Individual productivity is up. Firm-level results are flat. Both things are true at the same time, and the gap between them is where most enterprise AI spend goes to die.
Terence Tao, arguably the best living mathematician, described his own AI usage on the Dwarkesh Podcast. AI helps him plenty with the secondary work, the literature searches and reformatting and plotting, but the help stops short of the core, the actual solving of the problem. What shows up instead is richer papers: more code, more pictures, plots that took hours now taking minutes. The tell is what he did before AI: he wouldn’t have put the plot in the paper at all, he’d have described it in words, and that would have been fine. When Dwarkesh pushed him on whether he’s 2x more productive, Tao answered that productivity isn’t one-dimensional: today’s papers would take five times longer without AI, but he wouldn’t have written them that way in the first place.
The 5x is real, but it’s 5x on scope Tao chose to add, and the richer papers may even be better papers, easier to read, easier to build on. The value is the untraceable kind: nobody can tie a plot to a theorem, a citation, or a grant. That’s what the enterprise is buying at scale. AI raised the output from a B to an A+, and nobody can say what the A+ is worth.
The evidence
The Bank of Korea published research on exactly this (BOK Issue Note 2026-12). Workers who adopted generative AI saved an average of 1.5 hours a week, about 4% of working time. The correlation between time saved and output growth: 0.00. The researchers call it the “productivity disconnect.”
The exception is the interesting part. Productivity gains showed up in three groups: the self-employed, professionals, and the heaviest AI users. The BOK’s own read is that these groups have more job autonomy and stronger performance incentives than everyone else: their gains had somewhere to go. Everyone else saved time, and the gains show up nowhere the data can see.
The barriers the BOK identified: weak reward systems, rigid organizational decision-making, and AI that speeds up individual tasks while the workflow around them stays fixed. Keep that last one in mind.
Why A+ work doesn’t matter
Most enterprise work has a quality bar. The memo needs to be clear enough. The report needs to be accurate enough. The analysis needs to support the decision. Call it a B.
Give everyone in that system an AI subscription and they’ll produce A+ work, because doing better work is what skilled people know how to do. The memo gets more polished. The report gets more charts. The deck gets tighter.
None of it flows to the P&L. The decision was going to get made either way and the bar was already met, so the firm captured zero incremental value and paid for every token.
Better work can still convert: an analysis sharp enough to change a decision, or to let the same team cover twice the ground, shows up as revenue per person, and it counts. The waste is the quality above what changes anything, the A+ where a B already did the job.
Between those two sits a wide band of maybe: work that got better in ways nobody can trace to an outcome. Maybe the sharper memo is winning business somewhere. Nobody can show it, and no CFO funds maybe with conviction.
This isn’t hypothetical. Enterprise AI is priced on consumption, that’s how OpenAI and Anthropic sell it, and the firm pays for every token the org burns. The line item is visible, it’s growing, and in fixed-role functions nothing on the revenue side moves in response. The CFO asking “should we cap token spend?” is asking a reasonable question, and “look how much more productive everyone feels” is not an answer.
The part AI doesn’t touch
There’s a second constraint. Even when the productivity gain is real, it’s bounded by how much of the total workflow it touches.
Amdahl’s Law, borrowed from parallel computing: the part of the job you don’t speed up sets the floor. Take a 100-hour job where AI can help with 30 of those hours. Make AI 10x faster and the job takes 73 hours; make it infinitely fast and the job still takes 70. The other 70 hours sit untouched, and no amount of model capability moves them.
And 73 hours is the best case, the one where somebody collects the saved hours and re-plans the job around them. In a fixed role nobody collects them: the job is still scoped at a hundred hours, the saved time stays scattered across individual weeks, and the firm measures zero. Amdahl sets the ceiling. Most firms never get near it.
This has happened before. Paul David traced the same pattern through factory electrification: motors arrived in the 1880s, and the productivity surge waited until the 1920s. Electric motors bolted into steam-era layouts, with their central drive shafts and belt systems, delivered almost nothing. The gains arrived only when factories were rebuilt around the new technology: distributed motors, single-story layouts, workflow redesigned around the machine instead of the machine wedged into the workflow.
Enterprises today are bolting AI onto steam-era org charts.
The conversion test
So when does individual AI productivity convert to firm value? The BOK’s exception groups had two things the rest of the sample didn’t: autonomy and performance incentives. That matches every deployment I’ve watched, and it reduces to two conditions that have to hold at once:
A direct line to the P&L. Commissioned salespeople, P&L owners, founders. When your output has a price on it, faster and better show up as money.
Autonomy to redirect saved time. If AI frees up 20% of your week and you have the standing to point that time at the next most valuable problem, the gain has somewhere to go. If your role is fixed and your queue is defined by someone else, the saved time evaporates into Slack.
One without the other fails in its own way: a commissioned seller locked inside a fixed process has a real gain the workflow absorbs, and an autonomous worker with no revenue line turns saved time into work nobody priced. Solo and near-solo builders have both conditions by default, which is why AI lands hardest for individual builders and small teams, and why the loudest productivity claims come from them.
If you have neither, the shape is familiar: a fixed role inside a mature company, output already at the bar, saved time with nowhere to go. AI makes your work better and your company’s economics worse. That seat is the token burn zone, whoever sits in it, and if it’s your seat, the one move you own is to make the saved time visible: take the next problem before it’s assigned.
What companies should actually do
“Cap token spend” and “token max” both treat AI as a cost line to be tuned. The cap instinct is half right, burn-zone spend deserves one, but a blanket cap starves the corner where spend compounds. The real move is structural.
Measure at the function level, not the individual level. Time-saved surveys measure the wrong thing; the BOK just measured their relationship with output at zero. The question is whether the function’s economics moved: revenue per head for revenue functions, cost per unit of work for cost centers, per hire, per close, per contract reviewed. When the gains are real they surface one of two ways, revenue per employee goes up or headcount goes down. If revenue is growing you get to do this without layoffs; if it isn’t, you don’t. That’s unpopular and true. One caveat if you sell hours: on time-and-materials work, saved hours are lost revenue until you reprice, so the conversion runs through the pricing model before it touches the org chart.
Reorganize around the improvement. Small teams with real autonomy and end-to-end ownership are the electric-motor factory layout. I made this case in How to surf the AI wave: bet on small autonomous teams that own a problem end to end, and fund them with tokens instead of headcount. A small team can redesign its own workflow. A function spanning three departments can’t. Some of that rigidity is load-bearing: regulatory review, segregation of duties, client sign-off. A lot of it is just coordination left over from before the tools changed, and the reorg work is telling those two apart.
Route spend to where it converts. Token budgets should follow the conversion test, not seniority or enthusiasm. A commissioned sales team with heavy AI usage is an investment; a back-office function producing A+ versions of B-bar deliverables is pure cost. Most token-spend dashboards can’t tell the difference. Fix that before you fix the budget. I covered what the spend itself buys, rung by rung, in Tokenomics.
Stop celebrating the wrong use cases. The most famous enterprise AI use case is writing better emails. The enterprise value of a better email rounds to zero, and summarization isn’t far behind: you can’t even measure it. If your AI success stories are polish stories, you don’t have success stories yet.
The bottom line
AI is the rare technology where the individual gains are immediate and the firm gains are conditional. The conditions are structural: output with a price on it, and the autonomy to redirect the time AI hands back. Companies that buy it as a per-employee productivity vitamin will get what the Bank of Korea measured: hours saved, output flat, spend up.
AI works. The paradox is that most organizations aren’t set up to take advantage of it, and getting set up is the harder work: the factory rebuilt around the motor.