The Productivity J-Curve: Why AI Gets Worse Before It Gets Better
Nearly every company that deploys AI hits a dip before it sees a payoff. The ones that win aren't the ones with the best models — they're the ones who plan for the dip instead of quitting in it.
In 2025, MIT's Project NANDA published one of the most-cited numbers in enterprise AI: roughly 95% of generative-AI pilots deliver no measurable impact on the bottom line, while only about 5% reach real financial value.1 The instinct is to read that as a verdict on the technology. It isn't. The researchers were explicit that the gap is determined by approach — not by model quality or regulation. The tools work. Most organizations simply abandon them before they pay off. That pattern has a name, and a century of history behind it.
What the J-Curve actually says
In a 2018 NBER paper later published in the American Economic Journal, economists Erik Brynjolfsson, Daniel Rock, and Chad Syverson described what they called the productivity J-curve.2 Their argument: general-purpose technologies — electricity, the computer, now AI — don't create value on their own. They create capacity. Capturing that capacity takes a second, larger wave of investment that is mostly invisible: redesigning workflows, inventing new processes, retraining people, rebuilding how decisions get made. Economists call these "intangible" investments, and they cost real money and time before they produce anything you can measure.
The result is a curve shaped like a J. Early on, a company pays the full cost of adoption — the learning, the rework, the false starts — while the gains haven't arrived yet, so measured productivity dips. Then, as the organization figures out what actually works, those intangible investments start paying off and productivity climbs past where it began. Adjusting for intangibles in computing alone, the authors estimated U.S. productivity was nearly 16% higher than the official statistics suggested. The gains were always coming. They were just lagged.
We have seen this movie before
The J-curve isn't an AI phenomenon — it's how transformative technologies always behave. Economic historian Paul David's study of factory electrification found it took roughly 40 years for electric power to show up in productivity statistics, not because the technology was weak but because factories had to be physically and organizationally rebuilt around it first.3 Robert Solow captured the same paradox for computers in 1987: "You can see the computer age everywhere but in the productivity statistics."4 In every case the lag was real, the eventual payoff was large, and the bottleneck was organizational, not technical.
Moving along the J-curve is not something that happens to a company. It is the company's job to decide where on the curve it wants to be, how fast to climb, and how to bank the learning so the next deployment starts higher.
The dip is where the 95% give up
Here is the uncomfortable overlap. The MIT failure rate and the J-curve are describing the same moment from two angles. The trough of the curve — the period of trial, error, and rework — is exactly where most AI initiatives are quietly shelved. Budgets get cut, champions lose patience, and a tool that was three months from compounding gets labeled a failure. MIT's researchers found the difference between the 5% and the 95% wasn't the model; it was whether tools were integrated deeply into real workflows, whether they adapted over time, and whether the people closest to the work — not just a central AI lab — were empowered to drive adoption.1 In other words: whether someone was actively navigating the curve.
Three things decide which side of the divide you land on
Technology creates capacity; organizations capture value. Buying the tool is the easy part. The majority of what decides the outcome is people and process — the split BCG describes in its widely-cited 10-20-70 rule, where 70% of AI value comes from people and process, not the algorithm.5
Adoption friction is systematic, not noise. The dip is predictable and, more importantly, manageable. It can be planned for, measured, and shortened.
The capability can be built. Experimentation turns uncertainty into advantage — but only if someone is running the experiments deliberately and accumulating what works.
How Fractional Agent helps you navigate it
This is the entire reason Fractional Agent exists. We're operators who became AI experts, not the other way around, and we've seen enough transformations to know the technology is rarely the thing that fails. The dip is.
Our model is built around a single role: the Fractional AI Agent Manager. Think of it the way you'd think of a fractional CFO — an experienced operator who owns a defined slice of your AI operations, deep enough to create impact, lean enough to be affordable. Where most engagements end at deployment, right at the top of the trough, ours begin there. The Agent Manager does the unglamorous work that actually moves a company up the curve: prompting and context engineering, quality assurance, output validation, and the steady iteration that turns a promising pilot into a dependable part of operations.
A few principles guide how we do it:
We start with visibility, not cost-cutting. The first engagement unifies the operational data you already have into one clear picture, so decisions about pricing, capacity, and customer mix sharpen immediately — value that shows up even while deeper automation is still maturing.
We plan for the dip out loud. Adoption friction goes into the plan as a managed phase with a timeline, not a surprise that derails the project halfway through.
We tie the plan to your outcomes. Every implementation is designed around a clear, expected EBITDA impact. If the financial case isn't there, we'll tell you not to do it.
We build your capability, not your dependence. The goal is an organization that gets stronger at AI over time, not one that quietly hands its judgment to a black box.
The J-curve is not a warning to stay away from AI. It's a map. It tells you the dip is coming, that it's temporary, and that the companies who reach the climb are simply the ones who didn't try to navigate it alone.
If you're somewhere in the dip — or trying to avoid landing there — that's the conversation we have every day.
1 MIT Project NANDA (2025). The GenAI Divide: State of AI in Business 2025. Found roughly 95% of enterprise generative-AI pilots delivered no measurable P&L impact, with the gap driven by approach — depth of workflow integration and adoption ownership — rather than model quality.
2 Brynjolfsson, E., Rock, D., & Syverson, C. (2021). "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies." American Economic Journal: Macroeconomics, 13(1), 333–372. (NBER Working Paper 25148, 2018.)
3 David, P. A. (1990). "The Dynamo and the Computer: An Historical Perspective on the Modern Productivity Paradox." American Economic Review, 80(2), 355–361.
4 Solow, R. (1987). "We'd Better Watch Out." New York Times Book Review, July 12, 1987.
5 Boston Consulting Group. The 10-20-70 rule for AI transformation (10% algorithms, 20% technology and data, 70% people and process).
Want to navigate the dip instead of quitting in it? Let's talk.