WRITER published a survey last month with two numbers that do not obviously belong on the same page.
87% of leaders say their AI super-users, the top quartile of adopters, are at least 5x more productive than their peers.
Only 29% of the companies employing those people report significant ROI from generative AI.
Both numbers are true enough to matter. Both are directionally supported elsewhere. And sitting them next to each other makes the AI productivity conversation much more interesting than the usual "AI is overhyped" versus "AI changes everything" argument.
The individual productivity effect is real. The organisational value effect is not arriving at the same speed.
That is the super-user paradox.
The individual win is not the weak part of the argument
It is tempting to dismiss vendor research, especially when the vendor is selling enterprise AI software. I would not build a thesis on one survey either.
But WRITER's 5x headline is not the only evidence in the room. Stanford's 2026 AI Index summarises task-level gains across customer support, software development, marketing, and consulting. The gains are real, measurable, and uneven. AI helps structured work inside the model's capability frontier. It can hurt work outside that frontier.
That distinction matters because it explains why the anecdotes feel so convincing. You know someone who writes faster now. You know someone who turns a day of research into an hour. You know someone who has quietly built a stack of agents, prompts, and workflows that makes their old job description look obsolete.
Those people are not imagining the productivity jump.
They are also not enough to change the company.
The organisation is the bottleneck
The harder you hold the ruler to actual P&L, the smaller the number gets.
97% of executives in WRITER's sample say they deployed AI agents in the past year. 29% report significant ROI from generative AI. McKinsey's 2025 State of AI puts the stricter denominator at 6%: companies where AI contributes more than 5% to EBIT and the organisation reports significant value from AI.
That is not an adoption gap. It is a translation gap.
Individual employees can save hours and produce better task-level output while the firm still records little or no financial return. The missing layer is workflow redesign: decision rights, governance, incentives, executive ownership, and operating systems that convert faster individual work into collective performance.
This is where the enterprise AI conversation has been too tool-centred. The question is not whether a company has agents. The question is whether the organisation has changed enough for agentic work to be absorbed.
If decision rights are unclear, AI exposes it. Faster drafts arrive, and nobody knows whose approval matters.
If data governance is weak, AI magnifies it. Retrieval systems surface inconsistent ground truth, and the error rate compounds.
If incentives are misaligned, AI accelerates the misalignment. Super-users get promoted, laggards get threatened, and the company still cannot turn individual output into revenue, margin, or operating performance.
The best counterargument is still useful
There is a strong objection here: maybe we are just early.
Erik Brynjolfsson's productivity J-curve argument says general-purpose technologies depress measured productivity during the investment phase, then deliver gains later. Firms spend on systems, retraining, and reorganisation before the output shows up.
He may be right. US productivity growth in 2025 was stronger than the decade average, and the AI dividend may be starting to appear in the macro data.
But even if the J-curve is turning, it is not turning evenly.
The harvest is concentrated. Brynjolfsson himself points to a small cohort of power users automating end-to-end workstreams. Deloitte's more optimistic 2026 State of AI report says 66% of organisations report productivity and efficiency gains, but only 20% report revenue growth attributable to AI. The distribution is the story.
A harvest concentrated in a narrow cohort of people and firms is not a broad enterprise harvest. For everyone else, it is a competitive disadvantage.
What the 6% did differently
The companies getting meaningful value are not merely buying more software.
They connect AI use to revenue growth, cost efficiency, productivity, or risk reduction. They pick priority use cases instead of letting a thousand tools bloom. They assign executive owners with real authority. They track KPIs against the use case, not against adoption theatre.
McKinsey's 2026 State of Organizations report puts the operating model change plainly: capturing AI value depends as much on people as on technology. One executive in the report uses a rule of thumb I think more boards should internalise: for every dollar spent on AI technology, spend five on people, training, workflow redesign, and organisational change.
That is the sentence most AI budgets are still missing.
The uncomfortable implication
For professionals, being a super-user is still a career advantage. It is one of the few asymmetric opportunities inside large organisations right now.
But the ceiling on individual productivity is the ceiling of the organisation's workflow. Beyond a point, a super-user in a flat organisation is producing output the organisation cannot absorb.
For operators and investors, the same lesson applies at company level. The market is pricing AI adoption as if procurement converts into value. The evidence says it does not. Workflow redesign converts into value. Executive ownership converts into value. A serious people-and-operating budget converts into value.
AI productivity is real.
The ROI is organisational.
That is the gap most companies have to cross now.
Read the full essay on the Learn Hub: The Super-User Paradox.
“AI productivity is real. The ROI is organisational.
