The slide is confident. It usually is. Somewhere in the middle of a board pack, between the cybersecurity update and the ESG dashboard, sits a single line that twelve months ago was going to change everything. Enterprise AI. Deployed. There is a rollout map on the screen with a great many green ticks on it, one for every region, one for every function, because the license was bought at planetary scale and the model was pushed to every knowledge worker on the payroll in a quarter that somebody put on their objectives. The chief technology officer speaks with the measured calm of a person who has spent a large budget and would like the room to feel that it was spent well. Adoption is up. Everyone has access. We are, the slide says, an AI-enabled organisation now. The board nods. The photograph is taken. And nobody in the room asks the only question that matters, which is what any of it changed on the profit and loss.
Because the honest answer, across most of the global economy in 2025, was nothing. Nothing measurable. Nothing you could take to a chief financial officer and defend with a straight face. The model was superb. The rollout was flawless. The return was a rounding error.
Ninety-Five Percent Is Not a Technology Problem.
The number that should have ended the celebration arrived mid-2025, from MIT. Project NANDA, in its report on the state of AI in business, examined enterprise generative-AI pilots and found that about ninety-five percent of them produced no measurable return on the profit and loss. Not a soft return. Not a return that would show up eventually if everyone was patient. No return. This against an estimated thirty to forty billion dollars of enterprise investment, which is a great deal of money to convert into a slide with green ticks and no consequence. Roughly five percent captured real value. Everyone else had bought the future and filed it under pilot.
The instinctive reading of a ninety-five percent failure rate is that the technology was not ready. That the models hallucinated, that the use cases were premature, that another two versions would fix it. MIT looked at exactly that explanation and rejected it. The failure, they concluded, was not model quality. The frontier models were more than capable of the work being asked of them. The failure was a learning gap: the tools, and more importantly the organisations wrapped around the tools, did not learn from or adapt to the actual workflows of the people meant to use them. The model was dropped into the business like a grand piano delivered to a house with no one who plays. Impressive in the hallway. Silent in practice.
Two further findings in the same body of work tell you where the value actually went, and where it did not. More than ninety percent of firms had employees quietly using personal AI tools to get their work done, a phenomenon the report calls shadow AI, which is a polite name for the fact that the workforce had already found value the enterprise deployment had failed to deliver, and had gone around the official rollout to get it. And fewer than forty percent of organisations had scaled AI beyond the pilot stage at all. So the picture is not one of a hesitant workforce refusing a powerful tool. It is one of a workforce so hungry for the tool that it smuggled its own in, sitting inside organisations that bought the sanctioned version at scale and never built the thing that would have made it matter.
The Gap Was Never Access. It Was Depth.
If the failure were really about access, then giving everyone the same model would have produced roughly the same result for everyone. It did not, and the second piece of evidence is the one that should keep a chief human resources officer awake. OpenAI, in its own study of enterprise AI across roughly nine thousand workers in nearly a hundred companies, measured what happened when the same tools were handed to everyone. The frontier users, the top few percent by depth of adoption, were getting something close to six times the productivity of the median employee using the identical software. Same license. Same interface. Same model. A sixfold gap. The power users reported saving more than ten hours a week, week after week, while the person one desk over saved almost nothing and privately suspected the whole thing was overhyped.
Read that carefully, because it dismantles the entire theory of the rollout. The variable that produced a sixfold difference in return was not the technology, which was held constant. It was not access, which was universal. It was depth of use and learned orchestration: knowing which task to hand the model and which to keep, how to sequence a request, how to chain the tool into a real piece of work rather than treating it as a novelty search box. That is not a feature you can buy. It is a capability you have to build, deliberately, in a specific person, doing a specific job. The gap between the five percent of firms that won and the ninety-five that did not is the same gap, at the level of the institution, as the gap between the power user and the median employee. It is a learning gap. It has always been a learning gap.
And here is where the organisational story becomes uncomfortable, because the enterprise did have a plan for the human side. The plan was a launch email, a lunch-and-learn, a library of generic prompt tips, and a completion metric to prove it happened. This is the same architecture the learning function has used to roll out every tool for twenty years, and it produced the same thing it always produces: awareness without capability. People knew the tool existed. They had been to the webinar. They could not, in the flow of Tuesday's actual work, use it to do anything that changed a number. The organisation had confused distributing access with building capability, which are not adjacent activities. They are different disciplines, and only one of them was funded.
The Five Percent Built an Enablement Layer.
The scale of the misallocation is visible in what the market chose to spend money on. Microsoft, in mid-2025, announced Elevate, a four-billion-dollar AI-skilling initiative with the stated aim of credentialing twenty million people within two years. Set aside the specific vendor. The number is the point, and so is the shape of it. The global enablement spend has become enormous, genuinely board-level, and yet the overwhelming instinct behind it is credential volume: how many people can we certify, how fast, how many badges can we count at the end. Twenty million credentials is a procurement target. It is not, on its own, an architecture. A credential confirms that someone attended. It says nothing about whether a claims adjuster, a credit analyst or a regional sales lead can now do their actual job six times faster, which was the entire prize.
The World Economic Forum has put a floor under how large this problem is about to become. By 2030, on its estimate, roughly fifty-nine percent of the global workforce will need reskilling or upskilling. Nearly six in ten. That is not a training backlog to be cleared with a bigger content library and a longer list of courses. It is a structural demand that the current model of enablement, measured in completions and credentials, is architecturally incapable of meeting, because it was built to distribute information and the task in front of it is to build capability inside the workflow. The two things look similar on a slide. They could not be further apart in a profit and loss.
What the winning five percent did was not mysterious, and it was not a better model. They built the layer everyone else skipped: the enablement layer that turns a general-purpose model into role-specific performance. That layer is systematic rather than optional. It is role-based, because the way a model creates value for an underwriter is nothing like the way it creates value for a software engineer, and a single generic training deck serves neither. It is workflow-embedded, meaning the capability is built at the point of real work rather than in a classroom the learner leaves and forgets. It has a feedback loop, so that what the best users discover gets captured and taught rather than dying as one person's private trick. It is measured against outcomes that reach the profit and loss, not against attendance. This is the unglamorous machinery that converts a license into a result, and it is precisely the machinery that a rollout obsessed with access and credential counts never thinks to build.
This is the layer SSUNDAR builds, and it is the reason we treat an AI deployment as a capability-architecture problem long before we treat it as a technology one. We read the request to roll out a model the way we read any request: as a symptom. The stated need is access. The real need, almost always, is the systematic, role-based, workflow-embedded capability that stands between a model and a return, and that no license, however planetary its scale, has ever contained. We do not add another badge to the twenty million. We build the enablement layer that makes the model finally do the work it was bought to do. It is slower than sending the launch email. It is the only thing on the entire program that the chief financial officer would recognise, twelve months later, as having changed a number.
The model was never the variable. The variable was whether anyone built the layer that teaches a specific person to turn it into performance, and ninety-five percent of the market decided that layer was optional.
You did not buy the wrong model. You bought the model and skipped the only thing that was ever going to make it pay, and then you called the silence a technology problem.