If the first move in an AI program is cutting heads, the company has already misunderstood the technology.
The fastest way to fail an AI transformation is to treat it like a reduction-in-force. That instinct is understandable. AI arrives, boards see efficiency, finance sees labor cost, and suddenly “transformation” becomes a euphemism for shrinking payroll. But that is lazy management dressed up as strategy.
The real question is not how many people a company can remove. It is how work actually gets done. That sounds softer than it is. In practice, it is the harder, more serious question, because jobs are visible on an org chart, while work lives in the messier layer underneath: tasks, handoffs, exceptions, systems, approvals, rework, judgment, and tacit knowledge.
That is why the current debate is so distorted. The World Economic Forum found that 41% of employers expect workforce reductions where AI can automate certain tasks. But the same research shows the dominant response is not cutting; it is capability building: 77% plan to reskill and upskill existing workers to work alongside AI, 69% plan to recruit people who can design and improve AI tools, and 47% expect to move employees from AI-disrupted roles into other positions. Even among employers who see disruption coming, the center of gravity is redesigning work, not simply deleting it. (weforum.org)
That makes sense once you look at the evidence at the task level. Anthropic’s Economic Index shows AI use is concentrated in specific kinds of work, especially software development and technical writing. Roughly 36% of occupations show AI use in at least a quarter of their tasks, but only around 4% show AI use across three-quarters of tasks. And the pattern still leans toward augmentation, not replacement: 57% of observed usage is collaborative, versus 43% that looks more like direct automation. The ILO’s latest policy brief pushes the same point further: the potential for augmentation is about six times greater than the potential for full automation. (anthropic.com)
This is the part many executives still miss. “Paralegal,” “claims handler,” or “customer support rep” is not the right unit of analysis. A role is a container. Inside it are ten or fifty different activities, and AI will hit them unevenly. One task disappears, another speeds up, a third becomes more important because humans now spend more time on exceptions, escalation, and judgment. If leadership starts with layoffs, it freezes the role before it understands how the role is changing.
The economic signal is already pretty clear. PwC’s 2025 AI Jobs Barometer found that since 2022, productivity growth in industries most exposed to AI has nearly quadrupled, while workers with AI skills command a 56% wage premium. Wages are rising about twice as fast in industries most exposed to AI, and PwC reports that job numbers and wages are still growing across virtually every AI-exposed occupation, including the more automatable ones. That should end the fantasy that the only value story is labor subtraction. The bigger prize is work redesign that makes people more productive, more valuable, and more scalable. (pwc.com)
The catch is that work is changing faster than most management systems can see. PwC found that the skills employers seek are changing 66% faster in occupations most exposed to AI than in the least exposed ones. The World Economic Forum says nearly 40% of the skills required on the job are set to change by 2030, with skill gaps already the top barrier to business transformation. In other words: the problem is not mainly that companies will have too many people. It is that they will have the wrong work architecture and the wrong skill mix if they keep managing by static job titles. (pwc.com)
This is where workforce digital twins stop being a nice idea and become a management necessity. At OWI Labs, we think the first deliverable in an AI transformation should not be a cost-out target. It should be a computational model of work: who does what, in which sequence, with which tools, under what constraints, at what error rates, and with what dependency on human judgment. That model lets leaders test change before they impose it.
A serious work model quickly exposes the real options:
- Automate the repetitive subtask, not the whole role.
- Augment the judgment-heavy step, instead of pretending judgment is automatable.
- Re-bundle fragmented activities into higher-value hybrid roles.
- Move people into adjacent work where institutional knowledge still matters.
That is a very different playbook from announcing headcount cuts and hoping productivity shows up later.
It is also a better way to manage risk. When companies skip the work-understanding phase, they usually automate the visible part and ignore the invisible load: exception handling, quality control, compliance exposure, coordination debt, customer escalation, and managerial overhead. The result is familiar. Output looks faster on paper, while the system becomes more brittle underneath. Then the company blames AI, when the real failure was managerial abstraction.
The organizations that win this decade will not be the ones that use AI to remove labor fastest. They will be the ones that understand work deeply enough to recombine labor, software, and machine intelligence into something better. That requires more discipline than a layoff memo. It also creates more value.
AI transformation should not start with layoffs because layoffs are what companies do when they do not understand the work well enough to redesign it.