Past the Prompt: Why AI Skill Demand Is Shifting From Using to Building, and How Leaders Can Close the Skills-Support Gap
Download the article here:
Listen to a review of this article:
Abstract: Employer demand for artificial intelligence (AI) skills is moving quickly from basic use toward hands-on building. Workday's October 2026 Global Workforce Report finds that requisition demand for simple prompting peaked in January 2026 and then fell 25%, while demand for building AI tools, automating workflows, and AI engineering rose 51% between September 2025 and July 2026. In the same report, 79% of workers say they know which skills they need, but only 66% say their employer helps them develop those skills. This article examines that mismatch: the shrinking shelf life of entry-level AI skills, the gap between employee awareness and employer support, and the consequences for performance and wellbeing. Drawing on field experiments, training-transfer research, and labor-market evidence, it proposes five evidence-based responses: skills-based capability mapping, funded and time-protected reskilling, learning by building, champion and manager networks, and outcome-aligned expectations. It closes with three pillars for durable AI capability.
For a while, being "good with AI" at work meant knowing how to write a decent prompt. Organizations ran lunch-and-learns, circulated prompt libraries, and counted logins. That phase is closing faster than most learning calendars can adjust. Workday's October 2026 Global Workforce Report, based on skills data from job requisitions at roughly 550 enterprise employers, shows that demand for basic AI skills such as simple prompting rose through late 2025, peaked in January 2026, and then fell 25% over the following months. Over the same broad period, demand for hands-on skills such as building AI tools, automating workflows, and AI engineering climbed 51% between September 2025 and July 2026 (Workday, 2026). In Workday's own summary, employers have moved from wanting people who can use AI to wanting people who can build with it.
That shift would be manageable if organizations were well equipped to help people make it. The same report suggests they are not. In a global survey of about 6,000 employees and business leaders, 79% of workers said they know what skills they need to succeed, yet only 66% said their employer actually helps them develop those skills, a 13-point gap (Workday, 2026). Leaders, meanwhile, are signaling that they intend to grow capability rather than cut their way to productivity: 40% expect AI to help them get more out of the employees they already have, compared with 28% who expect it to reduce headcount. Put those numbers together and a clear tension appears. Leaders are betting on their current workforce, workers know what they need to learn, and the bridge between the two, employer-supported development, is the weakest part of the structure.
This article is about that bridge. It does not revisit whether AI adoption is spreading (it is) or whether AI will take jobs (the evidence is mixed and the subject of a large literature). Instead, it asks a narrower and more practical question for people leaders and learning professionals: how should organizations design learning systems that move a workforce from AI literacy to AI fluency and, for a growing share of roles, to AI building? The answer matters because the skills that earned a premium eighteen months ago are now table stakes, and the skills that are rising in value are harder to teach. Building an internal tool, automating a multistep workflow, or wiring an assistant into a team's data requires practice, feedback, and protected time, none of which a one-hour webinar provides.
There is also a quieter signal in the Workday data worth taking seriously. Mentions of management and leadership skills in requisitions fell 7%, and mentions of training skills fell 13%, between September 2025 and July 2026 (Workday, 2026). Organizations appear to be asking for less of exactly the capabilities that would help workers through this transition. If managers and trainers are the people who convert intent into practice, deprioritizing them while demanding faster reskilling is a design flaw, not a cost saving.
The sections that follow define the key terms, describe the drivers and distribution of the shift, examine consequences for organizations and individuals, and then set out five evidence-based responses, each illustrated by documented organizational practice. The article closes with three pillars for building capability that lasts longer than the next tool release.
The AI Skills Shift Landscape
Defining AI Literacy, Fluency, and Building
Much of the confusion in corporate AI learning comes from treating "AI skills" as a single thing. It helps to separate at least four levels of capability, which can be thought of as rungs on a ladder.
AI literacy refers to a working understanding of what AI systems can and cannot do, how to use approved tools safely, and how to give them basic instructions. The Organisation for Economic Co-operation and Development (OECD) distinguishes this kind of general AI literacy from the specialized expertise held by AI professionals, and warns that current training supply may not be sufficient to meet the growing need for general literacy alone (OECD, 2025). Simple prompting sits on this rung.
AI fluency goes further. A fluent user integrates AI into everyday tasks, judges the quality of outputs, and knows when not to rely on the tool. The importance of that judgment is well documented. In a preregistered field experiment with 758 Boston Consulting Group consultants, Dell'Acqua et al. (2023) found that AI users completed 12.2% more tasks and worked 25.1% faster on tasks inside what the authors called the "jagged technological frontier," but were less likely to produce correct solutions on a task deliberately chosen to fall outside it. Fluency, in other words, is not just speed with a tool. It includes the discernment to recognize where the frontier lies for a given task.
AI building is the capability Workday's requisition data show rising most sharply: creating custom assistants, automating workflows, connecting tools to data, and, at the technical end, AI engineering (Workday, 2026). Building does not always require writing code. Many enterprise platforms now let non-engineers configure assistants or chain steps together. But it does require a different mental model, closer to process design than to conversation.
AI stewardship is the fourth rung in this article's framework, and it is the least formalized. It describes people who redesign work and teams around AI, govern risk, and develop others. Microsoft's 2025 Work Trend Index captures a version of this with its idea of the "agent boss," a person who builds, delegates to, and manages AI agents, and it reports that 41% of leaders expect their teams to be training agents within five years (Microsoft, 2025). Stewardship also overlaps with the management and training skills that, in the Workday data, are receding from job requisitions.
Two further terms recur below. A skills-based organization is one that organizes work and talent decisions around skills rather than fixed job titles (Jones et al., 2022). Training transfer is the extent to which what is learned in training is applied and maintained on the job, a long-studied construct in industrial-organizational psychology (Baldwin & Ford, 1988; Blume et al., 2010).
Finally, the idea of a skill's half-life, the time it takes for a skill to lose much of its market value, is used here as a metaphor rather than a measured quantity. The evidence does not support precise half-life estimates for AI skills. What it does support is that skill requirements are changing faster in AI-exposed work. PwC's analysis of close to one billion job advertisements found that skills sought by employers were changing 66% faster in occupations most exposed to AI (PwC, 2025), and its 2026 edition reports that skills for the most AI-exposed jobs are changing more than twice as fast as for the least exposed jobs (PwC, 2026). The rise and fall of demand for prompting inside a single year is a vivid illustration of the same pattern (Workday, 2026).
Prevalence, Drivers, and Distribution
The broad context is one of rapid, sustained skills churn. The World Economic Forum's Future of Jobs Report 2025, based on a survey of more than 1,000 employers representing over 14 million workers in 55 economies, estimates that 39% of workers' core skills will be transformed or become outdated between 2025 and 2030 (World Economic Forum, 2025). AI and big data top its list of fastest-growing skills. The same report offers a sobering way to picture the training task: if the global workforce were 100 people, 59 would need training by 2030, of whom 29 could be upskilled in their current roles, 19 could be upskilled and redeployed, and 11 would be unlikely to receive the training they need.
Three drivers help explain why demand is shifting from using to building.
The first is commoditization. As AI tools become embedded in mainstream software and their interfaces improve, basic prompting becomes something most knowledge workers can do with little instruction. Microsoft and LinkedIn's 2024 Work Trend Index, which surveyed about 31,000 knowledge workers across 31 markets, already found that 75% of knowledge workers used AI at work and that 78% of AI users were bringing their own tools (Microsoft & LinkedIn, 2024). When a skill becomes nearly universal, it stops differentiating candidates, and employers stop listing it.
The second driver is the search for measurable return. Organizations that invested in licenses and literacy programs are now looking for productivity gains that show up in processes, not just in individual drafts. That requires people who can redesign a workflow, automate handoffs, or build a team-specific assistant. Microsoft's 2025 survey found that 47% of leaders list upskilling existing employees as a top workforce strategy for the next 12 to 18 months, and 78% are considering hiring for AI-specific roles (Microsoft, 2025).
The third driver is the spread of AI work beyond technology functions. Lightcast's analysis of more than 1.3 billion job postings found that, as of 2024, 51% of postings requesting AI skills were outside information technology and computer science, and that postings listing AI skills advertised salaries about 28% higher, nearly $18,000 per year (Lightcast, 2025). PwC reported an average wage premium of 56% for jobs requiring AI skills in its 2025 analysis, up from 25% the previous year (PwC, 2025). These estimates use different methods and should not be compared directly, but they point in the same direction: AI capability carries a market premium, and that premium is no longer confined to engineers.
The distribution of support is uneven. In the 2024 Work Trend Index, only 39% of people who used AI at work had received AI training from their employer (Microsoft & LinkedIn, 2024). Boston Consulting Group's 2025 survey of more than 10,600 workers across 11 countries found that 72% used AI regularly, but regular use among frontline workers had stagnated at 51%, only 36% of respondents felt adequately trained, and just 25% of frontline workers said their leaders provided enough AI guidance (Boston Consulting Group, 2025). The Workday gap between knowing and being supported (79% versus 66%) is therefore not an anomaly. It is the latest reading of a persistent pattern in which training lags adoption, and adoption lags ambition.
Organizational and Individual Consequences
Organizational Performance Impacts
The performance case for closing the gap rests on robust experimental evidence that AI can raise productivity when people know how to use it well. In a large field study of generative AI in a real workplace, Brynjolfsson et al. (2025) studied the staggered rollout of an AI conversational assistant to 5,172 customer-support agents and found a 15% average increase in issues resolved per hour, with larger gains in speed and quality for less experienced and lower-skilled workers. The authors interpret part of this effect as the AI tool spreading the tacit practices of top performers to everyone else. In an online experiment with 453 college-educated professionals, Noy and Zhang (2023) found that access to ChatGPT cut time on writing tasks by about 40% and raised rated quality by about 18%, with the biggest gains among those who started out weaker.
These studies are encouraging, but they also carry a warning for leaders. The gains came from fairly well-defined tasks in which the tool's strengths were matched to the work. Dell'Acqua et al. (2023) showed what happens when the match is wrong: on a task outside the frontier, consultants using AI did worse than those working without it. The organizational implication is that literacy without judgment can destroy value as easily as create it. A workforce trained only to prompt may produce more output while making more confident errors.
At the macro level, PwC's 2026 analysis reports that productivity growth is 40% higher at the companies most exposed to AI than at the least exposed (PwC, 2026). Such correlations do not prove causation, and the most AI-exposed firms may differ in many other ways. Still, combined with the experimental evidence, they suggest a plausible mechanism: the organizations that convert AI access into process change, which is exactly what building skills enable, capture more value.
The skills-support gap therefore has a direct cost. If 40% of leaders expect AI to help them get more from current employees (Workday, 2026), that expectation depends on employees acquiring the next rung of capability. Without employer support, organizations face three predictable outcomes. Some employees will self-teach, usually unevenly and outside approved tools. Some will stall at literacy and plateau. And some will leave, taking newly acquired skills to an employer willing to pay the premium documented by Lightcast (2025) and PwC (2025). There is also a talent-supply angle. Employers in Workday's data reported that roles often stay open because candidates lack the right skills (Workday, 2026). Building capability internally is often faster and cheaper than competing for scarce external builders, especially when, as LinkedIn's 2025 Workplace Learning Report notes, 49% of learning and talent respondents agree their executives worry that employees lack the skills to execute business strategy (LinkedIn, 2025).
Finally, the decline in requisition demand for management and training skills (Workday, 2026) creates a capacity problem. Learning research consistently finds that a supportive work environment is one of the strongest predictors of whether training transfers to the job (Blume et al., 2010). Managers and peer trainers are the most important part of that environment. An organization that thins out these capabilities while asking for faster reskilling is reducing the very conditions that make reskilling work.
Individual Wellbeing and Stakeholder Impacts
For individuals, the shift from using to building is both a threat and an opportunity, and workers seem to see it that way. In Workday's August 2026 AI@Work Pulse survey of 5,944 workers, 62% of those who use AI for nearly all their work said they think it will make their current skills less valuable, yet 76% of the same heavy users expect AI to open up new career opportunities (Workday, 2026). That combination, worry about current skills alongside optimism about new ones, is precisely the psychological state in which good development support matters most. With support, uncertainty can be channeled into growth. Without it, uncertainty tends to become anxiety.
Confidence is not the constraint. Sixty-five percent of workers in the Workday survey said they were confident they could learn new skills if asked to (Workday, 2026). The constraint is access: to time, to tools, to projects, and to people who can coach. When those resources are distributed unevenly, the gains from AI are likely to be distributed unevenly too. The BCG finding that frontline regular use has stalled at 51% while overall use reached 72% (Boston Consulting Group, 2025) hints at a two-speed workforce in which office-based employees accumulate AI capability and frontline workers fall further behind. The World Economic Forum's estimate that 11 of every 100 workers are unlikely to receive the training they need by 2030 (World Economic Forum, 2025) puts a number on the people most at risk.
There is a wellbeing dimension as well. Microsoft's 2025 survey found that 80% of the global workforce, including leaders, say they lack enough time or energy to do their work (Microsoft, 2025). Asking people to reskill on top of full workloads, without protected time, converts a development opportunity into another source of strain. And because AI tools tend to deliver their biggest boosts to less experienced workers (Brynjolfsson et al., 2025; Noy & Zhang, 2023), withholding structured support from those workers may also mean withholding the equalizing benefits that the research has documented.
Other stakeholders are affected too. Customers and clients bear the risk when partially trained employees deploy AI outside its frontier. Managers carry the burden when expectations rise faster than support. And the employment relationship itself is tested. Workday's data show employees staying put in a cautious labor market but becoming less likely to recommend their employer, with internal moves falling at 57% of employers (Workday, 2026). Visible, credible investment in people's AI capability is one of the few levers that can turn reluctant staying into genuine commitment.
Evidence-Based Organizational Responses
The evidence suggests that closing the gap between knowing and being supported requires a system rather than a course. Table 1 summarizes five mutually reinforcing responses, the research base for each, the practices that make them work, and organizations that have documented relevant approaches. Each is then examined in turn, along with a common pitfall.
Table 1: Evidence-based responses for moving workforces from AI literacy to AI building
Response | Evidence Base | Key Practices | Illustrative Examples |
Map AI capability with skills-based architecture | Skills-based organization research; labor-market analyses of AI skills churn and premiums | Define capability rungs by role; refresh skills data frequently; link roles to learning pathways | AT&T; Amazon |
Fund and protect time for employer-supported reskilling | Global employer and worker surveys on training coverage, dosage, and adequacy | Fund certifications; schedule protected learning hours; target frontline and less-experienced staff | Walmart; Accenture |
Learn by building in the flow of work | Deliberate practice; training transfer meta-analysis; learning-in-the-flow-of-work and 70-20-10 critiques | Project-based sprints; sandboxes; skill badges tied to validated work products | PwC |
Scale through champions, communities, and managers | Transfer climate research; field evidence on tacit knowledge diffusion; psychological safety | Identify power users; embed local experts; equip managers to coach; share reusable builds | Moderna; JPMorgan Chase |
Align expectations, incentives, and pathways with outcomes | Jagged frontier experiment; career development and internal mobility evidence | Reward outcomes, not usage; make AI work visible in reviews; open internal pathways for builders | Shopify; Duolingo |
Map AI Capability With a Skills-Based Architecture
The first response is diagnostic. Organizations cannot close a gap they cannot see, and most still describe AI capability in vague terms. A skills-based architecture makes it explicit: which rung of AI capability each role needs, which people currently sit on which rung, and which pathways lead upward.
The case for skills-based approaches is promising but still developing. Deloitte's research on the skills-based organization, drawing on surveys of 1,021 workers and 225 business and HR executives across 10 countries, found that fewer than one in five organizations used skills-based approaches to a significant extent, that only 10% of HR executives said they effectively classify skills into a taxonomy, and that 79% of workers were open to organizations collecting data on their demonstrated skills (Jones et al., 2022). The same research found that only 14% of executives strongly agreed their organization used its workforce's skills to their fullest potential. Taken together, these findings suggest a large untapped opportunity and a significant data gap. For AI skills specifically, the speed of change documented by PwC (2025, 2026) and Workday (2026) makes static, annually updated taxonomies particularly ill-suited. A capability map built in early 2025 would have listed prompting as an emerging skill; by mid-2026 the market had already moved on.
Effective approaches:
Role-level capability rungs: Specify, for each role family, whether the target is literacy, fluency, building, or stewardship, rather than setting one AI standard for everyone.
Live demand signals: Refresh skill priorities using internal requisition data and external labor-market data at least quarterly, so that the map tracks shifts like the post-January 2026 decline in prompting demand.
Demonstrated rather than declared skills: Assess building capability through artifacts people have produced, such as a working automation or assistant, rather than self-ratings or course completions.
Visible pathways: Publish, for each rung, the learning resources, projects, and roles that lead to the next one, so the 79% of workers who know what they need can see how to get it.
Shared ownership: Make the capability map a joint product of HR, learning, and business leaders, not a learning-department artifact.
AT&T offers an early and well-documented illustration of a skills-led workforce overhaul. As the company shifted from cables and hardware toward internet and cloud services, it launched an initiative known as Workforce 2020 to help its roughly 280,000 employees, whose average tenure was 22 years, build new digital skills (Donovan & Benko, 2016). Writing in Harvard Business Review, AT&T's chief strategy officer and a Deloitte co-author described an approach that combined an online self-service platform, courses in new technologies, tuition support, and online master's degrees developed with Georgia Tech and Udacity, alongside changes to performance metrics and compensation. At least half the workforce had been given new roles with the expectation that they earn credentials for them. The program predates generative AI, but its architecture, a clear view of where the business was going, a mapping of roles to future skills, and a self-service route for employees to move toward them, is precisely the scaffolding that AI capability now requires.
Amazon took a tiered approach that maps neatly onto the capability ladder. In 2021 the company announced plans to invest $1.2 billion to provide free skills training to 300,000 U.S. employees by 2025, through a portfolio ranging from full college tuition for front-line employees to technical apprenticeships and Machine Learning University, an internal program taught by Amazon employees that gives staff with technical backgrounds graduate-level machine learning skills (Amazon, 2021). The point is not the size of the budget but the structure: different programs serve different starting points and destinations, which is what a role-level capability map implies.
A common pitfall is building an exhaustive skills taxonomy and then letting it freeze. Taxonomy projects can consume years, and in AI the specific skills shift within months. A lighter map that names capability rungs and is refreshed frequently is usually more useful than a detailed catalog that is out of date by the time it launches.
Fund and Protect Time for Employer-Supported Reskilling
The second response addresses the gap that Workday's survey makes most visible: workers know what they need, but too few say their employer is helping. Closing that gap is partly about money and largely about time.
The research on training coverage paints a consistent picture. The World Economic Forum (2025) reports that the share of workers completing training as part of long-term learning strategies rose to 50%, up from 41% two years earlier, and that 85% of employers plan to prioritize upskilling. Yet in the 2024 Work Trend Index only 39% of AI users had received training from their employer (Microsoft & LinkedIn, 2024), and in BCG's 2025 survey only 36% felt adequately trained (Boston Consulting Group, 2025). BCG's data also speak to dosage: employees who received five or more hours of training, especially in person and with coaching, were significantly more likely to become regular users (Boston Consulting Group, 2025). That is a modest threshold, which suggests that many programs fall short of even basic sufficiency. Building skills will require far more. The OECD (2025) adds a system-level concern: the available supply of AI training may not be keeping pace with the need even for general literacy, let alone advanced building skills.
Effective approaches:
Protected learning time: Schedule regular, manager-endorsed hours for AI practice inside the working week rather than expecting learning to happen after hours.
Funded credentials with labor-market value: Cover the cost of recognized certifications so that employees gain portable proof of capability, which signals that the investment is in them, not just in the tool.
Targeted reach: Prioritize frontline, less experienced, and lower-paid employees, who research suggests stand to benefit most and are least likely to be reached by default (Brynjolfsson et al., 2025; Boston Consulting Group, 2025).
Sufficient dosage: Treat a few hours of introduction as a floor, and plan for sustained practice toward building capability.
Transparent commitments: Publish what support the organization offers, so employees can tell the difference between real investment and exhortation.
Walmart has made one of the most explicit commitments to frontline AI capability. In September 2025, the company's chief people officer announced that Walmart would offer a free AI certification, developed with OpenAI and delivered through Walmart Academy, to U.S. frontline and office-based associates beginning in 2026, with interim AI training available through the company's Live Better U education benefit (Morris, 2025). Walmart described the certification as designed to help associates both at work and in their personal lives, and cited a nearly $1 billion commitment to skills training through 2026. The design choice matters: by offering a recognized credential at no cost, Walmart directly addresses the support side of the gap Workday documents, and by including frontline associates, it targets the population BCG found most at risk of stalling.
Accenture illustrates both the scale that is possible and the hard trade-offs that can accompany it. On its September 2025 earnings call, chief executive Julie Sweet reported that over 550,000 of the firm's employees had been trained in the fundamentals of generative AI and that its AI and data professionals had grown to 77,000 from about 40,000 two years earlier (Accenture, 2025). She described upskilling as the firm's primary talent strategy, while also stating that Accenture was exiting, on a compressed timeline, people for whom reskilling was not, in the company's experience, a viable path to the skills it needed. Leaders can read this example two ways. It shows that very large literacy programs can be delivered quickly. It also shows that literacy at scale does not automatically produce building capability, and that organizations under time pressure may turn to replacement when development pathways are too slow.
A common pitfall is treating a single, organization-wide literacy push as the end of the job. Basic training is necessary, but Workday's data show that the market has already moved past it. If funding and time stop at the first rung, the organization will have paid to teach a skill whose market value is falling, while leaving the rising skills to chance.
Learn by Building in the Flow of Work
The third response concerns how people learn the skills that are rising in demand. Building an AI tool or automating a workflow is closer to a craft than to a body of knowledge. It is learned primarily through doing, with feedback, on real problems.
Several streams of research converge here. The literature on deliberate practice, originating with Ericsson et al. (1993), argues that expert performance develops through effortful practice designed to improve specific aspects of performance, with immediate feedback. Later meta-analytic work cautions that practice explains less variance in performance than the original theory implied, particularly in professional domains (Macnamara et al., 2014), which suggests that practice is necessary but not sufficient, and that aptitude, task design, and context also matter. Training-transfer research offers a further insight. Blume et al.'s (2010) meta-analysis of 89 studies found that motivation and a supportive work environment had stronger relationships with transfer when training targeted open skills, those requiring judgment and adaptation, than closed skills such as learning a specific piece of software. AI literacy is closer to a closed skill. AI building is an open one. That implies that building capability depends heavily on what happens back on the job.
The popular 70-20-10 heuristic, which holds that most workplace learning is informal and experiential, is often cited to justify on-the-job approaches. Practitioners should use it with care. Clardy's (2018) review concluded that the evidence behind the 70% figure is weak and that the rule should be abandoned as a basis for policy. The lesson is not that experiential learning is unimportant, but that it should be designed rather than assumed. Bersin and Zao-Sanders (2019) make a similar argument for "learning in the flow of work": because the urgency of work tends to crowd out the luxury of learning, organizations need to build learning into everyday tasks, with agreed time and accessible knowledge systems.
Effective approaches:
Build sprints on real problems: Organize short, time-boxed projects in which employees build an assistant or automation for a genuine pain point in their own work.
Safe sandboxes: Provide approved environments, data, and tools where people can experiment without risking production systems or sensitive information.
Validated skill badges: Award recognition for working products that have been reviewed, not just for course completion, and separate knowledge credentials from demonstrated skill credentials.
Reuse libraries: Make it easy to share and adapt what others have built, so learning compounds across teams.
Structured reflection: Include brief retrospectives on where AI failed or misled, to build the frontier judgment that Dell'Acqua et al. (2023) showed is essential.
PwC's Digital Accelerators program is one of the most fully documented examples of learning by building at scale. Recognized in Training magazine's 2021 Hall of Fame, the program selected about 1,100 staff from 3,500 applicants for its inaugural class and grew to roughly 2,000 participants (Training Magazine, 2021). After a two-and-a-half-day onboarding, participants spent around 10 hours per week over about two years in virtual learning in areas such as intelligent process automation and AI, combined with team projects, before deploying to client and internal work. Participants could earn knowledge badges requiring about 50 hours of learning and skill badges requiring 150 to 200 hours plus validation. Crucially, what people built was shared through the firm's Digital Lab, which had accumulated more than 5,500 assets, including automations, data workflows, AI models, and visualizations, with more than 3.2 million runs and downloads; Digital Accelerators developed 56% of those assets. One participant-built validation workflow was reported to save eight hours per run and eliminate more than 150 manual steps.
PwC's design addresses each of the research insights above. It provides sustained practice over months rather than days, ties learning to real projects, validates skills through demonstrated work, and creates a shared repository that turns individual learning into organizational capability. PwC US later extended this commitment to generative AI, announcing in 2023 a $1 billion investment over three years that included plans to upskill 65,000 people on AI tools and capabilities (PwC, 2023).
A common pitfall is the hackathon that goes nowhere. One-off build events generate enthusiasm and prototypes, but without a path to deployment, a library for reuse, and recognition for maintained solutions, the prototypes die and participants learn that building is a side show. The value lies in the system around the build, not the event.
Scale Through Champions, Communities, and Managers
The fourth response recognizes that AI capability spreads socially. No central learning team can personally coach thousands of employees through building their first automation. Peers, power users, and managers can.
The evidence for social and managerial channels is strong. Blume et al. (2010) identified a supportive work environment, including supervisor and peer support, as a reliable predictor of transfer. Brynjolfsson et al. (2025) found that an AI tool could diffuse the practices of top performers to less experienced colleagues; champion networks do something analogous through people rather than software. BCG's finding that only 25% of frontline workers say their leaders provide enough AI guidance, and that adoption and optimism are markedly higher where leaders are engaged (Boston Consulting Group, 2025), underscores the role of managers. Yet LinkedIn's 2025 Workplace Learning Report found that 50% of learning and talent professionals said managers lacked proper support for career development, and that only 15% of employees said their manager had helped them build a career plan in the past six months (LinkedIn, 2025). Microsoft's 2025 survey adds that 51% of managers expect AI training or upskilling to become a key team responsibility within five years (Microsoft, 2025). Experimentation also requires psychological safety, the shared belief that a team is safe for interpersonal risk-taking, which Edmondson (1999) linked to team learning behavior. People will not show colleagues a half-working automation if doing so feels risky.
Effective approaches:
Find the power users: Identify employees who are already building, through contests, usage data, or nominations, and formalize their role as champions with time and recognition.
Embed local experts: Place champions inside business units and geographies rather than in a central team, so help is close to the work.
Equip managers to coach: Give managers enough fluency to recognize good AI use, ask useful questions, and protect learning time, rather than expecting them to be experts.
Create open forums: Run office hours and internal communities where people can share builds and failures.
Normalize imperfection: Celebrate useful experiments and lessons from failures to build the psychological safety that learning requires.
Moderna's rollout of generative AI shows how a champion model can accelerate the move from use to building. According to OpenAI's account of the partnership, Moderna set a goal of full adoption and proficiency among employees with digital access within six months, ran an AI prompt contest that identified its top 100 power users, and organized them into an internal cohort of Generative AI Champions (OpenAI, 2024). The company supported this with office hours in every business line and geography and an internal AI forum with about 2,000 active weekly participants. The building outcome is notable: within two months of adopting ChatGPT Enterprise, Moderna employees had created 750 custom GPTs, and 40% of weekly active users had created GPTs themselves. In the language of the capability ladder, a large share of users moved beyond literacy to building within weeks, with champions doing much of the teaching.
JPMorgan Chase has paired formal training with a distributed network of local experts. The bank began training new hires in prompt engineering in 2024 (Fabbro, 2024) and later rolled out its LLM Suite generative AI assistant to about 200,000 employees (Abrego & Chan, 2024). Business Insider reported that the bank offered courses and in-person sessions and relied on "superusers," described as roughly 10% to 20% of employees who were eager to help, embedded in groups as local experts. Executives described healthy competition between teams, with an AI copilot piloted in the private bank prompting the investment bank to ask to join, and the stated aim of moving beyond drafting and summarizing toward tools built into daily workflows. That aim, from generic use to embedded tools, is exactly the shift Workday's data describe.
A common pitfall is relying on volunteer enthusiasm without structure. Champions who are given a title but no time, recognition, or connection to the business soon burn out, and their help reaches only their friends. Champion networks work when they are treated as part of the operating model, with explicit capacity and visible leadership backing.
Align Expectations, Incentives, and Pathways With Outcomes
The fifth response concerns what organizations ask for and reward. Expectations shape learning. If AI capability is invisible in goals, reviews, and career decisions, employees will rationally treat it as optional. But if expectations are poorly designed, they can push behavior in the wrong direction.
The jagged frontier finding is central here. Because AI improves performance on some tasks and degrades it on others (Dell'Acqua et al., 2023), any incentive that rewards AI use for its own sake risks encouraging people to apply it where it does harm. Expectations should be framed around outcomes: better work, faster cycle times, new capabilities, and well-judged decisions about when not to use AI. Career consequences matter too. LinkedIn's 2025 research found that organizations it classified as career development champions were 42% more likely to be generative AI frontrunners and 32% more likely to be deploying AI training programs (LinkedIn, 2025). The direction of causality is unclear, but the association suggests that organizations which take careers seriously also take AI capability seriously. Pathways are where this becomes tangible. Workday's data show internal moves falling at 57% of employers (Workday, 2026), which limits the routes through which newly skilled people can apply what they have learned. Even brief, project-based internal opportunities can give builders a place to use their skills.
Effective approaches:
Outcome-based expectations: State clearly that AI capability is expected, and define success in terms of results and judgment rather than frequency of use.
Visible work in reviews: Invite employees to document what they built or automated and its impact, so building is recognized without mandating usage.
Headcount and resourcing discipline: Encourage teams to consider automation and AI-enabled redesign when planning work, while guarding against reflexive cuts that erode learning capacity.
Project-based pathways: Open short-term assignments and gigs where newly capable builders can apply skills across teams, even when formal internal moves are slow.
Feedback loops: Revisit expectations as tools and evidence change, and be willing to adjust when a policy produces unintended behavior.
Shopify made AI expectations unusually explicit. In a memo sent to staff in March 2025 and later posted publicly, chief executive Tobi Lütke described reflexive AI use as a fundamental expectation, said AI usage would factor into performance reviews, and asked teams to show why AI could not meet a need before requesting more headcount or resources (Palmer, 2025). The memo also asked teams to imagine how their area would look if autonomous AI agents were already part of the team, a prompt that points directly toward building and stewardship rather than simple use.
Duolingo's experience shows why expectations need to be tied to outcomes. In April 2025, chief executive Luis von Ahn announced that the company would become "AI-first" and that employees would be assessed on their AI use. A year later, he said the company had backtracked: AI use no longer factored into performance evaluations, because the metric was pushing people toward AI even where it did not fit their work rather than holding them accountable for results (Munis, 2026). He emphasized that performance depends on doing the job well, with AI as a tool that can help. Read together, the Shopify and Duolingo examples show the value of clear expectations and the risk of measuring the wrong thing.
A common pitfall is measuring usage rather than capability. Login counts, prompt volumes, and license utilization are easy to track and easy to game. They say little about whether someone can build a useful tool or knows when to set AI aside. Where organizations do track AI activity, it should inform support, not substitute for judgments about performance.
Building Long-Term AI Capability
The five responses address the immediate gap. Sustaining capability over years, as tools and requirements keep changing, requires deeper organizational foundations. Three pillars stand out.
Skills Intelligence as Continuous Infrastructure
The first pillar is treating skills data as infrastructure rather than as a project. The rise and fall of prompting demand within a single year (Workday, 2026), and evidence that skills in AI-exposed jobs are changing more than twice as fast as in less exposed jobs (PwC, 2026), mean that any static view of AI capability will be wrong within months. Organizations need a continuous loop: external labor-market signals and internal requisition data feeding a capability map, which drives learning offers, which generate evidence of demonstrated skills, which in turn inform staffing and career decisions. Deloitte's finding that only 10% of HR executives say they effectively classify skills (Jones et al., 2022) suggests most organizations are far from this. The goal is not perfection but responsiveness. A good test is whether the organization would have noticed, within a quarter, that demand for prompting had peaked and demand for building had taken off. If the answer is no, the skills infrastructure is not yet doing its job.
This infrastructure should also be candid about what it does not know. Many of the most widely cited statistics on AI skills come from vendor and consulting surveys whose samples and methods vary. Skills intelligence works best when it triangulates across sources, including the organization's own data on which learning offers actually produce deployed tools and better outcomes.
Rebuilding the Teaching Organization
The second pillar is restoring and strengthening the people who teach. Workday's finding that mentions of management and leadership skills fell 7% and training skills fell 13% in job requisitions (Workday, 2026) should concern any leader who expects the workforce to reskill rapidly. The research on training transfer is unambiguous that the work environment, especially manager and peer support, shapes whether learning sticks (Blume et al., 2010). The examples from Moderna and JPMorgan Chase show that much of the teaching of AI building happens through champions and local experts (Abrego & Chan, 2024; OpenAI, 2024).
A teaching organization makes coaching and knowledge-sharing part of people's jobs rather than extra-curricular favors. It includes AI coaching in manager expectations and development, gives champions formal capacity, and treats the people who build reusable assets for others, as PwC's Digital Lab contributors did (Training Magazine, 2021), as multipliers worth recognizing. Microsoft's finding that about a third of leaders and managers are considering AI trainer roles (Microsoft, 2025) suggests some organizations are beginning to formalize this. The risk is that, under cost pressure, organizations do the opposite, and cut the capability that makes every other investment in AI pay off.
Judgment, Stewardship, and Learning to Learn
The third pillar is cultivating the capabilities that keep their value as specific tools come and go. If prompting can rise and fall within a year, so might any particular building technique. What persists is judgment about where AI helps and where it harms (Dell'Acqua et al., 2023), the ability to redesign work rather than simply speed it up, and the habit of continuous learning. Workday's spokesperson framed the question as whether people can learn fast enough to solve the next problem (Workday, 2026), and that framing applies to organizations as much as to individuals.
Stewardship, the top rung of the capability ladder in Figure 1, deserves deliberate development. It includes governance of risk, ethical and quality standards for AI-built tools, and the ability to develop others. Organizations can grow stewardship by rotating experienced builders into roles where they review, maintain, and improve tools built by others, by including AI governance in leadership development, and by explicitly valuing the mentoring of less experienced colleagues. Over time, this produces a workforce that is not merely trained on today's tools but capable of adapting to tomorrow's.
Conclusion
The AI skills market is telling leaders something important. The capability that most organizations spent 2024 and 2025 teaching, basic use and simple prompting, is losing its distinctiveness. Demand is moving toward people who can build: who can turn a recurring task into an automation, wire an assistant into a team's work, and know when the tool is out of its depth. Workday's October 2026 data capture this shift sharply, with demand for simple prompting falling 25% from its January 2026 peak and demand for hands-on building skills rising 51% (Workday, 2026). The same data reveal the organizational weak point: 79% of workers know what skills they need, but only 66% say their employer helps them develop those skills.
Closing that gap is achievable, but not through the methods that built basic literacy. The research on training transfer, deliberate practice, and AI productivity points toward learning systems that combine clear capability maps, funded and protected time, real building projects, social learning through champions and managers, and expectations tied to outcomes. The organizations profiled here, from AT&T and Amazon to Walmart, PwC, Moderna, and JPMorgan Chase, did not succeed by running more webinars. They redesigned how learning connects to work, and in the most instructive cases, Duolingo among them, they adjusted when early choices produced the wrong behavior.
Leaders who expect AI to help them get more from their current workforce are making a reasonable bet. But it is a bet on development, and development requires investment.
Retire literacy as the finish line: Treat basic AI training as a floor and set explicit, role-specific targets for fluency, building, and stewardship.
Close the support gap visibly: Fund credentials and protect learning time, prioritizing frontline and less experienced employees who have the most to gain.
Teach building by building: Use real projects, safe sandboxes, and validated skill badges, and create libraries so that each build teaches the next person.
Invest in the teachers: Formalize champion roles, equip managers to coach AI work, and resist the drift away from management and training skills.
Reward outcomes, not usage: Recognize what people build and the judgment they show, and avoid metrics that push AI into places where it does harm.
Keep the skills map live: Refresh capability priorities at least quarterly using internal and external demand signals.
Open pathways for builders: Provide project-based opportunities where newly skilled people can apply what they have learned, even when formal internal moves are slow.
Research Infographic
References
Abrego, M., & Chan, B. (2024, November 22). JPMorgan's AI rollout: Jamie Dimon's a "tremendous" user and it's caused some "healthy competition" among teams. Business Insider.
Accenture. (2025, September 25). Fourth quarter fiscal 2025 financial results: Conference call transcript. Accenture plc.
Amazon. (2021, September 9). Amazon to pay full college tuition for its front-line employees as part of $1.2 billion investment in education and skills training for U.S. workforce [Press release].
Baldwin, T. T., & Ford, J. K. (1988). Transfer of training: A review and directions for future research. Personnel Psychology, 41(1), 63–105.
Bersin, J., & Zao-Sanders, M. (2019, February 19). Making learning a part of everyday work. Harvard Business Review.
Blume, B. D., Ford, J. K., Baldwin, T. T., & Huang, J. L. (2010). Transfer of training: A meta-analytic review. Journal of Management, 36(4), 1065–1105.
Boston Consulting Group. (2025). AI at work 2025: Momentum builds, but gaps remain. Boston Consulting Group.
Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942.
Clardy, A. (2018). 70-20-10 and the dominance of informal learning: A fact in search of evidence. Human Resource Development Review, 17(2), 153–178.
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.
Donovan, J., & Benko, C. (2016). AT&T's talent overhaul. Harvard Business Review, 94(10), 68–73.
Edmondson, A. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383.
Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363–406.
Fabbro, R. (2024, May 20). JPMorgan Chase is training every new hire on AI. Quartz.
Jones, R., Cantrell, S., Griffiths, M., & Hiipakka, J. (2022). The skills-based organization: A new operating model for work and the workforce. Deloitte Insights.
Lightcast. (2025, July 23). New Lightcast report: AI skills command 28% salary premium as demand shifts beyond tech industry [Press release].
LinkedIn. (2025). 2025 workplace learning report: The rise of career champions. LinkedIn Learning.
Macnamara, B. N., Hambrick, D. Z., & Oswald, F. L. (2014). Deliberate practice and performance in music, games, sports, education, and professions: A meta-analysis. Psychological Science, 25(8), 1608–1618.
Microsoft. (2025). 2025 Work Trend Index annual report: 2025: The year the Frontier Firm is born. Microsoft WorkLab.
Microsoft & LinkedIn. (2024). 2024 Work Trend Index annual report: AI at work is here. Now comes the hard part. Microsoft WorkLab.
Morris, D. (2025, September 4). Donna Morris to associates: Being people-led means investing in you. Walmart.
Munis, J. (2026, April 13). ‘I’m not going to force you’: Duolingo CEO backs off from evaluating employees on their AI usage. Fortune.
Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192.
OECD. (2025). Bridging the AI skills gap: Is training keeping up? OECD Publishing.
OpenAI. (2024). Moderna: Accelerating the development of life-saving treatments [Customer story]. OpenAI.
Palmer, A. (2025, April 7). Shopify CEO says staffers need to prove jobs can't be done by AI before asking for more headcount. CNBC.
PwC. (2023, April 26). PwC US makes $1 billion investment to expand and scale AI capabilities [Press release].
PwC. (2025, June 3). AI linked to a fourfold increase in productivity growth and 56% wage premium, while jobs grow even in the most easily automated roles [Press release].
PwC. (2026). Two futures for jobs in an AI era: 2026 Global AI Jobs Barometer. PwC.
Training Magazine. (2021, October 12). Training Top 100 Hall of Fame outstanding training initiatives: PwC's Digital Accelerators program. Training.
Workday. (2026, October 5). Workday Global Workforce Report: AI is rewriting jobs more than it's cutting them [Press release].
World Economic Forum. (2025). The future of jobs report 2025. World Economic Forum.
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). Past the Prompt: Why AI Skill Demand Is Shifting From Using to Building, and How Leaders Can Close the Skills-Support Gap. Human Capital Leadership Review, 40(1). doi.org/10.70175/hclreview.2020.40.1.1
