Maria runs a forty-person logistics company outside Nashville. Last spring, she did what felt like the responsible, modern thing: she bought AI seats for the whole team — copilots for the dispatchers, drafting help for the office manager, a forecasting tool for herself. The invoice arrived on time, every month, as invoices do. The savings, though, have been harder to locate. Her dispatchers still work late on Fridays. The office manager still retypes half of what the tool drafts. When Maria looks at her P&L, the AI line is easy to find. The payoff line is not.
Maria is not behind the times. She is the times.
A McKinsey survey of 1,719 professionals globally, covered this week in Fortune’s CFO Daily, found that nearly nine in ten respondents say their organizations now regularly use AI in at least one business function, and 44 percent say they are scaling it enterprise-wide. Adoption, by any historical standard, is blistering. And yet — the productivity gains aren’t translating into operating profit for most companies. Nine in ten are using it. Almost half are scaling it. And the bottom line, for most, hasn’t moved.
This is the AI productivity paradox, and it deserves to be stated plainly: corporate America is spending billions on a technology that has not yet shown up where spending is supposed to show up — in profit.
Even the Federal Reserve is watching for proof. In his Jackson Hole speech, Fed Chair Kevin Warsh pointed to AI token prices — the units measuring the data AI models process — as a potential clue to whether AI is delivering the productivity gains businesses are betting on, according to Fortune’s CFO Daily. Think about what that means: the central bank is looking past the press releases and the pilot programs, hunting for a market price that would reveal whether all this activity is creating real economic value. When the Fed chair is reduced to reading token prices like tea leaves, you know the official statistics haven’t caught up with the story.
Meanwhile, the human side of the paradox is getting stranger. Researchers at MIT tracked students writing essays over four months, dividing them into a ChatGPT group, a search-engine group, and a no-tech group — and the no-tech group showed the strongest brain connectivity while the AI group showed the weakest, as reported in The Hustle’s September 11 edition. In the first session, 83 percent of AI users couldn’t produce a single quote from the essay they had just written. Read that slowly: the tool that was supposed to make us more capable left most of its users unable to remember what they had supposedly created. If that pattern rhymes even faintly with what is happening inside companies — workers assisted into a kind of productive amnesia — it would go a long way toward explaining why the P&L doesn’t reflect the adoption charts.
And then there is the question of who does the entry-level work while all of this sorts itself out. BCG’s James Tucker, who leads the firm’s global corporate finance and strategy practice and talks to hundreds of CFOs a year, told Fortune’s CFO Daily that entry-level hiring hasn’t stopped — but the junior finance job is being rewritten in real time. CFOs, in his telling, aren’t giving up on junior talent; they’re giving up on the old way of developing it. The apprenticeship ladder — the reconciliations, the models built cell by cell, the slow accumulation of judgment — is being dismantled and reassembled around AI workflows, and nobody is quite sure the new version produces the same craftsmen.
So here is the uncomfortable question, asked in the spirit of honest bookkeeping: what would prove the gains are real?
My take is that we should demand the receipts, and the receipts have names. First: operating margins. If AI is making companies more productive, margins should expand — the same revenue with less cost, or more revenue with the same cost. Second: unit labor costs. Productivity, properly measured, means more output per hour worked; if hours aren’t falling or output isn’t rising, the “productivity” is marketing. Third: prices. In a competitive economy, genuine productivity gains eventually reach the customer as lower prices or better products — Maria’s shippers should, in time, pay less for the same delivery. And fourth, the one we talk about least: wages. If workers are truly more productive, their pay should reflect it; a productivity boom that never reaches paychecks is, for most families, indistinguishable from no boom at all.
There is a hopeful version of this story, and it deserves its due. Every general-purpose technology in history — the steam engine, electricity, the computer — arrived with a long, confusing lag between adoption and measurable productivity. Factories kept their belts-and-shafts layout for decades after electrification because nobody had redesigned the factory yet. It is entirely possible that we are living inside that lag right now: the spending is real, the reorganization is incomplete, and the gains are coming but not yet countable. Warsh watching token prices may one day look like an early, clever attempt to see around the corner of the official data.
But “the gains are coming” is a hypothesis, not a result — and hypotheses don’t pay Maria’s invoices. Until operating profit moves, a healthy skepticism isn’t Luddism; it’s stewardship. The CFOs authorizing these budgets, the small business owners signing these contracts, the workers whose jobs are being “rewritten in real time” — they are all owed evidence, not enthusiasm.
It’s not enough to just buy the tools and declare ourselves transformed — we must listen for the evidence in margins and paychecks, learn to measure what actually matters instead of what is easiest to demo, and contribute our skepticism as a form of care: for the budgets we manage, the teams we lead, and the economy we all share.
The paradox will resolve, one way or another. Our job is to make sure we notice which way — and to insist, politely but firmly, on seeing the proof.
