Two years into the generative AI cycle, the mood inside many large organizations has changed. The early demonstrations were astonishing. The board-level ambition was expansive. The first wave of pilots was often useful. Yet when leadership teams ask what has changed in the cost base, the answer is frequently smaller than the investment, attention, and internal excitement would suggest.
The problem is rarely the model. It is the design of the work around the model. Most organizations have used AI as a new instrument inside an old process. A person drafts something; AI drafts it faster. A person summarizes a document; AI summarizes it faster. A team assembles a report; AI accelerates parts of the workflow. The point gain is real, but the system remains largely intact.
That approach can create value. It can remove friction, shorten a queue, or reduce the time spent on recurring work. But it does not change the architecture of the process. The same approvals remain. The same handoffs remain. The same review layers remain. The same operating cadence remains. The organization gets a faster version of the old way of working and mistakes that for transformation.
Efficiency gains in core functions are increasingly available when organizations move from AI augmentation to zero-based, AI-native process redesign.
Why the first wave feels smaller than expected
The first wave of enterprise AI work was shaped by understandable caution. Companies started with contained use cases because they were easier to approve, easier to explain, and easier to govern. Legal teams could review them. Technology teams could sandbox them. Functional leaders could show progress without changing the power structure of the organization.
That was the right place to start. It is not the right place to stop. A claims team that uses AI to draft customer letters faster has improved a step. A claims team that redesigns intake, triage, evidence review, exception handling, and quality control around AI has changed the work. Those are different ambitions, and they produce different economics.
The underwhelming result is therefore not a failure of imagination by technologists. It is a failure of management design. The most valuable AI opportunities sit across functional boundaries, inside decision rights, and beneath workflows that were built around human capacity constraints. Those constraints are changing. The operating model has not caught up.
What zero-based redesign means now
Zero-based redesign borrows the discipline of asking a first-principles question: if we were building this process today, with the tools now available, what would we design? The answer is almost never a clean automation of the current process. It usually involves removing steps, combining roles, changing what gets escalated, and deciding where human judgment is most valuable.
The phrase can sound like cost reduction dressed in new language. That is too narrow. In this context, zero-based redesign is a way to separate inherited work from necessary work. It asks which activities exist because the business still needs them, and which exist because the old process required them. Once that distinction is visible, AI becomes a design variable rather than a productivity feature.
Consider finance planning. Many teams begin by using AI to summarize variance explanations or draft commentary. That can help. But the larger question is whether the planning process still needs the same number of cycles, spreadsheet reconciliations, business-unit handoffs, and commentary packs. If AI can detect anomalies, draft explanations, and flag decisions, the meeting architecture itself may need to change.
The hard parts are not technical
Zero-based AI work is difficult because it forces management choices. A pilot can be owned by a function. A redesigned process usually crosses functions. A tool can be sponsored by technology. A new way of working requires business ownership. A productivity improvement can be measured in hours. A redesigned process must be measured in cost, cycle time, quality, risk, and accountability.
Governance is the first barrier. Most companies approve AI use cases one at a time. That works for experimentation, but it fragments the value case. The operating question is not whether a specific tool is acceptable. It is who owns the future-state process, who can remove redundant controls, who can change policies, and who has authority to retire old work.
The operating model is the second barrier. AI changes where work should sit. It can move activity from specialist teams to the front line, from analysts to managers, from monthly cycles to live decision routines. If roles, spans, and decision rights do not change, the organization pays for AI and keeps paying for the old structure too.
Incentives are the third barrier. Leaders often say they want efficiency, but local incentives reward stability. A function that gives up work may lose budget. A manager who changes a process may inherit risk before savings are visible. A team that removes reporting may be challenged by stakeholders who equate volume with control. The economics will not move unless the incentive system makes the new behavior rational.
What good looks like
The organizations making progress share four traits. First, they start with a process map that includes decisions, handoffs, controls, and ownership — not just activities. That makes the work visible as a system. Second, they build the value case at the process level. They do not count minutes saved in isolated tasks and call it transformation.
Third, they assign a senior business owner with authority to change how work runs. Technology remains essential, but the business has to own the design. Fourth, they create an adoption cadence that looks more like an operating review than a technology rollout. The weekly questions are concrete: what work was removed, what decision moved, what control changed, what savings are showing up in the run rate?
This is where many programs lose momentum. They celebrate deployment before the economics appear. A model goes live, a workflow changes, and the organization moves on. The better test is whether headcount demand, cycle time, rework, quality, and management attention are actually changing. If those measures do not move, the process has not been redesigned. It has been decorated.
The window is narrowing
AI cost advantage will not remain evenly distributed. In the early phase, access mattered. Then experimentation mattered. The next phase will be defined by management speed: which organizations can redesign work, change operating routines, and capture savings before the next budget cycle locks in the old model.
The timing matters because process economics compound. A redesigned process lowers the cost base, but it also changes how quickly the organization can absorb the next wave of tools. A company that keeps adding AI to fragmented work will keep rediscovering the same barriers. A company that redesigns the work once can reuse the new architecture as the technology improves.
For leadership teams, the practical question is no longer whether AI can create value. It is whether the organization is willing to let the value disturb the process. That means fewer pilots and more process ownership. Fewer isolated use cases and more operating-model decisions. Fewer demonstrations and more hard choices about how work should run.
The companies that move first will not simply be more automated. They will be structurally lighter, faster in decision cycles, and clearer about where human judgment belongs. That is the promise of zero-based redesign. It is also the reason many organizations will struggle to capture it. The technology is moving quickly. Management systems have to move with it.