Most Companies Are Waiting for an AI Payoff They Haven't Built

Ask a company how its AI program is going and you will usually get a usage number. Licenses rolled out, employees trained, pilots launched. The number goes up every quarter, and it goes into the board deck.

Ask what changed in the P&L and the room gets quieter.

The professional services firm PwC asked 4,454 CEOs that question late last year. Fifty-six percent said AI had delivered neither higher revenue nor lower costs, and only 12% had seen both. This summer, the National Bureau of Economic Research surveyed about 9,000 executives in seven countries. Seventy-eight percent of their firms use AI.

Asked what it had done for productivity over the past three years, the executives put it at 0.2% a year.

None of this is news anymore. McKinsey & Company, Boston Consulting Group (BCG), MIT and a long list of economists found the same gap between adoption and results. Almost nobody disputes it.

What people disagree about is what the gap means. The most popular answer is also the most comfortable one. Give it time.

The comfortable answer has a Nobel laureate behind it

In 1987, the year he won the Nobel Prize, the economist Robert Solow observed that you could see the computer age everywhere but in the productivity statistics. The payoff did come, about a decade later. In February, Torsten Slok, chief economist at the investment firm Apollo, updated the line for our moment: "AI is everywhere except in the incoming macroeconomic data."

There is serious theory behind the patience. Stanford economist Erik Brynjolfsson and his co-authors call it the productivity J-curve. A general-purpose technology makes measured productivity worse before it makes it better, because companies first have to invest in things no balance sheet records well: new processes, new skills, a different organization.

The dip is the investment. The climb comes after.

Brynjolfsson believes the climb has started. He argued in February that U.S. productivity grew about 2.7% in 2025, nearly twice the average of the past decade, and that we are moving from investing to harvesting. Slok and others say the data doesn't show that yet.

I have no problem with the J-curve. I have a problem with how it's being used.

A lag only helps if you're on the curve

The J-curve doesn't promise that everyone who buys the technology gets the payoff eventually. It says the payoff follows the investment in redesign. Skip the redesign and you are not waiting in the dip. You are standing next to the curve, watching other people climb it.

The survey data fits that reading uncomfortably well. McKinsey's high performers, the roughly 6% that attribute at least 5% of EBIT to AI, were nearly three times as likely as everyone else to have fundamentally redesigned their workflows. In a BCG survey of 152 CEOs this summer, high performers were about seven times as likely to redesign workflows and reshape the business end to end. Only 26% of those CEOs said they had embedded AI in a broader business transformation, and 14% had defined what all their AI initiatives should do for the P&L.

That is correlation, not proof. Well-run companies may simply redesign more and report more value.

But none of the surveys I went through show the reverse: a large group of companies getting real results from AI without changing how they work.

BCG's September release says AI is starting to pay off, with nearly half of companies now generating value. Look at the categories underneath and the picture moves slower. The group BCG calls future-built grew from 5% to 7.5% in a year. The scaling group grew from 35% to 41%. That's movement. It isn't a stampede.

The best hard data comes from Denmark. Economists Anders Humlum and Emilie Vestergaard linked AI chatbot adoption to official earnings and hours records for about 25,000 workers in 11 exposed occupations. Work did change: employers reorganized tasks, and about one in six users at actively adopting workplaces took on entirely new ones. Earnings and hours did not move, and the study rules out effects larger than 2%.

The authors read that as the early dip of the J-curve. It might be. It is also exactly what you would see in a company that adopted the tool and kept the operating model.

Why standing still feels like climbing

If most companies aren't on the curve, why do so many believe they are?

Part of the answer is what gets measured. Adoption is easy to count and results are hard to attribute, so dashboards fill up with the easy number. The easy number always rises.

The other part is that people are bad at judging their own speed. In a randomized trial by the research nonprofit METR, 16 experienced software developers worked on real tasks with and without AI tools. With AI, they took 19% longer. Afterward, they estimated that AI had made them about 20% faster. It was a small study with early-2025 tools, but the size of the misjudgment is the point.

Executives show the same pattern at company scale. In a survey of about 750 U.S. financial executives, the productivity gains firms reported for 2025 were three times what their own reported revenue and headcount changes implied, 1.8% against 0.6%. The researchers attribute part of the difference to revenue that hasn't arrived yet, which is the lag story again.

Put those together and the lag story seals itself. Usage rises, everyone feels faster, the payoff is always one budget cycle away, and nothing in the reporting forces anyone to ask whether the work was ever redesigned.

The way out runs through a second gap

So the answer is to redesign. Fine. Here is the part the transformation decks tend to skip.

In 2026, redesign increasingly means handing work to agents, systems that don't suggest the next step but take it. In BCG's latest survey, 42% of companies expect to give AI agents real decision-making authority by 2030. Only 5% have the full set of controls BCG considers necessary to do that safely, from oversight and rollback to audit and cost limits. The research firm Gartner expects more than 40% of agentic AI projects to be canceled by the end of 2027, and inadequate risk controls are one of the three reasons it gives.

That is the trap. Companies that stay with adoption don't get transformation. Companies that chase transformation through agents walk into a governance problem they haven't built for. They close the value gap by opening a control gap.

The control gap is harder to see. A policy that says every agent needs human oversight passes an audit. It says nothing about what the agent does at two in the morning when a step fails, and it decides to try something else. Governance that lives in a policy document is the adoption number of risk management. Easy to count, and not the thing that matters.

Lag or failure?

For the economy as a whole, the lag story is probably right. General-purpose technologies have always paid off in the end, and AI will too.

For most individual companies, I don't buy it. A lag is something that resolves while you wait. This gap resolves only when someone decides to change how the company works and then takes responsibility for systems that act on their own. Most companies haven't made that decision. Calling it a lag lets them postpone it with a straight face.

So stop reporting adoption as progress. Report what changed in the P&L, and measure it rather than asking people how it feels. Before the first agent gets real authority, make sure someone can answer a simple question. Who can stop it, and how fast?

If you think I'm wrong and this is a lag, tell me what your company is doing while it waits.

I collected the underlying numbers, including the ones that cut against my argument, in a separate fact brief: AI Adoption vs. Business Transformation.