The Redesign Dividend: Why Most Organizations Still Aren't Getting Value From AI, and How Leaders Can Redesign Work for a Human-Agent Workforce
Download the article here:
Listen to a review of this article:
Abstract: Despite several years of heavy investment in generative and agentic artificial intelligence, most organizations report little enterprise-level value. Recent surveys of chief information officers and chief executives show that only a small minority of enterprises have realized significant or transformational value, that most AI use cases fail, and that many leaders struggle to demonstrate financial returns. This article argues that the gap is primarily a work design problem rather than a technology problem. Drawing on field experiments, sociotechnical systems theory, job design research, and economic studies of automation, it explains why tool-centric pilots and headcount-first strategies tend to disappoint. It then outlines five evidence-based responses: redesigning workflows end to end, writing explicit role-redesign plans for a human-agent workforce, installing stopping rules for low-value initiatives, reinvesting freed capacity, and governing AI economics through outcome measures. Organizational examples from financial services, retail, biotechnology, education technology, and software illustrate both promising practice and costly missteps. The article closes with three long-term capabilities that allow organizations to capture what it calls the redesign dividend.
Ask a room of senior leaders whether their organization is using artificial intelligence and nearly every hand goes up. Ask whether that use has changed the economics of the business and most hands come down. That pattern, repeated in boardrooms and survey data alike, is the defining management puzzle of 2026. Spending is rising, tools are improving at a remarkable pace, and employees in many functions now use AI daily. Yet the enterprise-level returns that justified the investment remain elusive for most organizations.
The numbers are sobering. In Gartner's 2027 CIO and Technology Executive Agenda survey of 2,002 CIOs and technology executives, fielded between May and July 2026, only 13% said their enterprise had realized significant or transformational value from AI so far, and Gartner estimated the likelihood of any individual AI use case succeeding at roughly one in five (Gartner, 2026a). A day later, Gartner reported that 44% of CEOs in its 2027 CEO Survey struggle to demonstrate AI-driven financial returns to stakeholders (Gartner, 2026b). McKinsey's latest global survey tells a similar story: 37% of respondents attribute any EBIT impact to AI, essentially unchanged from the prior year, and only about 6% qualify as AI high performers (Tinkoff et al., 2026).
What separates the small group of organizations that are getting value from the much larger group that are not? The answer emerging across these studies is not the choice of model, the size of the budget, or the number of pilots. It is whether leaders redesign the work itself. Gartner found that organizations with explicit plans to redesign roles for a human-agent workforce were 2.4 times more likely to be among its "vanguard" performers, yet only 4% of CIOs reported having such plans (Gartner, 2026a). McKinsey's analysis similarly identifies the fundamental redesign of workflows as one of the strongest contributors to meaningful business impact among the factors it tested (Singla et al., 2025).
This article calls the value that flows from deliberate work redesign the redesign dividend. The term is meant to capture a simple idea with uncomfortable implications. AI does not produce value by being present in an organization. It produces value when tasks, roles, decision rights, and performance measures are reorganized so that human and machine capabilities complement one another. Organizations that skip this step tend to pay twice: once for the technology and again for the disruption of bolting it onto work that was designed for a different era.
The argument is not new in its essentials. Seventy-five years ago, researchers at the Tavistock Institute observed that mechanizing British coal mining without redesigning the social organization of work produced lower morale, more absenteeism, and disappointing productivity (Trist & Bamforth, 1951). Economists have since documented that general-purpose technologies typically depress measured productivity before they raise it, because firms must first invest in complementary intangible assets such as new processes, skills, and organizational structures (Brynjolfsson et al., 2021). What is new is the speed of the current technology, the arrival of AI agents that can carry out multistep tasks with limited supervision, and the temptation for leaders under cost pressure to treat AI primarily as a headcount reduction tool.
The sections that follow define the key concepts, examine the prevalence and drivers of the value gap, review its consequences for organizations and individuals, and set out five evidence-based responses. A final section describes the long-term capabilities that allow organizations to keep capturing value as the technology continues to change.
The AI Value Gap Landscape
Defining the Value Gap and the Human-Agent Workforce
Several terms recur throughout this article and deserve precise definitions.
The AI value gap refers to the distance between an organization's investment in and adoption of AI and the measurable business value it actually realizes. Adoption is easy to observe: licenses purchased, pilots launched, employees using a chat assistant. Value is harder to observe and is defined here as sustained improvement in outcomes that matter to the enterprise, such as unit costs, revenue growth, quality, cycle time, customer outcomes, or risk.
Generative AI refers to models that produce text, code, images, or other content in response to prompts. Agentic AI refers to systems that can plan and execute sequences of actions toward a goal, such as reading a support ticket, querying internal systems, drafting a resolution, and closing the case, with varying degrees of human oversight. Gartner's language of a human-agent workforce describes an operating model in which work is distributed deliberately between people and AI agents, and in which roles are designed with that distribution in mind (Gartner, 2026a).
Work design refers to the content and organization of tasks, activities, relationships, and responsibilities within a job and across a workflow (Parker, 2014). It includes what a role holder does, how much autonomy they have, what feedback they receive, and how their work connects to others. Role redesign is the deliberate revision of these elements, as opposed to simply adding a tool to an unchanged job.
Stopping rules are predefined criteria for ending an initiative. In the AI context, Gartner distinguishes between guidelines for when a pilot should not proceed to production and guidelines for abandoning initiatives already in production that are not delivering value (Gartner, 2026a). The underlying idea draws on decades of research on escalation of commitment, the well-documented tendency of decision makers to keep investing in a failing course of action, particularly when they feel personally responsible for the original decision (Staw, 1976).
Finally, sociotechnical systems thinking holds that organizational performance depends on the joint optimization of the technical system (tools, equipment, processes) and the social system (roles, relationships, skills, and norms). Optimizing one while ignoring the other tends to produce disappointing results (Trist & Bamforth, 1951). This idea sits at the heart of the redesign dividend.
Prevalence, Drivers, and Distribution
The value gap is wide and remarkably persistent. In McKinsey's 2026 survey of 1,719 respondents across 97 countries, fielded in May and June 2026, nearly nine in ten respondents reported regular AI use in at least one business function and 44% said AI was scaling across their enterprise, up from 38% a year earlier (Tinkoff et al., 2026). Yet only 37% attributed any EBIT impact to AI, a figure that did not move from the prior year, and the share of high performers, defined as organizations attributing at least 5% of EBIT to AI and reporting significant value, remained at about 6%. Gartner's 13% figure for significant or transformational value is consistent with this picture (Gartner, 2026a).
An earlier and more contested estimate came from MIT's NANDA initiative, which reviewed more than 300 publicly disclosed AI initiatives, interviewed representatives of 52 organizations, and surveyed 153 senior leaders in the first half of 2025 (Challapally et al., 2025). The authors concluded that the vast majority of organizations were getting no measurable return from generative AI and that only about 5% of organizations evaluating enterprise-grade tools had moved them into production. The methodology relies heavily on self-report and a modest sample, so the precise figures should be treated with caution. Still, the report's diagnosis is useful. The authors argued that the central barrier was not infrastructure, regulation, or talent but learning: most deployed systems did not retain feedback, adapt to context, or improve over time, and most organizations had not adapted their workflows around them.
Several drivers help explain why the gap persists.
The first driver is task substitution without workflow change. Many deployments insert AI into a single step of an existing process, such as drafting an email or summarizing a document, while leaving the surrounding process untouched. Individual employees may save minutes, but the saved time disappears into the cracks of the workflow rather than appearing in the profit and loss statement. McKinsey's data show that nearly three-quarters of high performers have fundamentally redesigned workflows, compared with roughly one-quarter of other respondents (Tinkoff et al., 2026).
The second driver is pilot proliferation. When AI is cheap to try and politically attractive to champion, organizations accumulate pilots faster than they can evaluate them. Gartner found that only 31% of CIOs have guidelines for deciding when a pilot should not move to production, and only 17% have guidelines for abandoning low-value initiatives already in production (Gartner, 2026a). Without these guidelines, weak initiatives linger, consuming budget, attention, and goodwill.
The third driver is cost uncertainty. Unlike traditional software with largely predictable licensing costs, the cost of generative and agentic AI is probabilistic and determined at run time by usage volumes and model choices. Gartner reported that 58% of CIOs face increased pressure to deliver AI-driven cost improvement while 51% expect AI to increase the total cost of ownership of technology solutions (Gartner, 2026a). About one in five McKinsey respondents said AI operating costs, including token costs, constrain their use (Tinkoff et al., 2026).
The fourth driver is a headcount-first mindset. Under investor and board pressure, some leaders frame AI primarily as a way to reduce payroll. Expectations of reduction are widespread: 39% of McKinsey respondents expected AI to reduce their organization's headcount over the next year (Tinkoff et al., 2026). Yet reality has lagged those expectations. Only 14% reported that AI had contributed to an overall workforce decline in the past year, less than half the 32% who had predicted reductions in the previous survey. Workday's global survey of roughly 6,000 employees and business leaders found that 40% of leaders expected AI to help them get more from existing employees, compared with 28% who expected it to reduce headcount (Workday, 2026). When leaders design programs around cuts that do not materialize, they often miss the more substantial opportunity of redeploying capacity toward higher-value work.
The fifth driver is a readiness gap at the top. In Gartner's CEO survey, 77% of CEOs expected AI industrialization to have a high or medium impact on their strategic objectives through 2027, but only 20% reported high readiness for it (Gartner, 2026b). Interestingly, 63% of CEOs named radically redesigning the operating model as a top move to accelerate tangible AI returns. Leaders appear to recognize the need for redesign in principle while lacking concrete plans to carry it out, which is consistent with Gartner's finding that only 4% of CIOs have explicit role-redesign plans (Gartner, 2026a).
The gap is not evenly distributed. Large organizations are scaling agents faster than smaller ones: 40% of large organizations reported scaling agents, compared with 22% of smaller ones (Tinkoff et al., 2026). Within organizations, gains also vary by task and by worker. Field experiments consistently show that AI helps most on tasks within its capabilities and can harm performance on tasks just beyond them, a pattern researchers have described as a jagged technological frontier (Dell'Acqua et al., 2023). The uneven shape of that frontier is one reason generic, enterprise-wide rollouts tend to produce generic, unimpressive results.
Organizational and Individual Consequences
Organizational Performance Impacts
At the level of individual tasks, the productivity evidence for generative AI is strong. In a study of 5,172 customer support agents at a software company, access to a generative AI assistant increased the number of issues resolved per hour by about 15% on average (Brynjolfsson et al., 2025). The gains were concentrated among less experienced and lower-skilled agents, whose speed and quality both improved, while the most experienced agents saw small speed gains and slight quality declines. In a preregistered experiment with 453 college-educated professionals, access to ChatGPT for midlevel writing tasks reduced time taken by 40% and raised output quality by 18% (Noy & Zhang, 2023). In a field experiment with 758 Boston Consulting Group consultants, those using GPT-4 on tasks inside the AI frontier completed 12.2% more subtasks and worked roughly 25% faster, with higher-quality output (Dell'Acqua et al., 2023).
These are large effects. Why, then, do they so rarely add up to enterprise value? Three mechanisms appear to be at work.
First, task-level gains do not automatically become organizational gains. A consultant who finishes a deliverable faster has created capacity, but that capacity produces value only if someone decides what to do with it. If work intake, approval processes, and staffing models stay the same, the saved time is absorbed by additional meetings, more revisions, or simply slack. Economists describe this as the need for complementary investments in processes and organization before a general-purpose technology shows up in productivity statistics (Brynjolfsson et al., 2021). The early phase of that J-shaped curve is costly, because organizations are investing in intangible assets whose returns arrive later.
Second, the jagged frontier creates hidden quality costs. In the same BCG experiment, consultants who used AI on a task deliberately designed to fall outside its capabilities were markedly less likely to reach the correct answer than consultants working without AI (Dell'Acqua et al., 2023). When organizations deploy AI uniformly without mapping where it is reliable, they risk trading visible speed gains for less visible errors. Lisanne Bainbridge (1983) identified a version of this problem four decades ago in her analysis of industrial automation: the more a system is automated, the more critical and the more difficult the remaining human role of monitoring and intervening becomes, because operators lose the hands-on practice that sustains their judgment.
Third, poorly governed AI can increase rather than reduce costs. When usage-based costs scale with adoption and few initiatives are ever stopped, the portfolio grows in cost faster than it grows in value. Gartner warned that legacy funding and governance approaches built for cloud and software-as-a-service are poorly suited to AI economics and recommended governing those economics across the full lifecycle rather than at the point of purchase (Gartner, 2026a).
The good news is that the evidence also points to a substantial dividend for organizations that address these mechanisms. Gartner's vanguards, defined as CIOs who deliver better business outcomes while also reporting high CEO-perceived AI value, were 2.3 times more likely to have pilot stopping guidelines, 2.6 times more likely to have production stopping guidelines, and 2.4 times more likely to have role-redesign plans (Gartner, 2026a). These are associations rather than proof of causation, but they align with McKinsey's finding that workflow redesign is among the strongest predictors of business impact (Singla et al., 2025).
Individual Wellbeing and Stakeholder Impacts
The consequences of the value gap are not only financial. How organizations pursue AI value shapes the daily experience of work, career prospects, and trust.
The most detailed view so far of what happens inside a heavily AI-intensive workplace comes from Anthropic's own economics team, which analyzed internal data on how work and employment at the company changed between 2025 and 2026 (Wilmers et al., 2026). The findings illustrate both the promise and the strain of a human-agent workforce. In an open-ended April 2026 employee survey, 27% of respondents described productivity gains from Claude tools, and 20% described their work shifting from doing tasks directly to managing multiple instances of Claude. Code commits per engineer rose to roughly five times their level six months earlier. At the same time, calendar data showed that the median employee spent 20% more time in meetings than a year earlier, with technical individual contributors seeing a 16% increase. The authors also found that Claude sessions were idle around 80% of the time, waiting for human input, which suggests that human attention and coordination, rather than machine capacity, had become the binding constraint. The authors are careful to note that these patterns are descriptive, that tool adoption was not random, and that Anthropic's circumstances are unusual.
Even with those caveats, the Anthropic data point to an important shift in the nature of work. When employees move from executing tasks to orchestrating agents, their jobs change in ways classic work design research would predict. Hackman and Oldham's (1976) job characteristics model identified skill variety, task identity, task significance, autonomy, and feedback as core features that shape motivation and satisfaction. Supervising agents can raise autonomy and task significance, but it can also fragment task identity and reduce the hands-on skill use that many professionals find meaningful. Parker and Grote (2022) argue that digital technologies make work design more consequential rather than less, because the same technology can either enrich or impoverish jobs depending on the choices managers make.
The job demands-resources model offers a further lens (Bakker & Demerouti, 2007). It holds that strain arises when job demands, such as workload, coordination effort, and monitoring, outpace job resources, such as autonomy, support, and skill development. Rising meeting loads and the cognitive demands of reviewing machine output can be meaningful new demands. If organizations add agents without adding resources, they risk higher strain even as output rises.
Career pathways are also changing. At Anthropic, the share of junior employees among new technical hires fell from 27% to 9%, and the junior share of technical staff fell from 21% to 14% (Wilmers et al., 2026). The authors stress that these figures are descriptive rather than causal. Workday's data show broader pressure on internal mobility: moves to new roles inside the same company fell at 57% of employers year over year, and promotion rates were essentially flat (Workday, 2026). Employees perceive a gap between what they need and what they receive. While 79% said they know what skills they need to succeed, only 66% said their employer actually helps them develop those skills. Among workers who use AI for nearly all of their work, 62% believe it will make their current skills less valuable, even as 76% expect it to open new career opportunities.
These findings matter for organizational performance as well as for individuals. If junior roles disappear, the pipeline that produces tomorrow's experienced professionals narrows. Brynjolfsson et al. (2025) found that AI helped novices learn faster, effectively disseminating the tacit knowledge of top performers. An organization that eliminates entry-level roles may forgo exactly the learning mechanism that makes AI most valuable.
Customers and other stakeholders are affected too. When AI-only service models degrade quality, customers notice. When AI is framed as a vehicle for layoffs, employees and the public notice, and the resulting loss of trust can slow adoption. Research on psychological safety suggests that people are more willing to experiment, report errors, and learn when they believe they will not be punished for doing so (Edmondson, 1999). A workforce that fears replacement is unlikely to surface the honest feedback that AI systems and redesigned workflows need in order to improve.
Evidence-Based Organizational Responses
The research and organizational experience reviewed above point to five responses that distinguish organizations capturing the redesign dividend from those stuck in the value gap. They are mutually reinforcing: workflow redesign identifies where value lies, role redesign makes that value sustainable, stopping rules protect the portfolio, capacity reinvestment converts time savings into outcomes, and outcome-based governance keeps the whole system honest.
Table 1. Five evidence-based responses for closing the AI value gap
Response | Evidence Base | Key Practices | Illustrative Examples |
Redesign workflows end to end | Sociotechnical systems theory; productivity J-curve; McKinsey high-performer data | Map tasks across the whole process; locate AI where it is reliable; redesign handoffs, approvals, and intake | Moderna; Morgan Stanley |
Write explicit role-redesign plans for a human-agent workforce | Gartner vanguard data; job characteristics model; task-based economics | Define agent and human responsibilities; create new roles; broaden role scope deliberately; protect learning | IKEA; Walmart; Anthropic |
Install stopping rules at pilot and production stages | Gartner vanguard data; escalation-of-commitment research; MIT NANDA findings | Define value before launch; set kill criteria; share stopping decisions with business owners; revisit live systems | Klarna |
Reinvest freed capacity rather than defaulting to cuts | Workday and McKinsey workforce data; field experiments on novice learning; new-task creation | Decide in advance where saved time goes; redeploy into growth and quality work; communicate honestly | Salesforce; Duolingo |
Govern AI economics through outcome measures | Gartner CIO and CEO surveys; jagged frontier evidence; ironies of automation | Track unit economics and customer outcomes, not usage; manage run-time costs; integrate workforce and technology governance | Shopify; Anthropic |
Redesign Workflows End to End
The most consistent finding in the AI value literature is that organizations that redesign whole workflows outperform those that add AI to individual tasks. McKinsey's relative-weights analysis of 31 organizational practices found that fundamentally redesigning workflows was among the strongest contributors to meaningful business impact (Singla et al., 2025). In the 2026 survey, nearly three-quarters of high performers had done so, up from 55% a year earlier, compared with roughly one-quarter of other organizations (Tinkoff et al., 2026). Gartner's CEO research reached a similar conclusion, advising leaders to prioritize AI investments that change workflows, decision making, and how value is delivered, rather than simply accelerating existing tasks (Gartner, 2026b).
The theoretical foundation for this finding is old and sturdy. Trist and Bamforth's (1951) study of the longwall method of coal mining showed that introducing new machinery while fragmenting the work into narrow, isolated tasks undermined the very productivity gains the machinery was meant to deliver. Their insight, that technical and social systems must be designed together, applies with particular force to AI agents, which change not just how a task is performed but who coordinates with whom, who holds decision rights, and where errors are caught.
Workflow redesign also addresses the jagged frontier. Because AI is highly capable on some tasks and unreliable on others that appear similar, organizations need to map tasks carefully and place AI where it adds value while preserving human judgment where it is most needed (Dell'Acqua et al., 2023). That mapping is impossible when AI is deployed as a general-purpose assistant with no reference to a specific process.
Effective approaches:
Process-level mapping: Break each target workflow into tasks and decisions, then assess each for AI suitability, risk, and the human judgment it requires, rather than starting from the capabilities of a particular tool.
Redesigned handoffs: Revisit approvals, review steps, and intake queues so that time saved in one step is not lost waiting for the next.
Frontier testing: Pilot AI on representative tasks, including deliberately difficult edge cases, to learn where it is reliable before scaling.
Joint ownership: Assign a business owner and a technology owner to each redesigned workflow, with shared accountability for outcomes.
Integrated planning: Plan technology, process, and talent changes together rather than in separate cycles.
Moderna merged its human resources and information technology functions into a single People and Digital Technology department led by Tracey Franklin, who became chief people and digital technology officer (Nawrat, 2025). Franklin described the aim as shifting from separate workforce planning and technology planning toward what she called work planning, in which leaders decide how work should flow across people and AI agents working side by side. The company had developed more than 3,000 custom GPTs with OpenAI, and Franklin described the organization as still early in its agentic AI journey. The structural choice is notable because it places the design of work, rather than the deployment of tools or the management of headcount, at the center of planning.
Morgan Stanley Wealth Management took a workflow-specific approach. After fully rolling out an internal generative AI assistant that gave financial advisors faster access to the firm's research and intellectual capital, and reporting that 98% of advisor teams had adopted it, the firm launched a second tool called Debrief in June 2024 (Morgan Stanley, 2024). With client consent, Debrief takes notes during client meetings, highlights action items, drafts a follow-up email the advisor can edit, and saves a note to the firm's client relationship system. One advisor reported saving roughly half an hour per meeting. The firm framed the purpose explicitly in terms of freeing advisors to spend more time on client relationships, which it described as remaining central to the advisor role. The example shows how targeting a well-defined, high-frequency workflow step and connecting it to downstream systems creates value that a generic chat tool would not.
A common pitfall is the "bolt-on" deployment, in which a powerful tool is licensed for everyone and success is measured by log-ins. Usage rises, individual employees report time savings, and nothing in the operating model changes. Within a year, leaders struggle to point to any outcome that moved, and the initiative becomes a line item to defend rather than a source of advantage.
Write Explicit Role-Redesign Plans for a Human-Agent Workforce
Gartner's single most striking finding may be that organizations with explicit plans to redesign roles for a human-agent workforce were 2.4 times more likely to be vanguards, while only 4% of CIOs reported having such plans (Gartner, 2026a). Gartner's analyst characterized workforce redesign, rather than headcount reduction, as the key differentiator, and anticipated that organizations will need fewer distinct role types, each with a wider scope of responsibility. The CEO survey reinforced the point: 52% of CEOs named aggressively deploying AI agents as employees as a top move for accelerating AI returns, but Gartner advised governing AI, digital labor, and workforce redesign as a single integrated transformation (Gartner, 2026b).
Economic research on automation helps explain why role redesign matters. Acemoglu and Restrepo (2019) show that technology affects labor through two opposing forces: a displacement effect, as machines take over tasks previously performed by people, and a reinstatement effect, as new tasks are created in which human labor has a comparative advantage. Whether a technology raises or lowers labor's contribution depends on the balance between them. Organizations cannot control the macroeconomic balance, but they can control whether they actively create new tasks and roles within their own walls. Role redesign is, in effect, deliberate reinstatement.
Job design theory adds a second rationale. Roles that retain skill variety, task identity, autonomy, and feedback support motivation and performance (Hackman & Oldham, 1976). As agents absorb routine tasks, the remaining human work can become richer, with more judgment, relationship building, and creative problem solving, or it can become thinner, consisting mainly of monitoring and exception handling. The difference lies in deliberate design. Research on job crafting further shows that employees actively reshape the boundaries of their own work to make it more meaningful (Wrzesniewski & Dutton, 2001). Effective role redesign combines top-down structure with room for this kind of bottom-up crafting.
Effective approaches:
Explicit responsibility charts: For each redesigned role, specify which tasks agents perform, which humans perform, and which require joint work, including who is accountable for agent output.
New role creation: Identify emerging roles such as agent builders, agent supervisors, and workflow designers, and create pathways into them from existing jobs.
Broadened role scope: Consolidate narrow, fragmented roles into broader ones that give people end-to-end ownership of outcomes, consistent with Gartner's expectation of fewer, wider roles.
Protected learning tasks: Preserve tasks that build expertise for junior employees, even where AI could perform them, so the talent pipeline does not dry up.
Co-design with employees: Involve role holders in redesign, drawing on their knowledge of where AI helps and where it gets in the way.
IKEA offers one of the clearest examples of role redesign at scale. As its AI customer service bot, Billie, handled 47% of customer queries to call centres over a two-year period, the company trained 8,500 call centre workers as interior design advisers beginning in 2021 (Reuters, 2023). Remote design services delivered by phone or video brought in €1.3 billion in the 2022 financial year, about 3.3% of revenue, and Ingka Group set a goal of raising that share to 10% by 2028. Asked whether AI would reduce headcount, an Ingka executive said that was not what the company was seeing. Rather than treating chatbot deflection as a cost saving, IKEA used it to free staff for higher-value advisory work tied to a growth strategy.
Walmart, with about 2.1 million employees worldwide, has said it expects its headcount to remain roughly flat over the next three years even as AI changes the mix of jobs, eliminating some tasks and roles while adding others (Coacci, 2025b). Chief people officer Donna Morris described tracking which job functions are growing, shrinking, or holding steady in order to direct reskilling. The company created an agent builder position in which employees build AI tools to help merchants, and it joined OpenAI's certification program to build AI skills. Chief executive Doug McMillon framed the aim as helping employees enhance their current roles with new tools, while keeping people in roles that interact directly with customers.
Anthropic's internal data offer an early glimpse of what role change looks like when agent use is pervasive. The fact that one in five survey respondents described a shift from doing work to managing Claude instances, alongside sharp increases in meta-work, standing agent instructions, and coding by nontechnical staff, suggests that roles are being redefined in practice whether or not they are redefined on paper (Wilmers et al., 2026). The lesson for other organizations is that role change will happen with or without a plan. The question is whether it happens deliberately, with attention to motivation, learning, and accountability, or haphazardly.
A common pitfall is redesign by attrition: allowing roles to erode as tasks are automated, without updating job descriptions, performance expectations, or development paths. Employees end up with a residue of tasks that do not form a coherent job, managers lack criteria for evaluating them, and the most capable people leave for clearer opportunities elsewhere.
Install Stopping Rules at Pilot and Production Stages
If workflow and role redesign are how organizations capture value, stopping rules are how they avoid wasting it. Gartner found that organizations with guidelines for when an AI pilot should not move to production were 2.3 times more likely to be vanguards, and those with guidelines for abandoning low-value initiatives already in production were 2.6 times more likely (Gartner, 2026a). Yet these guidelines remain uncommon, at 31% and 17% of CIOs respectively. Gartner recommended that organizations define use-case value in advance, set explicit criteria for stopping, and treat those criteria as a shared effort with business partners willing to drop low-value investments.
The case for stopping rules rests on robust behavioral research. Staw's (1976) experiments on escalation of commitment showed that people who felt responsible for an initial investment decision allocated more additional resources to a failing course of action than those who did not. AI initiatives are especially vulnerable to escalation because they are often sponsored by senior leaders, publicly announced, and tied to narratives of competitive necessity. Without predefined criteria, every disappointing pilot can be rationalized as needing more data, a better model, or more time.
The MIT NANDA findings underscore the stakes. Of organizations that evaluated enterprise-grade AI tools, only a small fraction had moved them into production, and the authors attributed much of the failure to tools that did not learn from feedback or fit into workflows (Challapally et al., 2025). Stopping rules do not prevent failure, but they shorten its duration and free resources for the minority of initiatives that are working.
Effective approaches:
Value hypotheses up front: Before a pilot begins, document the outcome it is expected to change, the baseline, the target, and the time frame for evaluation.
Stage-gate criteria: Set explicit criteria for moving from pilot to production, covering value, quality, risk, cost at scale, and workflow fit.
Production sunset reviews: Review live AI systems on a regular schedule against their original value case and retire those that no longer earn their keep.
Separated sponsorship and evaluation: Have someone other than the original sponsor assess whether an initiative should continue, reducing the pull of escalation.
Celebrated stopping: Treat a well-reasoned decision to stop as a success that frees resources, not as a failure to be hidden.
Klarna's experience illustrates both the appeal of aggressive AI substitution and the value of being willing to change course. In early 2024, the company reported that its AI assistant had handled about three-quarters of customer service chats in its first month, roughly 2.3 million conversations, and was doing the equivalent work of 700 agents (Shibu, 2025). Klarna had paused hiring, and headcount fell by about 22% to roughly 3,500, mostly through attrition. By May 2025, chief executive Sebastian Siemiatkowski told Bloomberg that although AI was cheaper, it had produced lower quality, and that the company would hire human customer service staff again so customers could always reach a person. Klarna began recruiting fully remote service workers.
Klarna's reversal is sometimes presented as a failure of AI. It is more usefully read as a delayed stopping decision. The company eventually applied a quality criterion to an AI-only service model and adjusted. Had that criterion been specified in advance, with customer outcomes measured alongside cost per contact, the adjustment might have come sooner and at lower cost to customer experience and reputation.
A common pitfall is defining stopping rules only for pilots. Production systems acquire constituencies, integrations, and sunk costs that make them hard to retire, and Gartner's data suggest that production stopping guidelines are both rarer and more strongly associated with success than pilot guidelines (Gartner, 2026a). Organizations that never revisit live systems accumulate an expanding portfolio of AI costs with no corresponding discipline on value.
Reinvest Freed Capacity Rather Than Defaulting to Cuts
When AI saves time, someone must decide what happens to that time. The default in many organizations is to treat it as a headcount opportunity. The evidence suggests this default is often mistaken, not because cost reduction is illegitimate, but because it is frequently smaller and slower than expected and because it can foreclose larger gains.
Workday's data indicate that more leaders expect AI to help them get more from current employees (40%) than expect it to reduce headcount (28%) (Workday, 2026). McKinsey's data show a large gap between predicted and realized workforce reductions: 32% of respondents in the prior year predicted reductions, while only 14% reported that AI had contributed to a workforce decline (Tinkoff et al., 2026). MIT NANDA's interviews found that the best back-office returns, such as reduced spending on outsourced business processes and external agencies, came without material reductions in the internal workforce (Challapally et al., 2025). Gartner's analyst explicitly contrasted workforce redesign with headcount reduction as the vanguard differentiator (Gartner, 2026a).
Field evidence suggests why reinvestment can pay. Brynjolfsson et al. (2025) found that AI disproportionately helped less experienced workers, accelerating their learning. Noy and Zhang (2023) found that AI reduced inequality in output among workers. These effects make existing employees more capable and adaptable, which is valuable only if they remain in the organization. Acemoglu and Restrepo's (2019) reinstatement mechanism likewise depends on organizations creating new tasks for people to do.
Effective approaches:
Capacity plans before deployment: For each initiative, decide in advance whether freed capacity will go to growth, quality, risk reduction, new services, or cost reduction, and track whether it does.
Redeployment pathways: Build internal mobility mechanisms that move people from roles where demand is shrinking to roles where it is growing.
Contractor and vendor levers: Where cost reduction is needed, look first at outsourced spending, where MIT NANDA found substantial returns.
Honest communication: Tell employees clearly what AI will and will not mean for their jobs, and keep those commitments.
Growth-linked targets: Tie at least some AI value cases to revenue, customer, or quality outcomes rather than purely to cost.
Salesforce provides a mixed but instructive example. Chief executive Marc Benioff said in September 2025 that the company had reduced its support headcount from about 9,000 to about 5,000 because AI agents were handling much of the work, with roughly half of customer interactions now involving agents (Burleigh, 2025). A company spokesperson said Salesforce had stopped backfilling support engineer roles and had moved hundreds of employees into professional services, sales, and customer success. The example shows that redeployment and reduction can coexist, and that the balance between them is a leadership choice. It also illustrates the reputational sensitivity of such choices, since the announcement drew considerable public attention.
Duolingo shows how communication shapes outcomes. In April 2025, chief executive Luis von Ahn shared a memo describing an AI-first strategy that called for gradually reducing contractors doing work AI could handle and adding headcount only when teams could not automate more of their work (Coacci, 2025a). The memo drew strong public criticism and fears of layoffs. In an interview several months later, von Ahn said he had not given enough context, that the company had never laid off full-time employees and did not plan to, and that salaried staff were expected to accomplish more with AI. Duolingo also instituted weekly Friday sessions in which teams experiment with using AI to become more efficient. The episode suggests that how leaders frame AI's workforce implications can matter as much as the underlying policy.
A common pitfall is announcing savings before they exist. When leaders commit publicly to cost reductions based on pilot results, they create pressure to cut roles before workflows have been redesigned to function without them. The result can be degraded service, overloaded remaining staff, and quiet rehiring, a sequence that erodes trust with both employees and investors.
Govern AI Economics Through Outcome Measures
The final response addresses how organizations know whether they are capturing value at all. Gartner's CEO research recommended measuring outcomes rather than AI activity, tracking unit economics, decision velocity, speed to market, customer outcomes, and business impact (Gartner, 2026b). Its CIO research emphasized that AI costs are probabilistic and determined at run time, requiring new funding, governance, and cost-accountability models (Gartner, 2026a). With 44% of CEOs struggling to demonstrate financial returns, measurement is not a technical afterthought but a leadership necessity (Gartner, 2026b).
Measurement matters because the most visible indicators of AI progress, such as adoption rates, number of use cases, and self-reported time savings, are poorly correlated with business value. McKinsey found that 80% of respondents said AI had improved their individual productivity, yet only 37% reported any EBIT impact (Tinkoff et al., 2026). The gap between those two figures is, in a sense, the value gap expressed in a single comparison. Measures that capture individual activity will consistently overstate organizational value unless they are paired with outcome measures.
Measurement also guards against the hidden quality costs of the jagged frontier and the ironies of automation (Bainbridge, 1983; Dell'Acqua et al., 2023). If organizations track speed but not accuracy, or volume but not customer outcomes, they will systematically favor deployments that look productive while eroding quality.
Effective approaches:
Outcome scorecards: For each AI-enabled workflow, track a small set of business outcomes, such as cost per resolved case, error rates, cycle time, and customer satisfaction, alongside usage.
Run-time cost management: Monitor model and usage costs continuously, set budgets by workflow, and choose model tiers according to task value.
Quality sampling: Regularly review samples of AI-assisted output, especially on tasks near the edge of AI capability.
Integrated governance: Bring technology, finance, human resources, and business leaders into a single forum that governs AI, digital labor, and workforce design together.
Human-experience indicators: Track meeting load, workload, learning opportunities, and internal mobility in AI-intensive teams, not just output.
Shopify illustrates a strong leadership signal and the measurement questions it raises. In a memo shared with staff in late March 2025 and reported in April, chief executive Tobi Lütke described reflexive AI use as a baseline expectation for everyone at the company, said teams must demonstrate why they cannot accomplish work with AI before requesting more headcount, and announced that AI usage questions would be added to performance and peer reviews (Nolan, 2025). The memo made AI adoption an explicit organizational priority and created a gate on headcount growth. For other organizations considering similar mandates, the open question is what is being measured. A review question about AI usage captures activity. Whether that activity improves merchant outcomes, product quality, or unit economics requires separate measures, and organizations that rely on usage alone risk the activity trap that Gartner warns against.
Anthropic's internal workforce report illustrates the kind of measurement that can reveal hidden costs (Wilmers et al., 2026). By combining calendar data, code repository data, session logs, and surveys, the company's economists identified rising meeting time, high idle time in agent sessions, and a shrinking junior share of hires, alongside large gains in code output. Equally important, the report is candid about what the data cannot show, noting that more code does not necessarily mean better code or business impact and that the junior-hiring patterns are descriptive. That combination of rich measurement and stated limitations is a useful model for any organization trying to understand what AI is actually doing to its work.
A common pitfall is measuring what is easy rather than what matters. Dashboards that celebrate adoption curves and prompt counts reassure sponsors but give no signal about whether to scale, adjust, or stop. Over time, they can turn AI programs into exercises in demonstrating activity rather than creating value.
Building Long-Term Redesign Capability
The five responses above can close part of the value gap in the near term. Sustaining the redesign dividend over years requires deeper organizational capabilities, because the technology will keep changing and each new generation of models and agents will redraw the jagged frontier. Three capabilities stand out.
Work Design Literacy Across the Leadership Bench
Most managers were never trained to design work. They were trained to manage people within roles that someone else designed, often decades earlier. In a human-agent workforce, every manager becomes, in part, a work designer, deciding which tasks go to agents, how output is reviewed, and how roles evolve. Parker (2014) argues that work design affects not only motivation but also learning, health, and an organization's ability to balance exploration with exploitation, which makes it a strategic capability rather than an administrative one.
Building work design literacy means teaching leaders the core principles of job design, including the job characteristics that sustain motivation (Hackman & Oldham, 1976), the balance of demands and resources that protects wellbeing (Bakker & Demerouti, 2007), and the sociotechnical insight that technical and social systems must be designed together (Trist & Bamforth, 1951). It also means equipping them to support job crafting, so that employees can shape their own roles within a redesigned structure (Wrzesniewski & Dutton, 2001). Organizations can embed these principles in leadership development, in templates for AI business cases, and in the criteria used to approve new deployments. A business case that cannot describe how roles will change is not ready for approval.
Workday's job-posting data contain a warning here. Mentions of management and leadership skills in job postings fell 7%, and mentions of training skills fell 13%, between September 2025 and July 2026, even as demand for hands-on AI skills rose 51% (Workday, 2026). If organizations are deprioritizing exactly the managerial and developmental capabilities that redesign requires, they may find themselves with more technical capacity and less ability to organize it.
Integrated People and Technology Governance
The second capability is structural. In most organizations, technology decisions are made in one forum, workforce decisions in another, and financial decisions in a third. AI value depends on the intersection of all three. Gartner's recommendation to govern AI, digital labor, and workforce redesign as one integrated transformation (Gartner, 2026b) and its call for new funding and cost-accountability models (Gartner, 2026a) both point toward governance that spans functional boundaries.
Moderna's merger of human resources and technology into a single department is one structural answer (Nawrat, 2025), though not every organization will or should go that far. Lighter-weight alternatives include joint planning cycles in which workforce and technology plans are built together, shared portfolio reviews that apply common stopping rules, and integrated scorecards that combine financial, operational, and human indicators. The common principle is that no AI initiative should be approved, scaled, or stopped without considering its effects on work design and people, and no workforce plan should be built without considering what AI will change.
Integrated governance also creates the conditions for honest stopping decisions. When business owners, technology leaders, and people leaders jointly own an initiative's value case, it becomes easier to separate the evaluation of results from the sponsorship of the original decision, which is the most reliable way to counter escalation of commitment (Staw, 1976).
Learning Systems and Talent Pipelines
The third capability concerns learning, at both the system and the individual level. MIT NANDA's central diagnosis was that most AI deployments fail because neither the tools nor the organizations around them learn from experience (Challapally et al., 2025). The productivity J-curve suggests the same: returns depend on accumulated intangible investments in process knowledge and skills, which take time to build (Brynjolfsson et al., 2021).
At the system level, organizations need feedback loops that capture what employees learn about where AI works and where it fails, and that feed this knowledge into workflow and role design. Such loops depend on psychological safety, because employees must feel able to report errors, question AI output, and admit that a redesigned process is not working (Edmondson, 1999). Leaders who frame AI primarily as a threat to jobs make that candor less likely.
At the individual level, organizations need talent pipelines that keep producing experienced professionals even as entry-level tasks are automated. The Anthropic data on falling junior hiring and Workday's data on declining internal mobility suggest that this pipeline is under pressure (Wilmers et al., 2026; Workday, 2026). Organizations that want experienced judgment in five years must design roles today that let early-career employees build it, for example by pairing them with agents in ways that accelerate learning rather than replacing the learning entirely. Brynjolfsson et al.'s (2025) finding that AI helped novices absorb the practices of top performers suggests that well-designed human-agent roles can shorten, not eliminate, the path to expertise.
Closing the 13-point gap Workday identified between employees who know what skills they need and those whose employers help them develop those skills is one practical starting point (Workday, 2026). Others include protected development time, rotational assignments into new human-agent roles, and transparent internal job markets that make it easier for people to move toward growing work.
Conclusion
The AI value gap is real, wide, and persistent. Most organizations have adopted AI, most employees who use it report personal productivity gains, and most enterprises still cannot point to significant business value. The evidence reviewed in this article suggests that the gap is not primarily a matter of model quality, budget, or enthusiasm. It is a matter of work design.
Field experiments show that AI can make individual tasks dramatically faster and better, especially for less experienced workers, while also degrading performance on tasks just beyond its reach. Economic research shows that general-purpose technologies pay off only after organizations invest in complementary changes to processes and skills, and that the labor effects of automation depend on whether new tasks are created alongside displaced ones. Sociotechnical and job design research, some of it now three-quarters of a century old, shows that technology introduced without attention to how work is organized tends to disappoint. The newest survey data from Gartner and McKinsey confirm the pattern: the organizations getting value are disproportionately those that redesign workflows and roles, define value in advance, and stop initiatives that are not delivering.
The redesign dividend is available to any organization willing to do this harder work. It does not require abandoning cost discipline, and it does not require treating every job as permanent. It does require leaders to resist two tempting shortcuts. The first is pilot proliferation, the belief that enough experiments will eventually add up to transformation. The second is headcount-first thinking, the belief that cutting payroll is the primary or fastest route to AI value. Both shortcuts substitute visible activity for the less visible but more consequential work of redesigning how people and agents work together.
Redesign the work, not just the task: Map whole workflows, place AI where it is reliable, and rework handoffs, approvals, and intake so that time savings become business outcomes.
Put role redesign on paper: Write explicit plans that define what agents do, what people do, and how roles will broaden, and include pathways into new roles such as agent builders and supervisors.
Decide in advance when to stop: Set value hypotheses and stopping criteria for pilots and for production systems, and separate evaluation from sponsorship.
Choose where freed capacity goes: Plan reinvestment into growth, quality, and service before deployment, and treat headcount reduction as one option rather than the default.
Measure outcomes, not usage: Track unit economics, quality, customer outcomes, and human-experience indicators alongside adoption, and govern run-time costs continuously.
Protect the learning pipeline: Preserve developmental tasks for early-career employees and invest in internal mobility so that tomorrow's expertise is still being built.
Govern people and technology together: Bring technology, finance, and people leaders into shared forums so that AI, digital labor, and workforce design are planned as one system.
Research Infographic

References
Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks: How technology displaces and reinstates labor. Journal of Economic Perspectives, 33(2), 3–30.
Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775–779.
Bakker, A. B., & Demerouti, E. (2007). The job demands-resources model: State of the art. Journal of Managerial Psychology, 22(3), 309–328.
Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. Quarterly Journal of Economics, 140(2), 889–942.
Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The productivity J-curve: How intangibles complement general purpose technologies. American Economic Journal: Macroeconomics, 13(1), 333–372.
Burleigh, E. (2025, September 2). Salesforce CEO Marc Benioff says his company has cut 4,000 customer service jobs as AI steps in: "I need less heads." Fortune.
Challapally, A., Pease, C., Raskar, R., & Chari, P. (2025). The GenAI divide: State of AI in business 2025. MIT NANDA.
Coacci, J. (2025a, August 18). Duolingo CEO admits his controversial AI memo "did not give enough context" and insists the company never laid off full-time employees. Fortune.
Coacci, J. (2025b, September 30). Walmart CEO says he can't think of a single job that won't be changed by AI: Here's how today's workforce can prepare. Fortune.
Dell'Acqua, F., McFowland, E., III, Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality (Harvard Business School Working Paper No. 24-013). Harvard Business School.
Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383.
Gartner. (2026a, October 19). Gartner survey finds only 13 percent of CIOs say their enterprise has realized significant or transformational value from AI [Press release].
Gartner. (2026b, October 20). Gartner says 44 percent of CEOs struggle to demonstrate AI financial returns as technology becomes central to growth [Press release].
Hackman, J. R., & Oldham, G. R. (1976). Motivation through the design of work: Test of a theory. Organizational Behavior and Human Performance, 16(2), 250–279.
Morgan Stanley. (2024, June 26). Morgan Stanley Wealth Management announces latest game-changing addition to suite of GenAI tools [Press release].
Nawrat, A. (2025). Why Moderna merged HR and IT to better "architect the flow of work." UNLEASH.
Nolan, B. (2025, April 8). Shopify CEO tells employees to prove AI can't do jobs before asking for new hires. Fortune.
Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192.
Parker, S. K. (2014). Beyond motivation: Job and work design for development, health, ambidexterity, and more. Annual Review of Psychology, 65, 661–691.
Parker, S. K., & Grote, G. (2022). Automation, algorithms, and beyond: Why work design matters more than ever in a digital world. Applied Psychology, 71(4), 1171–1204.
Reuters. (2023, June 13). IKEA bets on remote interior design as AI changes sales strategy. Reuters.
Shibu, S. (2025, May 9). Klarna is hiring customer service agents after AI couldn't cut it on calls, according to the company's CEO. Entrepreneur.
Singla, A., Sukharevsky, A., Hall, B., Yee, L., Chui, M., & Balakrishnan, T. (2025). The state of AI in 2025: Agents, innovation, and transformation. McKinsey & Company.
Staw, B. M. (1976). Knee-deep in the big muddy: A study of escalating commitment to a chosen course of action. Organizational Behavior and Human Performance, 16(1), 27–44.
Tinkoff, D., Van der Veken, L., Chui, M., & Balakrishnan, T. (2026). The state of AI in 2026: On the road to ROI. McKinsey & Company.
Trist, E. L., & Bamforth, K. W. (1951). Some social and psychological consequences of the longwall method of coal-getting. Human Relations, 4(1), 3–38.
Wilmers, N., Catipon, R., Massenkoff, M., & McCrory, P. (2026). How employment is changing at Anthropic. Anthropic.
Workday. (2026, October 5). Workday Global Workforce Report: AI is rewriting jobs more than it's cutting them [Press release].
Wrzesniewski, A., & Dutton, J. E. (2001). Crafting a job: Revisioning employees as active crafters of their work. Academy of Management Review, 26(2), 179–201.

Jonathan H. Westover, PhD, Chief Research Officer (Nexus Institute for Work and AI); Co-Founder & Chief Workforce and Learning Officer (Future State University); Founder & CEO (Human Capital Innovations); Professor of Organizational Leadership & Change (UVU). Read Jonathan Westover's executive profile here.
Suggested Citation: Westover, J. H. (2026). The Redesign Dividend: Why Most Organizations Still Aren't Getting Value From AI, and How Leaders Can Redesign Work for a Human-Agent Workforce. Human Capital Leadership Review, 40(4). doi.org/10.70175/hclreview.2020.40.4.4






















