The “slop tax” and the capacity paradox
Why human-centered design is essential to enterprise AI
The discourse surrounding enterprise AI has fractured into two equally loud, equally unhelpful camps.
On one side sits a desperate echo chamber of technical hype: the belief that inserting an LLM wrapper into an inherently broken workflow, or handing over your codebase to an agent and turning your best engineers into “enter button” pushers and vibe code copy editors, will magically generate exponential ROI. On the other side is a mounting wave of catastrophic panic: the doom-laden narrative that uncontrollable models are on the verge of dismantling human civilization, and we are all powerless to stop it.
Both arguments share a fundamental flaw: they exist entirely in the clouds, detached from the messy, day-to-day reality of modern business operations.
While pundits bicker over existential threats and hockey-stick projections, enterprise practitioners are dealing with the cold reality on the ground. The bell on enterprise AI adoption cannot be unrung. Millions of dollars have already been sunk into token budgets, proprietary codebases have been handed over to third-party models, and core operational workflows have been re-architected. AI isn’t coming; it is already embedded within the plumbing of global enterprise infrastructure, and “automate more, ship more, profit more” is an ever-louder klaxon, coming from all corners of the tech sector. But klaxons don’t provide roadmaps.
Recent Gartner research indicates that enterprise AI initiatives are struggling to demonstrate consistent financial returns: 57% of the use cases evaluated by finance leaders had unclear returns, while 19% delivered negative returns, according to CFO.com. In the rush to demonstrate that they are embracing AI, organizations risk confusing enthusiasm with strategy and adoption with progress. The pressure to keep pace can make critical questions about purpose, readiness, risk, and value seem like resistance rather than the due diligence required to make sound decisions. As I argued in my August essay, From Could to Should: Human-Centered Design as Organizational Decision-Making, this is what happens when organizations rush implementation without first establishing a strategy. We asked “Could we?” and skipped “Should we?” Now organizations are confronting the consequences: substantial investments, expanding AI use, and no consistent evidence that the resulting systems are delivering the value promised.
The “slop tax” is the hidden cost of making AI-generated work reliable. The “capacity paradox” is what happens when an organization expects AI to replace staff or solve a staffing problem without having the human and operational capacity to implement, oversee, and sustain it, or a clear understanding of which use cases are likely to deliver the greatest value and how that value will be measured.
Reversing course from blind adoption to reactionary backlash would be costly, if it is even feasible for organizations that have already integrated AI into their operations. We do not need a retreat, nor do we need more costly hype. We need a calm, clear-eyed framework for human-led stewardship: a targeted roadmap for scaling AI at a measured pace, grounded in organizational readiness, operational capacity, and demonstrable value.
The historical playbook: dehumanization as optimization
To understand why current enterprise AI rollouts are triggering such intense operational friction (along with the daily lurid headlines), we need to look at the management logic organizations are bringing to this technology.
For more than a century, businesses have sought to improve efficiency by standardizing tasks, measuring output, formalizing process-improvement methodologies, and finding ways to produce more at lower cost. One influential approach was scientific management, developed by Frederick Winslow Taylor in the late nineteenth and early twentieth centuries. Taylor argued that work could be studied systematically to identify more efficient ways of performing it. In practice, his model concentrated the planning and control of work in management, while workers were expected to execute tasks according to prescribed methods.
This approach helped shape modern management, but it also reinforced a way of seeing work that can obscure the value of the person doing it. When a job is reduced to a list of measurable tasks, the judgment, experience, relationships, and problem-solving required to perform it well can become difficult to see, let alone quantify.
Over time, this task-based logic intersected with other forces, including automation, shareholder-value pressures, and corporate downsizing. Labor increasingly became something to optimize on a spreadsheet: a cost to reduce, a headcount to justify, or a collection of tasks that might be performed more cheaply elsewhere. Modern rightsizing did not originate with Taylor alone, but it can draw on the same underlying assumption: that efficiency can be measured by reducing the resources required to produce a given output.
This way of thinking extends well beyond any single management theory. Corporate performance metrics, business education, and management consulting can all reinforce it when they prioritize measurable cost reductions over less visible investments in expertise, organizational learning, and resilience. The problem is not any one institution or discipline, but the incentives that make immediate savings easier to quantify and reward than the long-term capabilities an organization needs to perform well.
The tension is visible in contemporary corporate leadership. In May 2026, Standard Chartered CEO Bill Winters drew criticism for speaking of replacing “lower-value human capital with financial and investment capital,” later apologizing for his remarks amid job cuts. In a different context, a 2026 investigation by 404 Media into Proton, a company known for its privacy and security products, reported that CEO Andy Yen pushed for faster AI adoption while dismissing requests for additional security testing as “loser thinking.”
These are different situations, but the tensions they expose are related. Human expertise can be treated as a cost to reduce, while the work required to exercise that expertise can be treated as an impediment to speed. Yet judgment, security, and verification are not incidental overhead when they are essential to the value a business delivers. For a company whose reputation rests on privacy and security, the ability to establish that its products are safe is part of the product itself.
The consequences become particularly serious when reducing costs is mistaken for improving performance. An organization can eliminate positions and report immediate savings while losing institutional knowledge, overloading the people who remain, weakening its ability to develop new experts, or shifting work into less visible forms of verification and rework. The spreadsheet may show a leaner operation. It does not necessarily show a more capable one.
Walmart’s multibillion-dollar warehouse automation effort illustrates what can happen when organizations approach automation as a labor-replacement exercise rather than an opportunity to understand and improve the work itself. Its robotic systems have encountered problems with routine tasks, product variations, packaging, and reliability, while introducing substantial operating costs and new staffing requirements. Some of these challenges might have been anticipated through a comprehensive assessment of warehouse operations that included the expertise of workers who understand those operations firsthand. Which tasks are genuinely suitable for automation? Where are the actual bottlenecks? What workarounds and human interventions keep the operation running? These are not questions that can be answered by looking at a task list or a labor-cost spreadsheet alone.
When generative AI arrived, many organizations applied this familiar management logic to a new technology. Driven by pressure to demonstrate rapid returns, they identified tasks people performed, mapped those tasks onto AI capabilities, and began treating automation as a straightforward route to lower labor costs and higher productivity. The underlying assumption remained largely unchanged: if technology could perform part of someone’s job, the organization could reduce the human resources devoted to it.
But a probabilistic language model is not a mechanical gear. A job is rarely just the sum of its visible tasks. It also includes the judgment required when something goes wrong, the contextual knowledge needed to interpret ambiguous information, the relationships that make collaboration possible, and the experience that helps people recognize when an apparently reasonable answer is wrong.
AI did not create this management problem. It gives organizations a powerful new way to reproduce it at scale.
The rise of the algorithmic janitor
The evidence of this structural failure is mounting. According to an Atlanta Federal Reserve working paper cited by The Guardian, 90% of executives surveyed said AI efforts had yet to translate into productivity gains at their companies. Meanwhile, an EY survey published in September 2026 found that 36% of senior AI executives at companies with at least $1 billion in annual revenue reported an AI incident or failure that caused a materially negative impact, including data loss, financial damage, brand damage, or operational disruption.
One contributor to disappointing returns may be an unaccounted-for “slop tax.” As Wharton School Assistant Professor Jessica Reif explains in The Guardian, generative models can produce “plausible-looking output” with weaknesses that are not immediately apparent. Errors can propagate through a workflow rather than being corrected at the point where they occur, allowing operational problems to compound.
The hidden cost is not just correcting bad output. It includes reviewing results, evaluating quality, managing exceptions, resolving failures, and remaining accountable for decisions made with AI assistance. Organizations may also need new tools, training, and evaluation practices, while determining which tasks are appropriate to delegate to a model in the first place. If success is measured by task completion time or apparent labor savings, these costs can remain invisible. When implementation outpaces the processes needed to support it, the work does not necessarily disappear. It moves, often to people whose additional effort is missing from the original business case.
But the cost of AI-generated output extends beyond the effort required to correct individual mistakes. There is a fundamental difference between using AI to resolve customer problems faster and using it to increase the volume of features shipped. The first begins with a problem and seeks to improve an outcome. The second begins with a capacity to produce and risks treating increased output as evidence of progress.
Get Sara Danielle Reiss’s stories in your inbox
Join Medium for free to get updates from this writer.
Shorter development cycles can be valuable when they help teams address customer needs, resolve defects, and improve the completion of real tasks. But speed alone does not establish that a feature is needed, that it works reliably, or that it improves the product. Those questions require research, validation, and evaluation. For organizations that already struggle to prioritize user research, the ability to generate features faster can become a justification for bypassing the work of understanding what customers actually need. Security review, accessibility testing, and quality assurance can likewise be recast as obstacles to delivery rather than the disciplines that make delivery successful.
There is also a systems-level cost. Software consists of interdependent components, and each addition must work within the wider system. A feature that functions correctly in isolation may behave differently when integrated with existing components, data flows, permissions, and workflows. A fix can resolve one defect while introducing another. As systems grow more complex, diagnosing failures, testing interactions, and establishing that changes have not introduced new problems can demand increasingly substantial effort.
AI may reduce the time required to generate code, but it does not automatically reduce the work required to understand, validate, secure, integrate, and maintain that code. If production accelerates faster than an organization’s capacity to evaluate and sustain what it produces, the resulting complexity and rework can erode the efficiency the technology was intended to deliver. More output is not inherently more value, and a feature that is cheap to generate may be expensive to own.
When employees are discouraged from raising legitimate concerns about security, quality, or reliability, the organization risks weakening the very feedback mechanisms it needs to determine whether its investment is working. Speed that depends on suppressing scrutiny is not a sustainable operating model.
Ground truth: the staffing and workflow paradox
I encountered this friction firsthand while evaluating and designing human-in-the-loop AI workflows within a highly regulated anti-financial crime (AFC) operations environment.
The initiative began as a proof of concept to reduce the manual workload involved in reviewing large volumes of unstructured financial information. The proposed AI agent could ingest thousands of pages of data, cross-reference information against regulatory sources, and generate preliminary investigation summaries. The goal was to reduce repetitive work so that a small team of AFC subject-matter experts could devote more time to specialized, cognitively demanding risk assessments, without being stretched thin by the volume of manual work.
But the ability to generate an output was only one part of the problem. Because a language model is probabilistic, its output can appear authoritative even when it misses a critical transaction or misinterprets a regulatory concept.
Investigators still needed to establish whether the machine’s output was accurate, traceable to appropriate sources, relevant to the jurisdiction in question, and suitable for the compliance decision at hand. Without explicit source traceability, permission boundaries, and structured paths for human review, they could be left reconstructing the basis of an output before they could responsibly use it. In a highly scrutinized environment, this verification burden is not incidental. It is an operational cost and risk that the business case must account for.
The deeper paradox emerged as leadership began treating the existence of the tool as a reason not to increase staffing, even though the team still needed additional capacity. What had been conceived as a way to relieve pressure on subject-matter experts risked becoming a justification for maintaining the very resource constraints it was intended to help address. The tool could reduce some manual work, but it could not eliminate the need for people to investigate, exercise judgment, validate outputs, and manage the work that remained.
The organizational challenges compounded this problem. When key AI leadership departed, the gaps in the support structure became glaringly obvious. There was no clear mechanism connecting ongoing business operations, governance, and implementation. Leadership uncertainty, unclear ownership, and gaps in compliance sign-offs stalled the project. The team was left with the prospect of introducing a probabilistic tool into an already stretched operation, without the organizational support needed to implement and oversee it responsibly.
My discovery work identified that the challenge was not simply to improve the model. It was to establish an information architecture and user-focused workflow that could make human accountability operational through clear escalation paths, deterministic permission structures, and explicit data lineage. The proposed direction was designed to support machine assistance while giving investigators a way to understand and verify the basis of its outputs. The initiative stalled before these recommendations were implemented.
The business case for such a system cannot be evaluated solely by how much material it processes or how many hours it might theoretically save on a spreadsheet. It must also account for verification, exception handling, oversight, and the organizational capacity required to sustain reliable use.
There is a longer-term cost, too: the erosion of the conditions through which expertise develops. Entry-level, high-volume work is often where junior operators build the judgment, contextual knowledge, and practical experience that allow them to become tomorrow’s senior subject-matter experts. If organizations automate those foundational opportunities without considering how that expertise will be developed elsewhere, they risk weakening the very capabilities they will eventually need. A workforce is not simply a collection of labor costs to optimize against a technology budget. It is an essential part of an organization’s future capacity.
The lesson is not that AI cannot help address staffing constraints. It is that automation cannot substitute for the organizational capacity required to make it work. A tool intended to free up human expertise can become a justification for withholding the resources needed to exercise that expertise. A system can be technically capable and still fail to deliver value when ownership, governance, operational support, and the authority to act on evidence are missing.
The business case for structural discipline
The same gap between ambition and evidence is visible in enterprise AI infrastructure spending. VentureBeat Intelligence’s October 2026 analysis found that 79% of surveyed companies running AI in production and reporting GPU utilization said their GPUs were busy half the time or less. Yet 70% of that group planned to adopt a new, additional, or replacement compute platform within the next year. In the September survey, 91% of companies planning a new platform had no return-on-investment figure reviewed by their finance team.
These findings do not establish that the planned investments are unnecessary. Low utilization can have multiple explanations, and a planned platform may replace existing infrastructure rather than add to it. But the findings raise a fundamental question about investment discipline: are organizations validating what they have already invested in before committing to the next layer of capability?
The disconnect between technological ambition and organizational readiness extends beyond financial returns. Microsoft CEO Satya Nadella has acknowledged that AI alignment does not automatically follow from increasing model capability, calling for deterministic system design, human controls, and reliable operating procedures. Against that backdrop, Microsoft’s voluntary buyout offers to a portion of its workforce raise a legitimate question: how will the company maintain the expertise, oversight, and operational capacity needed to implement those safeguards as its AI systems become more capable? Recognizing the problem is important, but recognition alone does not establish that the resources and processes required to address it are in place. Sounding the alarm is not the same as providing a roadmap.
This principle applies whether an organization is evaluating a model, redesigning a workflow, or investing in infrastructure. The question is not simply what the technology can do, but what the organization needs it to achieve, what it will take to achieve it, and how the result will be measured. Good design decisions are only part of the equation. Organizations also need the people, expertise, and operational processes to put those decisions into practice.
Before treating AI as a silver bullet for a capacity problem, organizations need to examine the full system in which the technology will operate:
- Identify the problem and the outcomes that need to improve before selecting a technology.
- Define which tasks AI can realistically perform and where human judgment remains necessary.
- Calculate the end-to-end costs of verification, error recovery, training, oversight, integration, and downstream work.
- Establish clear ownership, accountability, governance, and operational support so that the system can be implemented and sustained.
- Measure outcomes over an appropriate period, including quality, reliability, capacity, and financial returns, rather than adoption or output volume alone.
Human-Centered Design is a way to preserve and restore human agency. It helps organizations understand work as it actually happens, surface the trade-offs that technology-led decisions can obscure, and evaluate what needs to be true for an investment to deliver meaningful value. HCD is not a cosmetic corrective applied after technological decisions have been made. It is part of the discipline required to make those decisions well.
The organizations that realize lasting value from AI will not necessarily be those that adopt it fastest, spend the most, or automate the largest share of their work. They will be those that understand what the technology can improve, what human capabilities the work still requires, and what organizational conditions are necessary to make the investment worthwhile. The question is not simply whether AI can perform a task or reduce the cost of doing it, but whether the resulting system works better, delivers measurable value, and can be sustained by the people and processes responsible for it.
That is the difference between adopting a technology and designing an organization that can use it well.
