Enterprise spending on generative AI runs into the tens of billions. Research from MIT, based on interviews across dozens of organisations and analysis of several hundred initiatives, puts the share seeing no measurable return at close to nineteen in twenty.
That number gets quoted a lot, usually as evidence that the technology is overhyped. It is evidence of something narrower and more useful.
Adoption is not the problem. Almost everybody has adopted. The problem is that adoption and return turn out to be unrelated.
Four in five organisations have piloted a general assistant. Around two in five report some form of deployment. The share reaching production with measurable impact on the profit and loss sits at roughly one in twenty. That is not a slow start. That is a structural gap between using something and getting anything back from it.
The chasm sits between pilot and production
Follow the funnel and the shape is consistent. Most organisations evaluate. A minority pilot. A small fraction deploy at any scale.
General assistants show the highest adoption, because they are flexible and easy to start with. They also fail first against work that matters, and they fail for a specific reason: they carry no memory of the business, no persistence between sessions, and no connection to the process the work actually sits in.
One technology leader in the research described seeing dozens of demonstrations in a year and finding one or two genuinely useful. That ratio is worth sitting with. The demonstrations were not fake. They were untethered.
The shadow economy inside the business
While formal programmes stall, the workforce has already moved. Around nine in ten people use a personal AI tool for work. Fewer than half of businesses have bought an enterprise subscription of any kind.
Your team already knows what useful AI feels like. They found it themselves, in a browser tab, and it made their individual work faster. What it did not do is make the business any better at anything, because everything it learned stayed in that tab and left with them.
The gap is learning, not intelligence
The barrier is not regulation, infrastructure or model quality. Those are the explanations that get offered, and they are the comfortable ones because they are somebody else's to fix.
The actual barrier is that most enterprise deployments do not retain context, do not learn from correction, do not improve with use, and do not adapt to how the work is really done. Every session starts from nothing. Every correction is discarded. The business is running a system that gets no better while the market it operates in keeps moving.
The same research found people preferred human judgement over AI by wide margins for anything spanning weeks or carrying real consequence. That preference is rational. You would also prefer the colleague who remembers last quarter to the one introducing themselves each morning.
Where the return actually shows up
The organisations crossing the divide have one thing in common. They put learning systems into specific processes rather than general assistants into general use.
The reported gains cluster around faster qualification, better retention, and the removal of outsourced cost. That last one matters more than it first appears: the return tends to arrive by replacing an external cost structure rather than by reducing headcount, which is a very different business case and a considerably easier one to defend internally.
Procurement leaders in the research described a window of roughly eighteen months before these decisions harden. That is the part worth acting on.
What this means for a commercial team
The divide is a question of operational design. Moving across it means moving from tools to systems, from prompts to workflows, and from outputs to outcomes.
In commercial terms that means the knowledge of how you win has to live somewhere other than in individual heads and individual chat histories. It has to be written down, kept current from the systems you already run, and reachable by everyone who makes a decision. A pattern that appears across a handful of accounts has to be held long enough to become visible, and judged by somebody senior before the team acts on it.
None of that is a model problem. All of it is an architecture problem, and it is the reason the return is so unevenly distributed.