The number that gave the title
An executive survey with 1,719 respondents covering 97 countries (May–June 2026) says the following: close to three quarters of the high-performing companies report that they have fundamentally redesigned their workflows because of AI — last year that share was still 55%. Among the other respondents the same figure is one quarter.
The wording of the survey is precise too: these companies are "fundamentally redesigning workflows enabled by AI, rather than fitting AI into existing ones".
This is the biggest and most consistent difference between successful and unsuccessful rollouts — and the only one that is not a question of money.
What has to be said about it, though
This is cross-sectional self-reporting. We do not know whether the redesign caused the good result, or whether well-performing companies are simply more willing to redesign, or whether a third thing — money, executive attention, company size — causes both.
And the same survey finds differences on other factors as well:
| What the high performers do differently | Them | The rest |
|---|---|---|
| Fundamentally redesigned their workflows | ~75% | ~25% |
| Have a defined process for measuring impact | 52% | 33% |
| Senior leadership takes real accountability for it | 53% | 41% |
| Manage the risk of workforce displacement | 52% | 33% |
| Manage the risk of unintended behaviour | 56% | 40% |
| Manage cybersecurity vulnerability | 47% | 27% |
The gap is widest on workflow redesign, but the survey cannot separate these factors from one another. Anyone who says "this is the only thing that matters" is claiming more than the data supports.
The experimental evidence: the effect goes through the organisation of work
This is where the article's claim becomes defensible, because the three measurements below are not self-reporting but randomised or pre-registered experiments — and all three show the same mechanism.
One: AI does not replace the tool, it makes the team structure unnecessary. In a pre-registered field experiment, 776 professionals worked on real product-development tasks at a large consumer goods manufacturer, in a two-by-two design (with or without AI, alone or in pairs). The result: an individual working with AI produced a solution of the same quality as a two-person team without AI. And the authors specifically highlight that AI "breaks down functional silos": without AI the R&D people proposed technical solutions and the commercial people commercial ones, while with AI both groups proposed balanced ones.
If AI were a better word processor, this result would not come out. The result is about how many people a step needs.
Two: the gain arises through a change in the division of labour. In an experiment with 2,234 participants, human-human and human-AI teams were randomly paired and had them produce 11,024 ads, which were then measured in a real campaign with roughly five million impressions. The human-AI teams produced 50% more ads per head. The authors measured the mechanism as well: 25% more task-oriented and 18% fewer personal messages, 17% more delegation, 62% less direct text editing — and these were what explained the performance difference.
The tool would have been the same in both groups. What differed was the way the work ran.
Three: the gain comes from proven practice spreading. In the customer-support measurement covering 5,179 agents, the AI assistant raised the number of issues resolved per hour by 14% on average, but by +34% for beginners and by roughly zero for the experienced. The authors' explanation: the model "disseminates the best practices of more able workers".
That is not a property of a tool either. It is a channel for knowledge to flow along, one that did not exist before.
The counter-test: redesign on its own is not enough either
If the article stopped here, it would be one-sided. Two pieces of evidence say that rearranging the process does not by itself produce money.
The first is the Danish data. One study linked large, representative adoption surveys to Danish administrative payroll data. The result: two years after language models appeared, no effect larger than 2% can be detected either on earnings or on recorded hours worked — not among the most intensive users, not among the early adopters, not at workplaces that invested substantially, and not among those who themselves reported large gains. And in passing the same study also establishes that adoption goes together with occupational change and a rearrangement of tasks.
In other words: the process does get rearranged, but no money comes out of it. At least not in pay and hours.
The second is software development. According to an industry survey of nearly 5,000 people, 90% of respondents use AI at work and more than 80% of them say their productivity has improved — yet the same survey also shows that AI adoption remains negatively associated with delivery stability. The report's thesis in one sentence: AI does not fix a team, it amplifies what is already there. And without a strong control system the increased volume of change leads to instability.
The two results together say this: redesign is necessary but not sufficient. If you rearrange the process but do not build in the control, AI multiplies the weakness that is already present.
Is there research that separated the two?
Honestly: I did not find any. In no single experiment has anyone measured separately the effect of "the same process with a better tool" and "a different process with the same tool".
What can be claimed is three mutually reinforcing pieces of indirect evidence:
- Cross-sectional: among the companies that achieve a significant result, the share of those who redesigned is three times as large.
- Experimental, from the mechanism: the measured effect consistently works through a change in the organisation of work.
- Negative control: where the organisation and the control system do not change, the tool on its own brings no detectable result, and can even make things worse.
Not one causal number can be written on top of this. But this much is enough for the decision.
What this means in a five-person company
"Workflow redesign" sounds like a large-corporate phrase. In practice it is a single question:
Which step disappears, not which step gets faster?
An example. Producing a quote today is five steps: (1) the customer writes in, (2) somebody digs out the similar quote from last time, (3) rewrites it, (4) a second person reviews it, (5) it goes out.
Swapping the tool: the third step gets faster, because the machine drafts the text. The five steps stay five. The turnaround time is roughly the same, because the bottleneck is the fourth step, where a colleague has to get round to looking at it.
Redesign: the second step disappears, because earlier quotes become searchable and the draft is built from the incoming message. The fourth step, however, stays, and gets stricter — because now it is machine output that somebody is reviewing, and the liability is still the company's. Five steps become four, and the bottleneck has not moved, but the road to it is shorter.
If in the example above you only do the first of the two, the measurement will show nothing — and that is exactly what the Danish data sees at national level.
Three common misconceptions
"Redesign means we are laying somebody off." The measurements do not support this. According to an American business survey from August 2026, 61% of service-sector firms use AI, but only 4% of them have laid anybody off because of it — in manufacturing, not a single firm. Those same firms, on the other hand, retrained existing staff in more than a third of cases. In practice, redesign is retraining, not headcount reduction.
"If we rearrange the process, quality will improve automatically." In the advertising experiment the output of the human-AI teams became more homogeneous — the authors call it a "collapse in diversity". Where being distinguishable is business value (brand voice, a proposal, a negotiation), that is a loss, not a gain.
"First let us find the best tool, then we will plan." It works the other way round. The process map tells you which step you want to remove — and from then on the choice of tool is a technical question, not a strategic one. The method for choosing a model comes in once you already know what for.
What to take away
Swapping the tool is cheap and spectacular, which is why everybody starts there. The measurements, however, consistently show that the effect arrives through the division of labour: how many people a step needs, who gets what delegated to them, whether good practice spreads. If you only swap the tool, your process stays the same number of steps, and nothing will show up in the measurement.
But redesign is no magic either. A process rearranged without controls also produces errors faster. Where the human approval should sit is the subject of the human in the loop. What to measure along the way and what not to is unpacked by measuring the return. And the typical pitfalls, stage by stage of a rollout, are listed by the ten common mistakes.