The Discipline Gap: Why Most AI Investment Isn't Paying Off Yet

A framework for telling the difference between funding activity and funding outcomes, before the board makes you. · Juan Vegarra

After many years sitting on boards and advising executive teams through technology cycles, I've learned to recognize a particular kind of meeting. Everyone agrees the technology matters. Everyone agrees the company has invested seriously in it. And nobody in the room can point to a number that moved because of it. That meeting is happening in boardrooms constantly right now, and it is not a sign that AI doesn't work. It is a sign that most organizations have not yet learned how to tell the difference between deploying a technology and capturing value from it.

The scale of the gap is larger than most executives want to admit publicly. MIT's Project NANDA, published in mid-2025, found that ninety-five percent of organizations that deployed generative AI saw no measurable financial impact, against an estimated thirty to forty billion dollars in collective investment.^[1] A 2026 survey of two hundred U.S. finance chiefs by the professional services firm RGP found that only fourteen percent had seen a clear, measurable return on their AI spend.^[2] Forrester now reports that enterprises are postponing twenty-five percent of planned AI spend into 2027 as financial scrutiny tightens.^[3] The CFO who once waved pilots through with minimal questions has become, in the space of about eighteen months, the most demanding voice in the room.

None of this means the investment thesis is wrong. It means the industry spent two years treating deployment as the finish line, when deployment was always meant to be the starting line. The organizations quietly getting real returns from AI right now are not the ones spending the most. In several cases I've studied closely, they are spending less, because they have learned to distinguish funding activity from funding outcomes, and they kill the former ruthlessly.

Why the confusion persists

Part of the reason this distinction is so hard to hold onto inside a large organization is that activity is easy to measure and outcomes are hard to attribute. A pilot that gets deployed, gets a press mention, gets cited in the annual report as evidence of innovation, generates a satisfying sense of progress regardless of whether it moved a single number on the P&L. Outcome attribution requires the much harder work of isolating what changed because of the technology versus what would have changed anyway, and most organizations don't have the analytical discipline in place to do that work rigorously. So they default to counting activity, because activity is countable and outcomes are not, at least not without real effort.

This is not a new failure mode. It is the same trap that swallowed a decade of "digital transformation" initiatives before AI arrived, and it will swallow whatever comes after AI unless the underlying discipline changes. The technology is not the variable that determines whether an investment pays off. The rigor applied to defining and measuring the outcome is.

The pattern among the minority who are seeing returns

Three habits repeat across the organizations I've seen actually capture value, consistently enough to treat as a framework rather than a coincidence.

First, they fund use cases with a direct line to a P&L number, not a productivity story. Fraud detection, customer service deflection, supply chain optimization, and software development acceleration are surviving budget scrutiny in 2026 because someone in the room can name the dollar figure the initiative is supposed to move, before it starts.^[3] "Employees report saving time" is not that number, and finance chiefs have largely stopped accepting it as one. This is a healthy correction, not an overreaction. A time-savings claim that never converts into headcount discipline, throughput increase, or cost reduction was never really a business case. It was a satisfaction survey.

Second, they treat the pilot as a hypothesis test, not a soft rollout. A pilot that "worked" but has no defined path to production, no owner accountable for scaling it, and no kill criteria if it stalls has not succeeded. It has produced an expensive demo. The organizations who scale successfully define what success looks like in a business metric before the pilot begins, not after it concludes, which sounds obvious and is nonetheless the step most commonly skipped.

Third, they separate infrastructure spend from experiment spend and budget each differently. Data readiness, integration, and governance are capital investments in a capability, and deserve the scrutiny and patience of any multi-year infrastructure bet. Individual use-case pilots are experiments, and should be funded, timeboxed, and killed like experiments when they don't pan out. Organizations that blend the two into a single undifferentiated "AI budget" tend to either starve the infrastructure that would make future use cases succeed, or over-fund pilots that were never going to scale in the first place, because nobody was forced to make the distinction explicit.

The honest counter-case

It would be a mistake to conclude from any of this that slow, cautious AI adoption is automatically the safer path. Some of the sharpest capital destruction I've observed hasn't come from moving too fast on AI. It has come from organizations that used "we're waiting for the ROI case to be clearer" as a permanent posture, watched a faster-moving competitor establish a genuine operational advantage in a use case with real network effects, such as a customer-facing recommendation engine or a fraud model that improves with more transaction volume, and found that the gap had become very expensive to close by the time they moved. Discipline is not the same as delay. The framework above is about the quality of the bet, not the speed of it. An organization can move quickly and still fund outcomes rather than activity, if it does the definitional work up front instead of skipping it in the name of speed.

What this means for the next budget cycle

If you are heading into a planning conversation about AI investment for 2027, the useful question is not "how much should we spend." It is "can we currently tell the difference, initiative by initiative, between the ones funding an outcome and the ones funding activity." Most organizations, if they are honest, cannot yet answer that question with confidence. Building the capability to answer it — clear success metrics defined before the pilot, named business owners, a genuine kill process, and a separation between capital-grade infrastructure spend and experiment-grade pilot spend — is a more valuable investment right now than almost any individual use case competing for the same budget line.

The technology has matured faster than most governance models have caught up to it. That gap is where the money is currently disappearing, and it is also exactly where a disciplined minority is quietly pulling ahead. Which side of that gap is your organization actually on, and how would you know?

Juan Vegarra is the author of An Outsider's Playbook (forthcoming). The views here are his own. More essays · Write me