costs down. We have seen this pattern repeatedly across infrastructure, software, cloud services, and countless other technology markets. While leaders emerge and innovation cycles vary, capabilities tend to converge over time. AI models are unlikely to be an exception.
The evidence is already visible. Organizations have more high-quality model choices today than they did even a year ago. New models continue to emerge, existing providers continue to improve, and enterprises are increasingly adopting strategies that allow them to leverage multiple models simultaneously. In many cases, organizations are beginning to demand the ability to bring their own models, switch providers, and route different types of work to different intelligence engines based on cost, performance, privacy, or business requirements.
This trend has significant architectural implications. If organizations expect models to change, then they should avoid building operational systems that depend on a specific model. Context, governance, workflows, permissions, memory, and coordination should not need to be redesigned every time a new model becomes available. The operational architecture should remain stable even as the underlying intelligence layer evolves.
This leads to what I believe is one of the most important design principles for enterprise AI:
When the model changes, the operation should not.
Organizations that tightly couple operations to a particular model will find themselves repeatedly rebuilding systems as the market evolves. Organizations that separate operations from intelligence will be able to adopt new models, improve performance, reduce costs, and adapt to changing requirements without disrupting how work gets done.
This is ultimately why the harness becomes so important. The harness provides continuity while the intelligence layer continues to evolve. It allows organizations to benefit from advances in AI without constantly redesigning the operational systems that make that intelligence useful.
The Future Enterprise Stack
Many organizations still think about enterprise AI architecture as a direct relationship between applications and models. Applications send requests to models, models generate responses, and business value is created.
While this architecture works for experimentation, it becomes increasingly difficult to manage as organizations scale. Governance, coordination, memory, permissions, workflows, and oversight all become architectural concerns that cannot be solved within the model itself. As organizations deploy AI across more business processes, teams, and workflows, these operational requirements become increasingly important.
At the same time, organizations are gaining access to a growing number of high-quality models. Many enterprises are already adopting multiple models simultaneously and increasingly want the flexibility to select, replace, or route between models based on performance, cost, privacy, or business requirements. The growing demand for Bring Your Own Model (BYOM) capabilities reflects a broader architectural shift. Enterprises are increasingly recognizing that the intelligence layer will change over time, and they do not want their operations tied to a single model provider. Organizations want the freedom to evolve the intelligence layer without disrupting how work gets done.
For that reason, I believe the future enterprise AI stack will contain an additional layer. Applications and business processes will interact with an operational system responsible for context, coordination, and control. That operational system will maintain organizational knowledge, coordinate work across people and agents, and enforce the governance necessary to ensure accountability and trust. The operational layer will then leverage whichever models are most appropriate for a given task.
In this architecture, the harness becomes the operating layer between business operations and AI intelligence. Organizations gain the ability to evolve models, introduce new agents, and increase autonomy without redesigning the entire system each time the intelligence layer changes.
This separation ultimately becomes one of the most important design principles for enterprise AI. Intelligence will continue to evolve. Models will continue to improve. New providers will emerge and existing providers will compete aggressively for market share. The organizations that succeed will be those that separate operations from intelligence and build architectures where the model can change without forcing the operation to change alongside it.
The Next Phase of AI
Over the last several years, the AI industry has focused primarily on intelligence. Researchers, model providers, and technology companies have pushed the boundaries of what machines can understand, generate, analyze, and create. The pace of progress has been remarkable, and there is little reason to believe that innovation will slow anytime soon.
At the same time, I believe the industry is beginning to encounter a different challenge. Most organizations no longer question whether AI can perform useful work. The question is whether AI can be deployed in a way that is secure, governed, accountable, and scalable. In other words, the conversation is beginning to shift from intelligence to operations.
The organizations that create lasting advantage over the coming years will not simply be the ones with access to the most capable models. They will be the ones that build environments where intelligence can be operationalized consistently across teams, workflows, and business processes. As models continue to evolve, the ability to govern, coordinate, and scale their use will become increasingly important.
This is why I believe the most important developments in enterprise AI may not come from the models themselves. They may come from the operational systems that surround them. Models will continue to provide intelligence. The real challenge is turning that intelligence into outcomes that organizations can trust.
In my view, the AI Harness provides the operational layer that allows models, agents, and humans to work together effectively. As enterprises move beyond experimentation and begin deploying AI as part of their daily operations, this layer will become increasingly important.
The industry spent the last several years proving that AI could think. The next several years will be spent figuring out how AI works.
