Why Isn't AI Adoption Showing Up in Your P&L?

September 28, 2026

Why Isn’t AI Adoption Showing Up in Your P&L?

At one bank, a team accelerated its pull request pipeline with AI. Review time rose roughly 441%. Incidents per pull request rose roughly 243%. The team hadn’t gotten slower. It had gotten faster in exactly the wrong place. Speed didn’t disappear when it hit the next stage of the pipeline; it just showed up somewhere else, as a queue.

That anecdote, and the population-scale data behind it—Faros AI’s 2026 research spanning 22,000 developers across more than 4,000 teams—anchors AI-Driven Organizational Value, a paper in the Fall 2026 Enterprise Technology Leadership Journal from IT Revolution, authored by Rodo Abad, Jason Cox, Jeff Gallimore, Betty Junod, Tapabrata Pal, Brian Scott, and Max Reele. Its argument, in short: Freed-up capacity is not the same thing as captured value, and most organizations are measuring the half of that equation that doesn’t matter.

The Bottleneck Didn’t Disappear. It Relocated.

Follow the logic far enough, and it lands somewhere uncomfortable. In the paper’s own words: “The bottleneck was never typing code. Code was simply where the limit was easiest to see and cheapest to attack.” Clear the code-review stage, and the queue lands in validation. Clear that, and it lands in operations—someone has to own everything that ships. Clear that, and it reaches the hardest stations of all: the business’s ability to decide what’s worth building, and the customer’s ability to absorb the change at all.

What limits an enterprise was never the typing speed of its engineers. It’s the whole system’s ability to safely absorb speed. An organization that accelerates one function without widening the rest of the pipe has purchased congestion, not velocity.

The Metric Everyone’s Watching Is the Wrong One

Ask most leadership teams how AI adoption is going, and you’ll get a number: how many people are using it, how many use cases are registered. The paper doesn’t mince words about what that number actually measures: “Counting registered use cases is the AI-era equivalent of counting licensed seats.” It tells you that capability exists but tells you nothing about whether that capability reached anyone who needed it.

The more honest instrument, per the paper, is what the authors call the second-user test: How many AI-built artifacts are in regular use by someone other than the person who built them? That single filter separates a personal script—useful to exactly one person—from real organizational capability. It’s also worth watching the distribution in addition to the count. Building that is concentrated among a handful of enthusiasts is an adoption signal. Building spread across finance, legal, and operations is an optimization signal, a sign the capability is actually compounding.

This matters more than it sounds like it should, because natural-language interfaces have quietly removed a barrier that killed the last version of this idea. Low-code platforms and citizen-development programs made nearly the same promise a decade ago and mostly produced a graveyard of abandoned apps. The difference this time, the paper argues, is that every prior attempt still required the builder to learn some formal system, such as a canvas, a schema, or a constrained language. That barrier moved the problem instead of removing it, and the people with the deepest domain knowledge were consistently the least likely to clear it. Natural language removes the translation step. The specification is the artifact now, which also predicts the new failure mode: Citizen development failed by producing too little to matter. AI, if it fails, will fail by producing far too much, unmanaged.

Value Arrives With a Bill Attached

None of this capacity is free, and the paper is direct about the discipline that has to travel alongside it: “Your AI bill is the sum of decisions someone made. Cost discipline is making them on purpose.” Tokens are this era’s unit of consumption, and a meter is attached to every hour of freed-up capacity. Treating cost as an afterthought—something to reconcile at the end of the quarter rather than a design decision made up front—is how freed capacity quietly turns into an expense line nobody planned for.

What to Do With This on Monday Morning

Take whatever AI adoption metric currently shows up in your leadership dashboard and ask it the second-user question: Of the things being built, how many are actually being used by someone other than their creator? If you don’t know, that’s the gap. The paper’s companion piece in the same issue, “When Coding Is Free, Where Is the Bottleneck?” picks up exactly where this one leaves off—mapping where the widening work actually needs to happen once you’ve found the bottleneck this paper helps you name.

This article draws on “AI-Driven Organizational Value” by Rodo Abad, Jason Cox, Jeff Gallimore, Betty Junod, Tapabrata Pal, Brian Scott, and Max Reele, published in the Enterprise Technology Leadership Journal, Fall 2026 (IT Revolution).

No comments found