The value of AI isn’t a number

The conversation about AI has matured in a specific, welcome way. A year ago the live question was whether the technology was real. The serious version of the question now is whether it pays — the consensus has moved from promise to proof, from celebrating that a model is in production to asking what it returned. This is the right correction. It is also, as corrections tend to be, calibrated to the last mistake rather than the next one.

The fact underneath the shift is by now familiar: most large companies are using generative AI, and most of them report no meaningful effect on the bottom line. Naming that gap was useful. But the remedy on offer — measure the value, then go realize it — quietly assumes that “the value of AI” is a single quantity a company can total up, point at, and grow. That assumption is the problem. Value doesn’t behave like a total. It behaves like a distribution, and the shape of that distribution is the whole story.

The average is where the information dies

Ask for “the value of AI across the enterprise” and you’ll get a number — some blend of hours saved, tickets deflected, and margin moved, summed across every deployment and divided into a tidy quarterly line. The number isn’t false. It’s just built in a way that destroys the only information worth having.

Value is not spread evenly across an AI program. A few uses move something that matters a great deal; the long tail moves things that matter a little. Average the two and you get a figure dominated by the tail — there is far more of it — while the signal from the head, where the consequential value lives, washes out. The aggregate doesn’t summarize the program; it hides it. A company can post a respectable enterprise-wide number while every decision it actually cared about went untouched, and it can post a disappointing one while quietly transforming the three decisions that determine its year.

Diffuse by design

There’s a reason the tail is so heavy. The easiest AI to deploy is the broad, horizontal kind — the assistant everyone gets, the copilot bolted onto tools people already use. It ships fast, adoption climbs, and the benefit is real: a few minutes here, a cleaner draft there, a little less friction multiplied across thousands of people. Real value — and also diffuse, low-consequence, and nearly impossible to attribute. Exactly the kind an aggregate absorbs without a trace.

The harder, function-specific uses — the ones wired into a particular high-stakes workflow — are where the concentrated value sits, and they are the ones that stall. They demand data nobody cleaned, decisions nobody wants to own, and changes to how the work is actually done rather than a tool laid on top of it. So programs drift toward what ships, accumulate a shelf of diffuse wins, and report the sum. The structure of the easy path and the structure of the measurement conspire to point everyone away from the value and toward the volume.

Value concentrates in decisions

Here is the lens we’d offer instead. The value of AI is not distributed across tasks; it concentrates in decisions — and not all decisions, but the small set that are consequential: high-stakes, made under genuine uncertainty, expensive or impossible to reverse. These are the places where being slightly more right is worth a great deal and being confidently wrong is ruinous. A thousand saved minutes do not add up to one of them. The pricing call, the capital allocation, the diagnosis, the go/no-go on a program with a nine-figure tail — the marginal value of improving one of those dwarfs the entire diffuse layer beneath it.

This is why the aggregate framing isn’t merely imprecise; it points in the wrong direction. It rewards spreading AI thin across everything, because that is what makes the company-wide number rise. The concentrated view rewards the opposite: find the few decisions where precision is worth paying for, and aim the capability there. Most of what an organization should do with AI, measured honestly, is decline to deploy it where the value is real but small — so it can be deployed where the value is large and provable.

Measure the decision, not the deployment

If value concentrates in decisions, measurement has to follow it there. The unit is not the deployment and not the user; it is the decision. Did the decisions this capability touches come out better — against the version of the world where it didn’t exist — and were those the decisions that mattered? That is a harder question than “what did we save,” and it is the only one whose answer is worth acting on. It also resists the flattering aggregate, because you can’t average a handful of consequential decisions into a smooth line. You have to take them one at a time and say, of each, whether the judgment improved.

None of this lets a program off the discipline of proof — it tightens it. A diffuse productivity number is easy to produce, easy to believe, and tells you almost nothing about whether your important decisions got better. A decision-level account is harder to assemble and far less comfortable, and it is the whole difference between knowing your AI works and hoping it does.

Realizing value is an act of concentration

“Realize the full value of AI” sounds like a call to do more — more deployments, more coverage, more of the company touched. Read through the distribution, it’s the reverse. The full value is realized by concentration: identifying the decisions where precision pays, aiming the best of the technology at them, instrumenting those decisions so you can tell whether judgment actually improved, and refusing the diffuse expansion that inflates the aggregate while starving the head. Coverage is a vanity metric for AI the way it is for everything else. The companies that will look like they got value from AI in a few years are not the ones that deployed it most widely. They are the ones that deployed it where being right was worth the most — and could prove they were.

One step short

The move from promise to impact was the right move, and it stopped one step short. Impact is still a number, and numbers in aggregate hide exactly what you most need to see. The chain doesn’t end at impact; it ends at a decision — a commitment somebody has to make under uncertainty and then stand behind. That is where AI’s value is finally spent or wasted, and it is the only place worth measuring it. Everything upstream of the decision is just signal. The value was never the signal. It was always whether the decision got better.