In NVIDIA’s 2026 survey on retail and fast-moving consumer goods, 91% of respondents report using or evaluating AI; this does not measure global enterprises in general nor daily usage. AI can shift part of the work toward verification and reprocessing by users, but the distribution of this burden according to salary, bargaining power, or hierarchical position must be established case by case. In certain Asian organizational contexts characterized by high power distance, employees may be less inclined to provide upward feedback. According to the global Workday study published on January 14, 2026, nearly 40% of the time saved through AI is lost in correction, rewriting, and verification of low-quality outputs.
The essentials
- 91% of globalized enterprises use AI in 2026, compared to 55% in 2023 (Workday Global Research 2026).
- 40% of productivity gains are offset by correction costs imposed on downstream recipients, amounting to $9M in annual losses for an organization of 10,000 people.
- The mechanism is asymmetrical: gains go to those who deploy the tool, costs fall on those who cannot refuse the task.
- In Asia, hierarchy blocking upward feedback exacerbates the silent accumulation of these losses.
- Organizations that reconfigured their workflows before AI deployment avoided this cost transfer.
AI doubles the sender’s speed, not the system’s
A writer who generates a report in twenty minutes instead of two hours saves time. Their manager validates faster. But downstream, the quality controller, the local translator, the customer support agent, or the data entry operator receives a document with more errors, more subtle ones, and arriving sooner than expected. The speed of production has been decoupled from the quality of what is produced. The real gain is in the first link; the real cost is in the following ones.
The 2026 Workday survey documents this mechanism at scale. Workday estimates that nearly 40% of the time saved through AI is lost in reprocessing. Separately, NVIDIA reports that 91% of retail/CPG respondents use or evaluate AI. These teams typically do not appear in AI productivity dashboards. They can absorb a cost that adoption indicators do not capture.
The figure of $186 per employee per month represents the average cost of this reprocessing, calculated across all exposed positions. At the scale of an organization of 10,000 people, this exceeds $9 million annually. Most decision-makers who approved AI deployment did not account for this line item in their return-on-investment analysis.
In Asia, hierarchy amplifies the mechanism
The problem exists everywhere. In certain Asian contexts with high power distance, problem reporting can be more indirect; the specific effect on AI costs remains to be established.
In certain organizational contexts, reporting defective results from a tool provided by management can be difficult. An employee who receives poor-quality AI content may spend time correcting it without this burden being systematically reported. The tool is officially a success. Productivity gains are displayed on the sender’s side. Losses remain invisible on the receiver’s side.
The absence of reliable reporting channels and fear of retaliation can create information blind spots, but their form and intensity vary depending on organizations and national contexts. Certain Asian organizations may accumulate specific risk factors, but a comparative regional generalization is not established.
The economics of gain-sharing: three years of predictions
Daron Acemoglu, MIT economist whose recent work focuses precisely on AI governance in the workplace, formulates a thesis directly applicable to this mechanism. In an article published in The New Yorker in 2026, he argues that generative AI creates shared prosperity only if institutions or organizational norms actively weigh on the distribution of gains. According to Acemoglu, institutions, market structures, and norms influence how the benefits and costs of technological change are distributed.
The 2026 Workday data illustrate this mechanism. The measurable productivity gain goes to the sender, often a manager or qualified employee with access to high-performing AI tools. The correction cost goes to the receiver, often an operations agent with little latitude to negotiate their workload or escalate the problem. The technology amplifies inequalities along existing lines of power.
Acemoglu’s thesis has its critics. Tyler Cowen, economist at George Mason University whose work on stagnation and innovation are regularly cited, argues that transition frictions are temporary and that labor markets adjust. In this reading, the correction cost imposed on receivers is a temporary inefficiency that competition between employers will eventually correct, since an organization that loses $9 million in invisible reprocessing eventually gets beaten by a competitor that configured its workflows better.
This argument deserves to be taken seriously. But it presupposes that information circulates. Obstacles to reporting can limit the upward flow of information about downstream difficulties. The market only corrects what it observes. When the cost is silent and the one bearing it has no means to make it visible, spontaneous adjustment is delayed, or does not occur.
Organizations that avoided the cost transfer have one thing in common
The 2026 Workday report emphasizes the importance of adapting roles and processes when deploying AI. The formulation is simple. What it encompasses is structurally different.
Reconfiguring a workflow before deploying AI requires identifying who receives what the tool will produce, in what timeframe, and with what verification capacities. Standard deployment starts with the sender, measures the sender’s gain, and stops there. Rigorous deployment starts from the complete system, integrating downstream friction points, before defining acceptable speed parameters.
This distinction is not technical. It is organizational and political. It requires that downstream teams have had a voice in designing the deployment, which implies either a corporate culture favoring information reporting, or a formalized process that compensates for the absence of such a culture. Both exist. Both require direction intention, not simply a technology deployment.
Shared access to research infrastructure in Europe illustrates an analogous mechanism: when productive resources are designed by one group and used by another, governance of the interface becomes the central stake, more than the technology itself.
Several major Korean and Japanese companies have begun integrating downstream burden audits into their AI deployment protocols. LG CNS, the digital subsidiary of the LG Group, established in 2025 a committee for assessing side effects in which representatives of execution teams sit before any extension of a tool to a new business line. The approach remains minority in the region, but it signals that an organizational solution to the problem exists and works.
The next phase of adoption: what it will determine
The 2030-2032 horizon leaves open a question that current data do not yet resolve: will the productivity generated by AI translate into real wages and redistribution, or will it remain captured in capital margins and costs imposed on uncompensated executors.
Two trajectories are plausible, depending on the institutional and organizational choices of the next four years.
Gains measured at the task level do not always translate into measurable gains at the enterprise scale; tracking of reprocessing and verification burden varies by measurement tools and organizations. Organizations may retain indicators that do not capture downstream burden. Employees tasked with AI quality control may see their workload increase without this contribution being systematically factored into their evaluation.
In the second trajectory, enterprises that have understood the cost-transfer mechanism measure it, internalize it into their AI performance indicators, and rebuild their workflows to redistribute the burden. This involves compensating differently for AI verification work, which did not exist in this form before deployment. Studies conducted at Harvard Business School on task decomposition show that critical review work on AI outputs constitutes a distinct skill from traditional human proofreading: it requires knowledge of the tool’s biases, its failure modes, and an ability to distinguish plausible error from obvious error. This skill is not recognized in current classifications.
Latin American engineers accelerating American AI startups illustrate this recomposition: the AI value chain already distributes tasks globally, but control skills and their accompanying compensation follow logics still largely inherited from the pre-AI world.
The signal to watch by 2028 is the following: do enterprises that measure downstream costs integrate them into salary negotiations with the affected teams. If so, the second trajectory is underway. If downstream burden audits are not accompanied by reflection on cost and benefit distribution, the distributive effects of technological change described by Acemoglu may persist.
Governing deployment or letting hierarchy absorb it
The institutional pressure that Acemoglu calls for does not necessarily have to come from the state. It can come from organizations themselves, provided they have the right indicators and an internal governance structure that gives visibility to downstream teams.
Three levers have been tested with documented results. The first is downstream burden audit prior to deployment, modeled on LG CNS. The second is integrating a reprocessing cost indicator into AI project tracking dashboards, which a few major financial services enterprises in Australia and Singapore have begun experimenting with. The third is creating an AI quality control bonus, distinct from traditional performance bonuses, that explicitly recognizes verification work as a task in its own right.
None of these levers requires immediate regulatory intervention. All require a governance decision that management must make before deployment, not after. The experience of organizations that managed to avoid cost transfer suggests that this decision must be made when the tool is still in the pilot phase, when the workflow can still be reconfigured without internal political cost.
A minimum standard is emerging slowly in the most advanced enterprises: any large-scale AI deployment should include a mapping of receivers, a measurement of their reprocessing burden, and a review of their ability to report it. It spreads through professional contagion more than through regulatory prescription. It could accelerate if major consulting firms that oversee deployments in Asia integrated it into their standard methodologies.
The question that remains open is one of pace. The relationship between AI adoption in Asia and organizations’ ability to reconfigure their workflows remains to be established. Workday documents a gap between AI use and adaptation of roles and processes; the NVIDIA source examined does not allow confirmation of the same organizational maturity indicator.
Sources
- Workday Global Research 2026 & NVIDIA APAC Survey, summary and data via AI Business Weekly: https://aibusinessweekly.net/p/ai-productivity-statistics
- Daron Acemoglu (2026), Can A.I. Be Pro-Worker?, The New Yorker: https://www.newyorker.com/contributors/john-cassidy
- NVIDIA APAC Survey 2025, cited via the primary source (aibusinessweekly.net)
- Harvard Business School, studies on AI task decomposition (task-level study): Harvard Business School Working Knowledge, no direct URL available