Hi, I'm Charles Costa, researcher, strategist, and founder of Lexora Labs. Each week I dig into AI, knowledge management, and future-of-work research so you don't have to.
EDITOR'S NOTE
Last week we looked at whether the labs can be trusted to grade their own work. This week the question moves inside your own building.
You have read some version of "you can see AI everywhere except in the productivity statistics" a hundred times this year. As of last Thursday it is wrong.
The Bureau of Labor Statistics released second quarter numbers on August 6. Productivity grew 1.4 percent, and across the current business cycle it has run at 2.1 percent annualized against 1.5 percent for the last one. That is not stagnation.
So why does every operator I talk to sound like nothing is landing?
Because nine in ten managers report AI has had no impact on their own company, a fifth of firms capture three quarters of the returns, and the share of output reaching workers just hit the lowest level on record.
The gains are real, invisible to the people producing them, concentrated in a few companies, and not reaching workers. That is a very different problem from the one most people are writing about.
THIS WEEK'S READING:
Productivity and Costs, Second Quarter 2026: U.S. Bureau of Labor Statistics
Nonfarm business labor productivity rose 1.4 percent in the second quarter.
Why It Matters: This release quietly invalidates a lot of commentary you have been reading, including things I have written. The paradox framing assumes the gains have not arrived. They have.
The number almost nobody picked up matters more. Labor share of output, the percentage of what gets produced that reaches workers as compensation, came in at 52.9 percent, the lowest in a series starting in 1947. The productivity is real, and it is going somewhere other than payroll.
If the national accounts can see the gains, the obvious question is who can't.
Firm Data on AI: National Bureau of Economic Research
Nearly 6,000 senior executives surveyed across the US, UK, Germany, and Australia. Sixty-nine percent of firms report using AI, yet more than 90 percent of managers report no impact on employment over three years and 89 percent report none on productivity.
Why It Matters: One correction first. This is not a US study despite the coverage, and "no impact" is the residual bucket on a five point scale, so anyone who did not perceive a change worth sizing lands there.
The real point is that these are perceptions, not measurements. The official statistics say productivity is climbing. The people running these companies say they cannot detect anything. Both are correct, and the gap between them is the story. A companion paper, NBER 34984, finds the reverse, executives perceiving larger gains than were actually measured. Managers are bad instruments in both directions, and most companies are running their entire AI strategy off exactly that instrument.
If perception is unreliable, the results should show up in the spread between firms. They do.
PwC benchmarked 1,217 companies across 25 sectors. Twenty percent of firms capture 74 percent of all AI driven returns, and the most AI fit companies deliver financial performance 7.2 times higher than the rest.
Why It Matters: PwC opens on a scene you will recognize. Someone puts up a slide with a neat grid of AI pilots, the room nods, then somebody asks which of these increased revenue and the room goes quiet.
The finding I keep returning to is that companies with strong foundations see roughly double the improvement from the same increase in AI usage. Same tools, same spend, twice the output. A 2022 study in Technovation found the same pattern in high tech ventures, where adoption at low levels produced no significant revenue growth, and gains were larger at firms that had also invested in complementary systems.
The comfortable explanation is that adoption is still early. It isn't.
Monitoring AI Adoption in the U.S. Economy: Federal Reserve
Jeffrey Allen pulls together three surveys. About 18 percent of US firms had adopted AI as of year end 2025, and about 78 percent of the labor force works at a firm that has.
Why It Matters: It is easy to mislead yourself with the first number, and I did. Eighteen percent sounds like AI has barely arrived, but 95 percent of American firms have fewer than 50 employees, so counting firms mostly counts very small businesses. A separate survey of executives, weighted by employment rather than headcount of firms, puts it at 78 percent.
Four out of five American workers already go to work somewhere that has adopted AI. Access is not the constraint. The tools are in the building.
Fortune argues that task level gains are real but organizational restructuring lags, invoking the parallel to electrification.
Why It Matters: This is the argument I would have made three months ago, and the history still holds.
Factories installed electric motors in the 1880s. The gains did not arrive until the 1920s, because owners swapped out the steam engine and kept the central line shaft. The floor was laid out around the distribution of power, and the gains came only once someone rearranged it around the flow of work.
Here is the 2026 version. Engineers post on Blind about how much faster they are coding, then sit idle waiting on approvals from another team. The motor got faster, the floor plan did not change, and nothing reaches the customer sooner. Speed in engineering that is not matched in legal or procurement does not produce throughput, it produces a bottleneck that moved.
So why do companies keep bolting the tool on instead of rebuilding the process? Not stupidity. They need to meet earnings expectations, and there is enormous fear of missing out. Spending is legible and redesign is not. You can announce a purchase this quarter. You cannot announce a rebuilt approval workflow.
That is the whole thing, and it is not a technology problem. It is a measurement and incentive problem wearing a technology costume.
If you are an executive, change what you ask for before you change what you spend. Most AI reporting counts seats and pilots, which tells you nothing about whether anything got better. Ask which projects shipped on time, and for the ones that did not, where they stalled. That bottleneck list is your redesign backlog.
If you are earlier in your career, the exposure is not your title, it is your scope. What automates cleanly is the bounded task executed to spec, which is where the vanishing apprenticeship comes back around, because judgment is exactly what that narrow work used to teach.
I will not tell you whether the correction is coming, because nobody knows. But this is not a bubble in the dot com sense, which funded companies with no underlying substance. What is being built now will still be useful after expectations reset. Investment will come down. The money was not wasted.
The companies pulling ahead did not buy better tools, they rebuilt the floor around the work. That has been true since the 1880s. The only thing that changed is how fast the gap compounds.
Send this to whoever in your organization is being asked to justify last year's AI spend.
Have a take on any of this? Reply directly, I read every response.