Heavy AI Adopters Hired 10% More White-Collar Workers. The Rest Are Getting Less Than 10% Back.

In the first two years after heavy AI adoption, white-collar headcount at the highest-intensity adopters grew 10.2%, and entry-level roles grew 12%. A control group of similar companies that did not adopt as heavily showed no meaningful change in either. The figures come from a Ramp and Revelio analysis of roughly 22,000 U.S. companies, surfaced by the Financial Times and written up by ZDNet.

Read the same period through a different lens and the picture inverts. Bain's 2026 Automation and AI Pathfinder Survey found that nearly 40% of companies that bother to measure AI cost savings are landing below 10% returns, even though they targeted 11% to 20%. About 90% of those companies are still raising their AI budgets.

The split has a name. Brian Solis's Forbes column this week calls it the difference between treating AI as a tool, treating it as an automation layer, and treating it as a catalyst. Tool users buy a faster way to do tasks that already exist. Automation-layer users replace a step in a process and leave the process alone. Catalyst users rethink the work, the role, the operating model, and the shape of the business.

The headcount data tracks the catalyst group. When AI is used to restructure how work is done, the work expands. New entry-level roles appear because senior judgment is applied to more problems, not fewer. White-collar headcount grows because the company is willing to staff the new shape of the business.

The return data tracks the other two. Tool users get tool returns, which is marginal time savings on existing tasks. Automation users get automation returns, which is a cost line that drops, a process that runs faster, a margin that improves by a few points. Neither group reaches 11% to 20% because neither group changed the underlying business.

Forty percent of measurers are landing below 10% against a target band of 11% to 20%. The shortfall is not a rounding error. The 90% who keep raising the budget are buying more of the same posture and expecting a different number.

The standard reply is that the companies are early on the curve. They are still figuring out which model to use, which vendor to standardize on, which workflow to retire. Give it another year. The reply is convenient and partially true. It does not survive the time horizon in the Ramp/Revelio data, which already spans two years post-adoption and still shows a clean split between heavy and light adopters. If the curve were the only thing that mattered, the gap would close. It is widening.

Edward Zitron's newsletter makes a sharper critique. AI's economics are concealed by opaque token pricing, he argues, and companies claiming AI savings are often reading a vendor invoice that does not reflect the true cost of running the model, the inference overhead, or the human time spent supervising outputs. If Zitron is right even partly, the 40% landing below 10% are landing below 10% on a numerator that is already understated. Catalyst users are not exempt from this critique, but they have a different exposure: a company that has used AI to rethink the work is less dependent on a line item it cannot fully audit.

The Economist asked in July 2024 "what happened to the AI revolution?" at a moment when enterprise AI pilots were running hot and measurable production returns were thin. Two years later, the headcount data and the return data have answered the question. The deciding factor was the leadership team, not the model.

The test a leadership team can run this quarter is short. List the three to five AI deployments that have shipped to production. For each, write one line: did this change who does the work, what work gets done, or only how fast the same work gets done. The first two are catalyst-shaped. The third is where the 40% are landing.