The Gap Between Knowing and Doing
What Robotics Is Teaching Us About the Real Challenge of Enterprise AI
According to RAND Corporation’s 2025 analysis, 80% of enterprise AI projects fail to deliver their intended business value, twice the failure rate of conventional technology projects. McKinsey’s 2025 AI survey found that while 88% of organizations now use AI in at least one business function, only 39% report any measurable impact on earnings. In 2025, 42% of companies abandoned most of their AI initiatives, up from just 17% the year before.
Most explanations for these failures point to data quality, organizational readiness, or change management. These are real factors. But after years of working in AI Governance and Security Risk, I’ve come to believe there is a more fundamental problem, one that these explanations don’t fully capture. Organizations are confusing Intelligence with Capability. And that confusion is costing them billions.
A Lesson From Robotics
The distinction became clear to me through an unexpected source: a conversation with Joe Dong, co-founder and CTO of Chestnut Robotics. Joe observed that most people are amazed when they see a robot do a backflip. But in the robotics industry, the harder problem is usually not the backflip, it’s picking up a piece of paper from a table.
When I first heard this, it seemed counterintuitive. A backflip is obviously more complex. But the more I thought about it, the more I realized these are fundamentally different kinds of problems.
A backflip demonstrates what a system can do. Picking up a piece of paper verifies whether a system can actually complete a task. The former is a demonstration of Intelligence. The latter is a test of Capability. The former shows up in demo videos. The latter determines whether a system creates real value.
Joe explained that most people assume a robot’s intelligence is the neural network. But the robot’s intelligence is actually the entire system. Even if the model is perfectly correct, the robot can still fail, because of camera calibration errors, sensor drift, inconsistent hardware behavior, or latency in the control system. The model knows the answer. The system still can’t deliver the result. This is not a problem unique to robotics. It is the central problem of enterprise AI.
The Intelligence-Capability Confusion
For the past several years, the AI industry has competed primarily on one dimension: making models smarter. Every major release from the leading AI labs has focused on stronger reasoning, longer context windows, better benchmark performance.
This competition has produced genuinely remarkable results. But it has also created a dangerous assumption in the enterprise: that Intelligence naturally and automatically translates into Capability. It doesn’t.
Consider what happens when you give the same model to two organizations. One embeds it deeply into workflows and decision processes, redesigns how work gets done around it, and builds the data infrastructure to support it. The other lets employees use it occasionally to draft emails or summarize meetings. Same model. Radically different outcomes.
McKinsey’s research confirms this pattern quantitatively: organizations reporting significant financial returns from AI are twice as likely to have redesigned end-to-end workflows before selecting their modeling approach. The Intelligence was not the differentiator. The integration was.
The most successful AI companies of the past two years illustrate the same principle. Harvey in legal services, Cursor in software development, Perplexity in information retrieval, none of them won by offering a smarter model. They won by embedding AI into a complete workflow. The value users receive doesn’t come from the model itself. It comes from the combination of model, process, data, interface, and decision chain.
Between knowing and doing, between Intelligence and Capability, there are workflows, data pipelines, control systems, feedback loops, and countless organizational constraints. That gap is where most enterprise AI projects go to die.
Why This Confusion Is So Persistent
The Intelligence-Capability confusion persists for a reason: it is easy to see Intelligence and hard to see Capability.
When a model first demonstrates something impressive, summarizing a complex document, generating a detailed analysis, producing code that works, it is natural to conclude that the hard problem is solved. The system knows what to do. Surely value will follow.
But the real world has never worked that way. Knowing how to complete a task and actually completing it are separated by data quality, process integration, exception handling, user adoption, and organizational change. These factors don’t show up in demos. They show up in production.
There is a further complication: organizations lack the data to understand how AI will behave in their specific context. Unlike traditional software, which behaves predictably given defined inputs, AI systems make probabilistic decisions that interact with organizational processes in ways that are difficult to anticipate. There is no ready-made database telling an organization how its AI will perform in its specific workflows, where it will fail, what risks it will trigger, or how it will interact with existing systems. That understanding can only be built through deployment, observation, and iteration.
This is why so many AI projects stall after the pilot stage. The pilot demonstrates Intelligence. Production requires Capability. And the path from one to the other is longer and more demanding than most organizations expect when they approve the initial investment.
A Framework for Closing the Gap
Executives approving AI investments need a different set of questions, ones that probe Capability, not just Intelligence. Before committing to an AI initiative, consider three diagnostic questions:
1. Have we redesigned the workflow, or are we adding AI to the existing one?
AI deployed on top of a broken or inefficient workflow produces faster broken outputs. The organizations that generate measurable returns from AI invest significant time redesigning the work process before selecting the technology. If the answer to this question is “we’re adding AI to what we already do,” the project is likely to underperform.
2. Do we have the data infrastructure to support this system in production?
Gartner estimates that 85% of AI failures are attributable to data quality or data readiness problems. A model is only as good as the data it operates on. If the organization cannot answer clearly where the AI will get its data, how that data will be validated, and how the system will handle data gaps or errors, the project is not ready for production.
3. What is our plan for observability and iteration after launch?
This is the question most organizations skip. Deploying an AI system is not the end of the work, it is the beginning of a continuous process of monitoring, feedback, and improvement. Organizations that treat deployment as the finish line consistently underperform those that treat it as the starting line for learning.
These questions do not evaluate the Intelligence of the model. They evaluate whether the organization has the conditions to translate Intelligence into Capability.
The Next Decade of AI
The robotics industry is teaching us something important about the future of enterprise AI. In robotics, the frontier has shifted. Getting the model right is no longer the primary challenge. Getting the entire system, sensors, hardware, control systems, feedback loops, data pipelines, to work reliably in the real world is where the hard work now lives.
Enterprise AI is approaching the same inflection point. The models are capable enough. The limiting factor is no longer Intelligence. It is the organizational and technical infrastructure required to translate Intelligence into consistent, reliable, measurable Capability.
For executives, this reframe has significant practical implications. It shifts the investment question from “which model should we use?” to “what does our organization need to build to make any model work?” It shifts the success metric from “can the model perform this task?” to “can our system reliably deliver this outcome?” And it shifts the governance question from “is the AI smart enough?” to “is the AI embedded well enough?”
Most enterprise AI failures are not Intelligence failures. They are Capability failures, failures of workflow, data, process, and organizational readiness. Recognizing this distinction is the first step toward building AI programs that actually deliver.
The gap between knowing and doing is not a technology problem. It is a leadership and organizational challenge. And closing it may be the most important AI work of the next decade.
Reference: Innovator Coffee Podcast: Innovator Coffee EP-36 Beyond the model - The real challenge of robotics: