AI is everywhere in the headlines. It is not yet everywhere in people’s working lives.

A strange gap has opened in the economy. On one side, artificial intelligence is being described as the next great career accelerant: a tool that can help people write, code, analyze, design, summarize, sell, teach, research, and automate parts of their day. On the other side, millions of workers still are not using these tools in any meaningful way — not because they are lazy, incurious, or resistant to change, but because access is uneven, training is thin, workplace rules are confusing, and the benefits are distributed through existing inequalities.

The numbers show the gap clearly. In September 2025, Pew Research Center found that 21% of U.S. workers said at least some of their work was done with AI, up from 16% roughly a year earlier. But 65% still said they did not use AI much or at all in their jobs. Even among workers who were not using AI, more than a third said at least some of their work could be done with it. In other words, AI is not simply irrelevant to most workers. Many people are standing near the technology, but not yet able to use it well.

That matters because the upside is real, even if it is often oversold. In a large customer-support study published in The Quarterly Journal of Economics, access to a generative AI assistant increased productivity by 15% on average, with especially large gains for less-experienced and lower-skilled workers. In another widely cited field experiment with consultants, researchers working with BCG found that GPT-4 improved speed and quality on tasks inside the model’s capability frontier, while also showing that AI could hurt performance when people used it for tasks where the tool was unreliable.

So the issue is not whether AI can help some people do some jobs better. The issue is whether ordinary workers can actually reach the tools, understand them, trust them, and apply them safely to real career problems. Right now, for many people, the answer is still no.

The first barrier is not intelligence. It is access.

Public discussion often assumes that AI tools are available to anyone with curiosity and an internet connection. That assumption hides a lot. AI tools require devices, reliable broadband, accounts, sometimes paid subscriptions, and often the ability to experiment without fear of breaking a rule or exposing sensitive information.

Even in the United States, internet access remains unequal. Pew reported in January 2026 that about eight in ten U.S. adults subscribe to home broadband, but the divide by income remains persistent: adults in households earning under $30,000 are far less likely than those with higher incomes to have broadband at home.

That matters because AI is not a lightweight technology in practice. A worker trying to use AI for résumé rewriting, spreadsheet analysis, coding practice, interview preparation, or portfolio building needs more than occasional phone access. They need stable connectivity, a device that can handle multitasking, and enough privacy and time to experiment. A person using a shared phone on a limited data plan is not starting from the same place as a salaried professional with a laptop, enterprise AI license, and paid time to learn.

The global divide is even sharper. The IMF’s AI Preparedness Index evaluates economies across digital infrastructure, human capital and labor-market policy, innovation and economic integration, and regulation. The IMF found that wealthier economies tend to be better equipped for AI adoption than low-income countries.

This means “learn AI” is not a neutral instruction. For some workers, it means opening a company-approved assistant inside tools they already use. For others, it means finding a free chatbot after a long shift, on a phone, with no guidance, no data privacy support, and no clear connection to a promotion.

The second barrier is training — and the training market is not keeping up.

AI tools look deceptively simple. A chat box invites anyone to type a question. But career-enhancing use requires judgment: how to frame a problem, how to check outputs, how to protect data, how to decide when not to use AI, and how to turn a model’s answer into actual work.

The OECD warned in 2025 that demand for both specialized AI professionals and general AI understanding is rising, but only a small percentage of training courses in the countries it studied currently deliver AI content. The OECD also noted that governments are creating publicly funded AI training programs, but more needs to be done to make adult training suitable for an AI-driven workplace.

This is one of the most important reasons people cannot simply “get their hands on AI” and use it to advance. Access to the tool is not the same as access to the skill. A worker may know ChatGPT, Copilot, Gemini, Claude, or another system exists. They may even have tried one. But a few experiments do not automatically translate into better work, better pay, or a stronger career path.

Workers themselves appear to understand this. A 2025 Salesforce and Morning Consult survey found that 64% of workers globally supported more investment in general skills, while 53% specifically wanted AI-related training. Nearly two-thirds said they would likely take AI training if governments offered discounts or financial support.

The desire to learn is there. The structured support often is not.

The third barrier is that companies buy tools faster than they redesign work.

At the organizational level, AI adoption looks impressive. Stanford’s 2026 AI Index, drawing on McKinsey survey data, reported that in 2025 88% of respondents said their organizations used AI in at least one business function, and 79% said their organizations regularly used generative AI in at least one business function.

But organizational adoption is not the same as worker empowerment. A company can use AI in marketing analytics, software engineering, customer service, HR screening, or supply-chain forecasting without giving every employee meaningful access. AI may be concentrated in specific teams, restricted to managers, or embedded invisibly into enterprise systems. Workers may be affected by AI without being trained to use it.

McKinsey’s 2025 global survey found that companies are experimenting with and scaling AI agents, but also that only about one-third of respondents said their organizations were scaling AI programs across the enterprise. Larger companies were more likely than smaller ones to be at the scaling stage.

That leaves many workers in a half-adopted world. Leadership talks about AI transformation. Employees hear that AI skills matter. But the daily workflow remains unchanged, or the approved tool is hard to access, or the use cases are vague. The result is a workplace where AI is simultaneously hyped and inaccessible.

Small and medium-size businesses face a different version of the same problem. The OECD found that generative AI can help SMEs compensate for skill gaps and labor shortages; among SMEs that used generative AI and had experienced a skill gap, 39% said it helped compensate for that gap. Yet SMEs often have fewer resources for training, governance, and technical support.

So workers in smaller firms may be close to business problems where AI would help, but far from the formal systems that make AI safe and useful.

The fourth barrier is fear — and it is rational.

Many workers are not merely confused about AI. They are worried about what it means for their jobs.

Pew’s 2025 report found that 52% of U.S. workers said they were worried about future workplace AI use, while only 36% said they felt hopeful. Pew also found that only 6% thought workplace AI would lead to more job opportunities for them in the long run, while 32% thought it would lead to fewer opportunities. Lower- and middle-income workers were more likely than upper-income workers to say AI would reduce their job opportunities.

That fear can suppress adoption. If a worker believes that showing AI can do part of their job will make them easier to replace, they may avoid experimenting openly. If a manager frames AI as a cost-cutting tool rather than a capability-building tool, employees may treat it as a threat. If a company rewards output but not learning, workers may not risk the temporary slowdown that comes with learning a new system.

The fear is not irrational because the labor market is changing. The World Economic Forum’s Future of Jobs Report 2025 projected that 39% of workers’ existing skill sets would be transformed or become outdated over the 2025–2030 period. It also identified AI and big data, networks and cybersecurity, and technological literacy among the fastest-growing skill areas.

When workers hear that message without support, the career advice sounds less like “AI will help you grow” and more like “keep up or be left behind.”

The fifth barrier is unclear policy.

Many employees do not know what they are allowed to put into AI systems. Can they paste a customer email into a chatbot? Can they summarize a legal document? Can they use AI to draft code? Can they upload a spreadsheet? Can they use a personal account if the company has not provided a tool?

These are not trivial questions. Cisco’s 2025 Data Privacy Benchmark Study found that 64% of respondents worried about inadvertently sharing sensitive information publicly or with competitors, yet nearly half admitted to putting personal employee data or non-public company data into generative AI tools.

Security researchers and enterprise leaders now use the term “shadow AI” to describe workers using unapproved AI tools outside official oversight. IBM’s 2025 Cost of a Data Breach research found that high levels of shadow AI added $670,000 to the global average cost of a breach.

This creates a paradox. If companies ban AI broadly, workers lose access to tools that could help them. If companies provide no guidance, workers may use unsafe tools anyway. If policies are too vague, cautious employees avoid AI while bolder employees take hidden risks. None of those outcomes creates broad career mobility.

The solution is not simply more restriction. It is clear permission, safe tooling, practical examples, and role-specific training. Workers need to know not only what is forbidden, but what is encouraged.

The sixth barrier is that AI is unevenly useful.

One reason AI adoption feels confusing is that the technology is not equally good at everything. It can draft a decent email, summarize a meeting, generate code, brainstorm interview questions, analyze text, and help structure a project plan. It can also fabricate facts, miss context, reproduce bias, leak sensitive information if used carelessly, or produce confident nonsense.

The BCG-Harvard research on the “jagged technological frontier” is useful here. The study found that AI improved performance on tasks within the frontier, but could reduce performance when used for tasks outside it.

For workers, this means AI literacy is not just “prompt engineering.” It is task judgment. The valuable skill is knowing when AI is a calculator, when it is a brainstorming partner, when it is a drafting assistant, when it is a search companion, and when it is a liability.

That is hard to learn alone. It is also hard to learn from generic tutorials. A nurse, paralegal, salesperson, accountant, mechanic, teacher, project manager, and warehouse supervisor do not need the same AI playbook. They need examples connected to their work, their risks, and their career ladder.

The seventh barrier is time.

Learning AI takes time, and time is unequally distributed. Professionals with flexible schedules may experiment during work hours. Hourly employees may be expected to learn after work. Job seekers may be told to upskill while also applying for jobs, caring for family, or working multiple shifts.

This is why employer-led training matters. If AI skills are becoming part of work, then learning them should be treated as work. When training is pushed entirely onto individuals, the people with the least spare time fall further behind.

The World Economic Forum estimated that if the global workforce were represented by 100 people, 59 would need reskilling or upskilling by 2030, and 11 would be unlikely to receive it.

That is the shape of the problem: not a lack of motivation, but a mismatch between the scale of change and the systems available to help people adapt.

The eighth barrier is that career benefits are not automatic.

Even when people use AI, it does not automatically improve their careers. A worker may become faster without being promoted. A job seeker may create a better résumé without gaining access to stronger networks. A freelancer may produce more drafts while clients lower rates because they assume AI made the work easy.

This distinction matters. AI can improve tasks. Careers improve through recognition, opportunity, credentials, bargaining power, and trust. If organizations capture all the productivity gains while workers absorb the learning burden, AI may enhance output without enhancing careers.

That is why AI access has to be connected to advancement. Companies should define what AI fluency means for different roles, include it in promotion pathways, and reward workers who use AI responsibly to improve outcomes. Schools, workforce boards, unions, libraries, and community colleges should connect AI training to specific labor-market opportunities, not just abstract digital literacy.

So what would real access look like?

Real access would start with infrastructure: reliable broadband, devices, and affordable tools. It would continue with safe enterprise access, so workers do not have to choose between falling behind and using unapproved systems. It would include paid learning time, role-specific examples, and clear policies about data, confidentiality, and acceptable use.

It would also include honest communication. Workers deserve to know whether AI is being introduced to augment their roles, reduce headcount, change performance expectations, or create new internal opportunities. Trust is a prerequisite for adoption.

Finally, real access would treat AI literacy as a public workforce issue, not a private hobby. The OECD’s research points to the need for scaled-up AI training supply and more inclusive policy tools, including career guidance, public-private collaboration, and train-the-trainer programs.

That is the missing layer in much of today’s AI conversation. The technology is moving quickly, but the institutions that help people absorb technology — schools, employers, workforce systems, unions, libraries, professional associations, and government programs — move more slowly.

The bottom line

People are not failing to use AI because they lack ambition. Many are blocked by cost, connectivity, unclear rules, weak training, fear of job loss, lack of time, and workplaces that adopt AI at the top while leaving employees to figure it out at the edges.

The career promise of AI is real, but it is not self-executing. A tool does not become empowering just because it exists. It becomes empowering when people can access it, understand it, trust it, practice with it, and use it in environments where their gains are recognized.

If we want AI to enhance careers broadly, the question should not be, “Why don’t workers just learn AI?”

The better question is: What would have to change so that every worker has a fair chance to use AI well?

Amit Shrivastava