AI Has Made Output Cheaper. It Hasn't Changed What Work Is.
By Jeff Chung - Chief Strategy Officer
06/17/2026
96% of organizations using AI report productivity gains. Less than a third report meaningful organizational ROI. The gap between those two figures is the whole story.
That gap isn't a measurement error, and it isn't a "give it more time" problem. It's the shape of what's actually happening with AI right now, and it has a cleaner explanation than most of the conversation suggests.
Productivity is up. Outcomes aren't. Both numbers are honest
The productivity numbers are real. Drafts get written faster. Reports get assembled faster. Code gets generated, summaries appear, customer tickets close. That 96% is not a rounding artifact, and more than half of those organizations call the gains significant. It tracks with what most teams using AI day-to-day will tell you. The output really is cheaper.
The organizational numbers are also real, and they are harder. McKinsey's State of AI puts the share of companies where AI is showing up meaningfully in EBIT at around 6%.
Two true things, sitting next to each other. Productivity is up almost everywhere. Outcomes are not.
Every shape of AI in use today shares one assumption
It's tempting to read this as a tool problem. Better models, better prompts, better integrations. The next wave will close the gap.
The pattern in the data points the other way.
Consider the four shapes AI typically takes inside an organization today. AI dropped into a workflow, doing a step faster than it used to be done. Agent frameworks automating a sequence of steps that used to take a person walking through them. Copilots and assistants sitting next to a human as they do their job. Chatbots answering a question while the actual work continues somewhere else.
These look very different from the inside. They share one assumption: the work itself is fixed, and AI is the thing that has to fit into it.
That assumption is the ceiling.
Output is downstream of how work is structured
If the work doesn't change, AI can only compress the existing steps. The steps still happen. The handoffs still happen. The decisions still route through the same humans, in the same sequence, with the same boundaries. Individual productivity climbs. Organizational outcomes mostly don't.
McKinsey's own data makes this concrete in a way the headline numbers don't. Across 25 variables they tested for impact on AI-driven financial returns, the largest single effect came from one thing: redesigning the workflow itself. Companies seeing real returns were roughly three times more likely to have rebuilt the process around the AI capability instead of layering AI onto the old one.
The same pattern shows up from another angle. The small group of companies pulling sharply ahead are not running better prompts. They are shifting the shape of how work happens, handing more of it to AI within clear guardrails and letting more decisions run without a human in the middle.
That's not a better tool. That's a different operating model.
The productivity paradox is a description, not a flaw
Researchers have started calling this gap "the productivity paradox." Perceived productivity gains are larger than measured productivity gains.
The cleaner way to say it. When you ask a person, AI saved them time today. When you measure the company, those time savings haven't shown up in any line that matters.
The savings are real. They just aren't connected to anything.
That's not a flaw in the data. It's a description of what AI is currently being asked to do, which is to make existing work faster without changing what existing work is.
A different question is sitting underneath, mostly unasked
Most of the conversation about AI is still asking the same question. How do we get more of it into our current operations?
That question has a ceiling, and the data is starting to make it visible. A growing share of companies are quietly walking away from AI initiatives that never moved the numbers, scrapping pilots that worked in a demo and stalled in production. Those are mostly companies that ran into the ceiling without recognizing it.
A different question is sitting underneath, mostly unasked.
What would the work look like if it weren't designed for humans alone to do it?
Not a faster version. A different version. One where the assumptions about who decides what, who routes what, who owns what, and where the handoffs sit are all on the table. Where AI isn't a tool inside the existing operating model, but a participant the operating model is built around.
That's a harder question to ask. It's also the one the data is pointing at.
Output is one thing. Work is another
The 96% and the much smaller share are not in tension because someone is wrong. They're in tension because they are measuring different things. Output is one thing. Work is another. Most organizations have spent two years optimizing the first one. The second one is mostly where it was when this all started.
That's where the next move is
Chief Strategy Officer, Arkane Digital
Jeff advises organizations on AI transformation, focusing on connecting business strategy to practical implementation without fragmented or reactive adoption.