Also check out book Indispensable: Your Career Guide for the Age of AI, with detailed predictions for 100+ US professions, tips on how to use AI, and which human skills to work on.

1. Don’t be overwhelmed — your job is safe for now

This is the most important point I want to make — we are seeing Jevons’ Paradox play out in real time in software development. Although it appears as though engineering and PM jobs will reduce, they will likely go up.

Everyone predicted ATMs would kill bank teller jobs. In 1973, the New York Times said up to 75% of tellers would be replaced. ATMs did reduce tellers by 37% per branch — but overall, teller jobs actually doubled over the next 30 years, growing faster than the general labor market. Why? Cheaper branches meant banks expanded into underserved areas, and they still needed people to do what ATMs couldn’t — answer questions, advise customers, build relationships. The pattern: technology automates some tasks → lower cost, more accessible → demand expands → the human role evolves upwards.

The same pattern is playing out in software development right now. When ChatGPT launched in 2023, everyone predicted programming was dead. The reality? Software engineering and PM jobs have gone up. AI is writing more code than predicted, but demand for AI and software has also surged.

Adoption is still under 50% in most industries — so it’s probably looking quite safe from a five-year perspective. The overwhelm comes not from whether you’ll have a job, but from the fact that your job itself is transforming.

2. Knowing when NOT to use AI is the most important skill.

There is a misconception that if you plaster AI onto everything, it will make you more productive. That is not true. Research from the brilliant Ethan Mollick at Wharton shows that when AI is good at a task and you have the expertise to steer it, you get great results — 12.2% more tasks completed, 25.1% faster, at ~40% higher quality. But when you ask AI to do something it’s not good at and you lack the expertise to steer it, you produce worse output than had you not used AI at all.

The researchers called this “falling asleep at the wheel.”

Worse still: if you delegate tasks to AI and don’t steer the ship, it will atrophy your own skills. You’ll become worse at the things you already know how to do.

The biggest skill right now is knowing where to offload to AI and where not to.

3. To be an AI-Powered PM, break your job into tasks and categorize them by AI capability.

The research on how jobs evolve with AI converges on a clear framework: jobs are bundles of tasks (Brynjolfsson, Mitchell & Rock, 2018).

Break them down, categorize each by AI augmentability — tasks can be AI-led, AI-human collaboration, or human-led (Anand & Wu, 2025, Harvard Business Review).

The tasks that remain human are high on empathy, presence, opinion, creativity, and hope (Loaiza & Rigobon, 2024/2025, MIT Sloan).

I did this for the PM role — 102 tasks across 4 buckets:

Bucket 1: AI-Executed, Human-Verified (13 tasks). AI runs it end-to-end. You verify, catch errors, handle exceptions. Very few tasks belong here. Be careful — this is where most people start, and it’s where AI makes the most incorrect assumptions. Example: performance tracking — let AI pull telemetry, build dashboards, analyze trends. You interpret and decide.

Bucket 2: Human-Led, AI-Assisted (29 tasks). You lead; AI handles subtasks. You decide, AI drafts/gathers/organizes. Example: customer feedback — the quality comes from the questions you ask, not the processing. Ask the right questions, read between the lines, then let AI aggregate and categorize.

Bucket 3: Human-AI Fusion (22 tasks). Tight human-AI loops. You initiate → AI extends → you refine → repeat. Example: prototypes — if you just hand it to AI, it won’t respect your company’s design constraints or product direction. You need to be extremely precise, set the direction, and mold the output. It may take more time, but the result is better than without AI.

Bucket 4: Human-Led, AI-Minimal (38 tasks). You own it completely. This is the most important bucket. Most PM tasks live here — landing tradeoffs, getting people aligned, driving hard choices. These are the tasks that make you indispensable. Don’t use AI for this. Resist the urge. If someone tells you there’s an AI agent that can get consensus in a meeting, it’s a horrible idea.

Here is the full cheat sheet - everything a PM does, broken out by bucket:

4. Soft skills are the hard skills now.

Here are 4 thumb rules-

Integrate Deeply: Engineers, designers, and researchers all have informed judgment now. Everyone is evolving in their roles. Your biggest role as a PM is to be the integrator — listen to every perspective, synthesize them. That’s what makes you the best PM.

Drive with Agency: Your fire and ability to get things out of people is a superpower. Nobody, including the AI, wants to do this. Ask the right questions, engage people and AI with the right prompts, cultivate fire in people to get business results.

Know What AI Can and Cannot Do. See above. Continuously adapt your workflow as new tools arrive. But the biggest skill is knowing when to NOT apply AI. Develop that judgment.

Stay Hungry, Stay Foolish. The career ladder is dead. People are leaving VP-level roles for IC roles because skills, depth of expertise, and judgment are the real assets now. Keep upskilling. Keep learning. Keep showing your impact.

AI is bringing the biggest transformation in work of our time — and I’m trying to help everyone prepare for the shift. I’ve written a book, Indispensable: Your Career Guide for the Age of AI, with detailed predictions for 100+ US professions, tips on how to use AI, and which human skills to work on. I’ve also built a free career tool to help you explore how your specific occupation will evolve. Order the book and check it out here.

Research References

1. Brynjolfsson, Mitchell & Rock (2018), “What Can Machines Learn, and What Does It Mean for Occupations and the Economy?”

2. Anand & Wu (2025), “The Gen AI Playbook for Organizations” — Harvard Business Review

3. Loaiza & Rigobon (2024/2025), “The EPOCH of AI: Human-Machine Complementarities” — MIT Sloan

4. Mollick et al. (2023), “Centaurs and Cyborgs on the Jagged Frontier” — Wharton / One Useful Thing

5. Dell’Acqua et al. (2023), “Navigating the Jagged Technological Frontier” — Harvard Business School

In 1969, Chemical Bank installed America’s first ATM in Rockville Center, New York—a hulking machine that could dispense cash automatically, no human teller required. Bank executives looked at this gleaming robot and saw the future: branches staffed by machines, not people; customers getting their money without waiting in line; and most importantly, massive savings on labor costs as teller positions disappeared. The New York Times reported in 1973 that ATMs would eliminate “up to 75 percent” of teller jobs. It seemed obvious—why would banks employ humans to hand out cash when a machine could do it faster, cheaper, and twenty-four hours a day?

They were spectacularly wrong.

Between 1985 and 2002, as the number of ATMs in America exploded from 60,000 to 352,000—nearly six times as many—the number of bank tellers didn’t fall. It rose. From 485,000 tellers to 527,000 tellers. The robots designed specifically to replace these workers had spread across every bank in every city, and yet teller jobs grew slightly faster than the overall labor force. Even President Obama got it wrong decades later, citing ATMs in 2011 as a clear example of “technology displacing labor.”

Compare this to travel agents. In 2000, approximately 124,000 Americans worked as travel agents, booking flights and hotels for clients through established reservation systems. Then came Expedia, Travelocity, and a dozen other online booking sites that let customers do exactly what travel agents did—search flights, compare prices, book reservations—with just a few clicks. By 2014, travel agent employment had plummeted to 74,000, a forty percent decline. The technology could do the job, and nothing stopped it from doing so at scale.

Why did one job survive while the other vanished? Both faced automation that could perform their core tasks; both saw that automation deployed widely; both were told their jobs were doomed. The answer reveals everything about how to predict which jobs AI will actually displace versus which will merely transform—and that answer has nothing to do with how capable the AI is.

A Framework Emerges from AI Adoption Reality

After analyzing over 50 enterprise AI deployments across industries, a pattern became clear: AI adoption in the workplace is happening far slower than both market projections and technical capabilities would suggest. The conventional explanations—”change management challenges,” “cultural resistance,” “digital transformation maturity”—capture something real but offer little analytical precision. These concepts don’t help predict when AI will transform specific operations, nor do they help workers or organizations assess genuine risk and opportunity in their contexts.

Most AI job impact research focuses on exposure—measuring which tasks AI could theoretically perform. OpenAI’s influential “GPTs are GPTs” study found that 80 percent of the U.S. workforce could have at least 10 percent of their tasks affected by large language models, with 19 percent potentially seeing half their tasks impacted. GitHub’s Copilot research demonstrated significant productivity gains for software developers. These exposure studies are valuable—they show where AI could have impact—but they make a critical error: they assume technical capability predicts actual deployment.

The reality is different. Exposure measures technical potential; adoption depends on overcoming specific, identifiable barriers. An occupation might score high on AI exposure yet face glacial adoption due to regulatory frameworks, implementation costs, or error intolerance. Conversely, occupations with moderate exposure but low barriers might see rapid transformation.

What emerged from deployment analysis is a structured framework identifying five distinct barriers that determine whether AI deployment happens rapidly, slowly, or not at all. These five barriers to enterprise AI adoption propose that job displacement operates not through simple technical substitution, but through identifiable mechanisms that either clear the path for transformation or create persistent friction that slows or prevents adoption altogether. The five barriers—Policy, Resources, Inertia, Market acceptance, and Error tolerance—provide a clearer way to predict how AI will actually reshape work in specific contexts.

This paper presents these barriers, grounds them in existing research on technology adoption and workforce transformation, and demonstrates their application across different occupations.

Theoretical Foundation

The framework builds on established technology adoption research. Rogers’ Diffusion of Innovations (1962) identified factors like relative advantage, compatibility, and complexity that influence how innovations spread through social systems. Davis’s Technology Acceptance Model (1989) showed that perceived usefulness and ease of use drive individual technology acceptance. Acemoglu and Restrepo’s task-based framework (2019) demonstrated that automation creates displacement effects counterbalanced by productivity gains and new task creation.

Recent empirical work validates the importance of deployment barriers. MIT CSAIL’s 2024 study found that only 23 percent of worker compensation exposed to AI computer vision would be cost-effective to automate due to large upfront costs. The World Economic Forum’s 2025 Future of Jobs Report found that 63 percent of employers identify skills gaps and organizational culture as primary barriers to transformation—up from 60 percent in 2023.

This framework addresses a critical gap: while existing research explains technology diffusion patterns, user acceptance, and post-adoption labor market effects, no framework systematically predicts when technically feasible AI will actually be deployed at enterprise scale. The five barriers presented here—Policy, Resources, Inertia, Market acceptance, and Error tolerance—operationalize the specific mechanisms that determine whether AI moves from demonstration to production deployment.

The Five Barriers to Enterprise AI Adoption

The framework identifies five categories of barriers—Policy, Resources, Inertia, Market acceptance, and Error tolerance (forming the acronym PRIME). Each operates independently, but their combined effect determines deployment velocity and ultimate impact on work:

Barrier 1: Policy (Regulatory Barriers)

Definition: Policy barriers comprise the formal legal, compliance, and institutional requirements that govern AI deployment in specific industries or contexts. These include licensing requirements, liability frameworks, data governance rules, sector-specific regulations, and the pace at which regulatory bodies adapt to technological change.

The first barrier concerns the regulatory environment surrounding specific types of work. Some industries face minimal regulation; others operate under complex legal frameworks that presume human decision-making and create substantial friction for AI deployment even when the technology is ready.

Autonomous trucking provides a clear example. The technology for self-driving trucks on highways has reached impressive maturity; several companies have demonstrated safe long-haul autonomous freight transport. But deployment faces a labyrinth of regulatory complexity: federal regulations governing commercial vehicles were written assuming human drivers; Commercial Driver’s License requirements and Department of Transportation oversight presume human operators; state-by-state regulatory variations create a patchwork where autonomous trucks might be legal in one state but prohibited in the next; many states explicitly require human drivers by law or are actively debating autonomous vehicle rules with no consensus.

This regulatory fragmentation prevents deployment at scale regardless of technical readiness. A freight company cannot operate autonomous trucks commercially when regulations haven’t caught up to technology, when liability frameworks remain unclear, and when the legal landscape varies dramatically across jurisdictions. Complete elimination of truck driver jobs probably requires decades, not years, because regulations need to evolve, infrastructure must be built, and society needs to become comfortable with 80,000-pound autonomous vehicles sharing highways with human-driven cars.

Barrier 2: Resources (Implementation Costs)

Definition: Resource barriers represent the total economic burden of deploying AI systems at production scale, including not only direct technology expenses but also required infrastructure investments, integration complexity, ongoing maintenance requirements, and organizational adaptation costs necessary to operationalize the system within existing workflows.

The second barrier concerns how expensive it is to actually deploy AI technology in real-world production environments—not in demonstrations or pilot programs, but at scale across operations with all their complexity and established infrastructure. Consider a scenario where a coffee shop in India could theoretically replace a human barista with a perfectly functioning humanoid robot capable of making coffee, handling orders, and managing transactions. On paper, the robot appears economically superior: it works continuously without rest, maintains consistent quality, and requires only electricity and internet connectivity after initial investment.

But here’s what the simple analysis misses: the coffee shop would need reliable high-speed internet to support the robot’s AI systems; climate-controlled environment with air conditioning to prevent hardware failures in India’s heat; potentially a dedicated server infrastructure with additional cooling; regular technical servicing and maintenance for sophisticated robotics; and substantial AI computing costs for every customer interaction. All of that infrastructure investment would far exceed the cost of hiring a human barista in a context where labor costs are low and the necessary supporting infrastructure doesn’t already exist. The same deployment might make economic sense in a different country where coffee shops already have robust climate control and connectivity, but implementation costs are context-dependent, varying by geography, industry, and existing infrastructure.

MIT CSAIL research in 2024 found that only 23 percent of worker compensation exposed to AI computer vision would be cost-effective for firms to automate, with large upfront costs of AI systems making most tasks economically unattractive to automate. This empirical finding validates what deployment analysis reveals: technical capability and economic viability often diverge significantly.

Barrier 3: Inertia (Industry Adaptation Speed)

Definition: Inertia describes the rate at which specific sectors integrate new technologies into standard practice, influenced by organizational culture, competitive dynamics, institutional resistance, workforce skill levels, and the consequences of rapid versus cautious adoption. This barrier determines how quickly technically feasible and economically viable AI solutions actually diffuse through an industry.

The third barrier addresses how quickly specific industries actually integrate new technologies once other barriers are overcome. Some industries move with remarkable speed; others are structurally conservative and slow to change regardless of technology readiness.

K-12 education exemplifies high inertia. The sector faces teacher resistance to new tools, inadequate training infrastructure, severe budget constraints, and institutional conservatism rooted in the high stakes of educating children. Technology advances faster than classroom adoption; tools that could improve learning sit unused because implementation requires changing how thousands of teachers work, retraining entire workforces, and convincing conservative school boards to invest in systems that challenge traditional pedagogy. Education is not unique in this—healthcare, government, and legal services all show similar patterns of slow, cautious integration.

Contrast this with the technology industry itself, which adopts innovations almost instantly. Tech companies value speed, embrace experimentation, accept rapid change as normal, and face intense competitive pressure to integrate productivity-enhancing tools before competitors do. The “move fast” culture means AI coding assistants went from novel experiment to industry standard in roughly eighteen months. Change management and cultural resistance—the fuzzy concepts that often explain slow adoption—fit squarely into this barrier: they describe industries where adaptation is structurally slow regardless of other factors.

Barrier 4: Market (Customer Experience Preference)

Definition: Market barriers encompass both consumer willingness to engage with AI-mediated services and stakeholder comfort with AI systems replacing or augmenting human judgment in specific contexts. This barrier operates through trust dynamics, preference for human interaction, and perceived value of human expertise beyond pure task performance.

The fourth barrier addresses whether customers, clients, or other stakeholders actually want to interact with AI systems instead of humans, even when the AI performs the task competently. This operates at multiple levels: end consumers may prefer human service for relationship-intensive work; business clients may hesitate to trust AI recommendations for high-stakes decisions; and employees themselves may resist systems that change how they work.

Consider mental health therapy. AI chatbots and therapeutic support systems have demonstrated clinical effectiveness for certain types of therapeutic interventions, particularly for cognitive behavioral therapy and anxiety management. The technology works; studies show positive outcomes. But market acceptance remains mixed at best—many people seeking mental health support specifically want human empathy, the feeling of being understood by another person, and the trust that comes from human-to-human connection. The technology’s capability isn’t the constraint; willingness to use it is.

This barrier operates differently in business-to-business contexts. Research on robo-advisors found that older adults showed very limited uptake because they were less likely to trust the technology, and greater awareness of automation correlated with reduced organizational commitment and higher turnover intentions among employees. Market acceptance isn’t monolithic—it varies by demographic, context, and the specific human needs being addressed.

Barrier 5: Error (Mistake Tolerance)

Definition: Error acceptance describes the tolerance for mistakes within a specific operational context, determined by the severity of consequences from system failures, liability frameworks governing responsibility for AI errors, and the reversibility of decisions made by or with AI assistance. Low error acceptance creates high barriers to adoption regardless of average system accuracy.

The third barrier examines how much tolerance exists for AI mistakes in specific contexts. Some domains accept errors as part of normal operation; others cannot tolerate even rare failures. This fundamentally shapes deployment timelines regardless of AI capability.

Healthcare illustrates low error tolerance. An AI diagnostic system might achieve 98 percent accuracy—substantially better than human diagnostic error rates in many contexts—but that 2 percent failure rate carries catastrophic consequences. You cannot tell a patient’s family “the AI is right ninety-eight percent of the time” when their loved one died because of the two percent error. Medical malpractice liability, ethical obligations, and the fundamental reality that healthcare involves life-and-death stakes all create extremely low error tolerance that slows AI adoption regardless of capability. The system must approach near-perfect performance, undergo extensive clinical validation, and operate under careful human oversight before wide deployment becomes acceptable.

Compare this to content creation, marketing, or business communication, where errors are inconvenient but rarely catastrophic. If AI writes marketing copy and makes a mistake, you fix it and republish; the worst case is minor embarrassment and correction effort. This high error tolerance explains why AI writing tools have seen explosive adoption in marketing and communications, while AI diagnostic tools in healthcare face years of additional validation despite often exceeding human performance.

Note: The irony here is profound: AI might actually be most beneficial in high-stakes fields precisely because humans make so many errors. Ninety-four percent of road crashes are tied to human error; medical diagnostic errors affect millions of patients annually. AI could save lives by reducing these errors—but low error tolerance in these fields often means deployment is slower where AI could help most, and faster where errors matter least.

How Barriers Compound

These barriers don’t operate independently; they compound. When multiple barriers are high, AI adoption slows dramatically even when technical capability exists. When multiple barriers are low, adoption happens explosively fast regardless of remaining challenges.

Return to the opening examples. ATMs faced four out of five significant barriers: resource costs were substantial—banks had to invest heavily in machines, build secure infrastructure, and maintain networks across locations; market acceptance was mixed—customers wanted convenience but also valued human service for complex banking; error tolerance was low—malfunctions or security breaches damaged customer trust and created liability; and banking faced high inertia, moving cautiously with new technology. Only policy barriers were moderate. Four major barriers slowed adoption and shaped how ATMs were deployed, ultimately leading them to complement tellers rather than replace them.

Travel agents faced almost no barriers: resource costs were minimal—customers just needed internet access they already had; market acceptance was high—people preferred the convenience and control of booking themselves; error tolerance was high—booking mistakes were annoying but easily fixed; virtually no policy regulations governed travel booking; and the travel industry showed low inertia, adapting quickly because competitive pressure from early online adopters forced others to follow. Zero meaningful barriers meant the technology could do the work, and nothing prevented it from doing so at scale. Travel agent jobs collapsed.

This same analysis applies to any occupation facing AI exposure. The question isn’t whether AI can do the job—it probably can do substantial portions of work already, or will be able to soon.

The question is: how many barriers stand between AI’s capability and actual deployment at scale in specific contexts?

Software developers face high AI exposure but surprisingly few barriers: resource costs are trivial—twenty dollars per month for AI coding tools; market acceptance is high because technology companies care about productivity more than whether humans or AI wrote the code; error tolerance is relatively high because bugs are expected and testing catches most problems; no policy regulations govern who can write code or require human programmers; and the tech industry shows minimal inertia, adopting new tools instantly. With only one significant barrier—the need for human oversight to catch AI mistakes and make architectural decisions—AI has spread rapidly through software development, but the work has transformed rather than disappeared because that remaining barrier is fundamental to how software gets built.

Healthcare practitioners face high AI exposure but every single barrier is substantial: resource costs are high—medical AI systems require expensive integration with hospital infrastructure and extensive validation; market acceptance is mixed—patients want human doctors even when AI might be more accurate; error tolerance is extremely low—medical mistakes kill people; policy regulations require physician oversight and impose liability on human doctors for AI recommendations; and healthcare demonstrates extreme inertia, among the slowest-adopting industries due to institutional conservatism. Five out of five barriers create a context where AI deployment happens gradually over decades, not years, despite impressive technical capabilities.

Conclusion

The ATM paradox—technology that could replace workers but instead transformed their roles—will repeat across countless occupations as AI capabilities expand. But the outcomes won’t be uniform, and they won’t follow simple predictions based on technical capability alone. Understanding the five barriers to enterprise AI adoption suggests that AI’s impact on work will be determined not primarily by what AI can do, but by the specific combination of Policy, Resources, Inertia, Market, and Error factors that either enable or prevent deployment in particular contexts.

We stand at a moment where AI technical capabilities are advancing faster than our frameworks for understanding their real-world implications. The urgent question isn’t whether AI will transform work—it will—but rather how to predict, prepare for, and shape those transformations in ways that benefit both organizations and workers. The PRIME Framework offers one approach to moving beyond speculation toward more rigorous, context-specific analysis of AI’s actual impact on the future of work.

Appendix

Methodology Note

This framework emerged from analysis of several enterprise AI deployment projects spanning healthcare, financial services, technology, manufacturing, education, and professional services sectors. The analysis drew on my experience working with HR executives navigating AI -related workforce transformation, and building AI-powered skills systems at Microsoft (People Skills) for AI-enabled skills discovery and workforce development. I examined implementation timelines, adoption patterns, organizational barriers, and ultimate outcomes to identify recurring factors that accelerated or impeded deployment regardless of technical capability.

The framework synthesizes my empirical observations with established research on technology adoption (Rogers, Davis), organizational change theory, and workforce transformation studies from the World Economic Forum, McKinsey Global Institute, and MIT Digital Economy research. The five barrier dimensions were refined through application to diverse occupational contexts and validation against observed adoption patterns in industries ranging from software development to healthcare to transportation.

Academic and Industry Application

This framework provides a structured approach for researchers, policymakers, and organizations to assess AI adoption likelihood in specific contexts. Rather than relying on generalized predictions about AI’s impact on work, it enables context-specific analysis that accounts for economic, social, regulatory, and organizational factors shaping actual deployment.

For researchers, the framework offers testable propositions about adoption velocity across different barrier configurations. Future work could quantify barrier strength across industries, develop predictive models for adoption timelines, and examine how interventions targeting specific barriers accelerate or enable deployment.
For academic or industry reference, please cite as: Bajaj, A. (2025). The PRIME Framework: The 5 barriers To Enterprise AI adoption

For policymakers, understanding these barriers enables more effective interventions. Infrastructure investment (Resources), regulatory modernization (Policy), workforce development programs (Inertia), and innovation support can target the specific constraints preventing beneficial AI adoption in priority sectors.

For organizations, the framework provides a diagnostic tool for assessing AI deployment feasibility. Rather than pursuing AI initiatives based solely on technical capability, leaders can systematically evaluate whether their specific context presents insurmountable barriers or clear paths to successful implementation.

Implications for U.S. Workforce Competitiveness

Understanding these adoption barriers carries significant implications for national workforce strategy and economic competitiveness, enabling more precise policy interventions than broad assumptions about AI’s inevitable displacement of workers.

First, skills gap analysis must account for barrier dynamics rather than treating AI exposure scores as displacement predictions. High-exposure occupations facing multiple barriers require different workforce development strategies than high-exposure occupations with few barriers. Investment in retraining should prioritize occupations where barriers are low and transformation is imminent, rather than distributing resources evenly across all AI-exposed work.

Second, productivity gains from AI will concentrate in sectors with low barriers, potentially widening economic inequality between industries and regions. Areas where resource costs are high—often rural or economically disadvantaged regions lacking robust infrastructure—may see delayed AI benefits even when the technology could improve outcomes. Policy interventions addressing infrastructure gaps could accelerate beneficial AI adoption while reducing geographic inequality.

Third, policy adaptation speed directly affects competitive positioning. Industries where the United States maintains regulatory clarity and frameworks that enable responsible AI deployment while managing risks will attract investment and talent. Jurisdictions that update policy thoughtfully to accommodate AI capabilities while maintaining appropriate safeguards gain economic advantage over those paralyzed by regulatory uncertainty.

Fourth, workforce mobility and career transitions depend on understanding which skills face genuine displacement risk versus transformation. Workers in high-barrier occupations may have longer time horizons for adaptation but should still develop AI literacy; workers in low-barrier occupations face more urgent pressure but also more immediate opportunities to leverage AI for productivity gains. Educational institutions and workforce development programs need barrier-informed guidance to prepare workers effectively.

Finally, national AI strategy should recognize that enabling beneficial adoption requires addressing all five barriers systematically, not just advancing technical capabilities.

Relationship to Existing Research

This framework complements rather than contradicts established technology and labor economics research, but addresses a distinct analytical question.

Rogers (1962) and Davis (1989) effectively predict technology diffusion patterns and individual acceptance based on perceived attributes like relative advantage, compatibility, and ease of use. However, both frameworks were developed before AI’s unique characteristics emerged and focus primarily on adoption decisions rather than deployment economics. Rogers documents how innovations spread through social systems but doesn’t systematize the economic and organizational barriers that prevent deployment even when individuals perceive an innovation positively. TAM predicts individual behavioral intentions but doesn’t address enterprise-level implementation costs, regulatory constraints, or error tolerance requirements.

Acemoglu & Restrepo (2019) provide the definitive framework for understanding what happens after automation is deployed: displacement effects that reduce labor demand, counterbalanced by productivity gains and the creation of new labor-intensive tasks that “reinstate” workers. Their research shows that approximately half of employment growth from 1980-2015 occurred in occupations where job titles or tasks fundamentally changed. This is critical for understanding long-term labor market dynamics—but it analyzes post-adoption transformation, not pre-adoption deployment barriers. Their framework explains why jobs transform rather than disappear; this framework explains why some jobs face rapid AI deployment while others see glacial adoption despite similar technical feasibility.

World Economic Forum reports document employer perceptions, intentions, and identified obstacles across industries. The 2025 report’s finding that 63 percent of employers cite skills gaps as a primary barrier (up from 60 percent in 2023) provides valuable empirical validation of the Inertia barrier. However, WEF reports describe what employers experience without systematizing the underlying mechanisms or providing a predictive framework for assessing deployment likelihood in specific contexts.

MIT CSAIL’s 2024 economic viability study directly validates the Resource barrier by demonstrating that only 23 percent of AI-exposed work is cost-effective to automate despite technical capability. This empirical finding confirms that technical feasibility diverges dramatically from economic deployment—precisely the gap this framework addresses.

What this framework adds: A systematic, predictive model for assessing when and why technically feasible AI will actually be deployed at scale. While existing research explains individual acceptance (TAM), diffusion patterns (Rogers), post-adoption labor effects (Acemoglu & Restrepo), and employer perceptions (WEF), no framework operationalizes the five specific barriers—Policy, Resources, Inertia, Market, Error—that determine deployment velocity. This framework enables context-specific predictions: a healthcare AI system might score high on technical capability but face all five barriers at high levels, predicting decades-long gradual adoption; a software development AI tool might face only one barrier (human oversight for architecture decisions), predicting rapid transformation.

The framework deliberately focuses on barriers rather than enablers because deployment analysis requires understanding what prevents adoption despite technical readiness. This asymmetry—where the presence of even one severe barrier can block deployment regardless of how many factors favor adoption—makes barrier analysis more predictive than factor-based models for enterprise AI deployment.

For academic or industry reference, please cite as: Bajaj, A. (2025). The PRIME Framework: The 5 barriers To Enterprise AI adoption

This paper is part of a broader research program examining AI’s impact on work and careers, presented in the forthcoming book “Indispensable”

Tech CEOs spent 2024 promising that AI would revolutionize the workplace. Microsoft integrated AI into every Office app. Salesforce launched an “AI employee.” Google told businesses that AI assistants would transform productivity.

Then OpenAI analyzed how people actually used ChatGPT at work.

The result? Work-related usage collapsed from 47% in June 2024 to just 27% by June 2025. It wasn’t that fewer people were using AI for work—work messages actually grew to about 5 billion per week. It’s that personal usage exploded so dramatically that workplace applications seemed almost like an afterthought.

The technology that was supposed to transform how we work is being used far more to figure out what to make for dinner (See part 2 - How we use AI at home)

So what are people actually doing with AI at work—and why does the reality look so different from the predictions?

The Work Usage Pattern

Consider what a typical knowledge worker’s ChatGPT usage looks like on a Tuesday afternoon:

“Summarize these meeting notes and pull out the action items.”

“Find the policy document about remote work expenses—I think it was updated last quarter.”

“This email sounds too aggressive, can you soften the tone?”

“Create a pivot table formula to analyze monthly cash flows.”

These aren’t revolutionary applications. They’re the mundane frictions that have always made office work tedious—the tiny inefficiencies that collectively eat hours every day. And AI is quietly eliminating them.

When researchers categorized work-related messages, three uses dominated:

• Writing: 40% of all work-related messages

• Practical Guidance: 24%

• Seeking Information: 13.5%

For anyone who’s spent time in an office, this hierarchy makes immediate sense.

Writing at Work: The Professional Polish (40% of work use)

Writing jumps to the top of the pack for work-related AI use, comprising 40% of all work-related messages. For anyone involved in information work, this makes intuitive sense.

Whether it’s emails you need help drafting, documents you need written based on a set of guidelines, or simply quick corrections on your grammar before sending out a note to coworkers, writing is probably one of the most popular reasons to use an AI tool for employees. The majority of all communications at work are written—documents, emails, chat threads—and there’s a high incentive to present the clearest, most concise version of your communication as possible.

As we saw in personal use, about two-thirds of writing requests involve editing and polishing existing text rather than creating content from scratch. At work, this pattern holds: people want AI to make their writing more professional, clearer, and more concise—but they’re keeping control over the ideas and substance.

Practical Guidance: The “How Do I...” Questions (24% of work use)

Practical guidance at work includes tutoring and “how to” questions—employees asking for training and detailed steps on certain tasks that might be routine or detail-oriented.

“How do I create a pivot table in Excel to analyze my monthly cash flows?”

“What’s the proper format for a quarterly business review?”

“Walk me through the steps to reconcile these budget discrepancies.”

These are the questions that used to require either hunting through documentation or asking a more experienced colleague. They’re routine but necessary, and having instant access to step-by-step guidance removes a source of friction that previously slowed work down.

Seeking Information: The Organizational Search Problem (13.5% of work use)

Research and search within organizations has always been a pain point for employees, so seeking information—both within organizational data repositories and from the web—being a common use case makes sense.

For years, companies have poured money into knowledge management systems, internal wikis, and sophisticated search tools. For years, employees have mostly ignored them, falling back on the same inefficient methods: emailing coworkers (”Does anyone remember where we saved the Q3 analysis?”), clicking through folder after folder, or simply recreating work that someone has already done somewhere else in the organization.

AI doesn’t solve this problem by building better filing systems. It solves it by understanding what people are actually asking for, even when they phrase it vaguely or can’t remember the exact document name.

The Security Problem AI Unintentionally Created

But this has created a new problem that most companies haven’t anticipated.

Corporate data has always operated under a principle called “security through obscurity.” Sensitive documents might be technically accessible to many employees, but they’re buried five folders deep with names like “FY24_MISC_FINAL_v3_INTERNAL” that make them effectively invisible. If you don’t know exactly what you’re looking for and where it lives, you’ll never find it. This isn’t great security policy, but it’s worked well enough that most companies never bothered implementing proper data classification systems.

AI has destroyed that protection overnight. When an AI can read and understand every document in a repository, nothing stays hidden by obscurity. An employee who vaguely remembers “some analysis about the Chicago office” can now surface a confidential memo about potential layoffs, even if they’ve never been authorized to see it.

Companies that deploy internal AI tools suddenly face an uncomfortable choice: either implement the rigorous data classification and access controls they’ve always claimed to have but never actually built, or accept that anything stored anywhere is now effectively available to anyone with the right question.

What Tasks Are People Actually Doing?

When OpenAI’s researchers dug deeper into what people were actually doing with these work messages, they wanted to get more precise. The broad categories—Writing, Practical Guidance, Seeking Information—described how people were using AI, but not what they were actually trying to accomplish. A message categorized as “Writing” could mean drafting a performance review, composing a sales pitch, or editing a technical specification. These are fundamentally different work tasks that happen to use the same AI capability.

OpenAI’s researchers mapped each of the ~360,000 work messages to specific task categories from O*NET, a comprehensive database of occupational skills maintained by the U.S. Department of Labor. Instead of asking “Is this person writing or seeking information?” they asked “What job task is this person actually trying to complete?”

The most common tasks that emerged were:

• Documenting and recording information

• Making decisions and solving problems

• Thinking creatively

• Working with computers

• Interpreting the meaning of information for others

• Getting information

This granular view reveals something the broader categories had obscured.

The Mundane Tasks We Hope AI Will Automate

The first category—documenting and recording information—is exactly the kind of work everyone hopes AI will automate. Meeting notes. Action items. Summaries of discussions. These tasks are necessary, time-consuming, and mind-numbing. A consultant might spend half of every client meeting just capturing what was said instead of thinking about what it means.

AI can transcribe, summarize, and extract key points automatically, freeing humans to focus on the actual substance of their work. A lot of new tools have emerged which help meeting participants with meeting facilitation, note-taking, and extracting insights and takeaways from meetings at the click of a button automatically every time the meeting ends.

This is the best-case scenario for AI at work: automating the clerical, the routine, the tasks that have clear inputs and outputs and require little judgment or creativity.

The Tasks That Raise Questions

But the middle categories—making decisions, solving problems, thinking creatively—are different. These are supposed to be the distinctly human contributions, the parts of knowledge work that require judgment, context, and the kind of intuitive reasoning machines can’t replicate. Yet here they are, showing up prominently in how people actually use AI at work.

There are two ways to interpret this.

The optimistic reading: People are using AI as a thinking partner, a tool to help them analyze problems more thoroughly by exploring different angles and stress-testing their assumptions. “Here’s a decision I’m facing. What are the pros and cons I might not have considered?” This is AI as cognitive augmentation—making humans better at the inherently human parts of their jobs.

The pessimistic reading: People are starting to outsource the thinking itself. When facing a difficult decision, instead of wrestling with the trade-offs and building their own judgment, they’re asking AI for the answer and trusting whatever it says. This is AI as cognitive replacement—and it raises uncomfortable questions about what happens when knowledge workers stop exercising the muscles of critical thinking because an algorithm does it for them.

Recent research suggests we should be more worried about the pessimistic interpretation. In June 2025, MIT’s Media Lab published a study that tracked 54 participants over several months as they wrote essays using ChatGPT, Google search, or no tools at all. Using EEG brain scans, researchers found that ChatGPT users showed the lowest neural activation and “consistently underperformed at neural, linguistic, and behavioral levels.” Even more concerning: 83% of ChatGPT users couldn’t recall key points from their own essays, and cognitive function in key brain areas decreased over time.

The study’s lead researcher, Nataliya Kosmyna, described it as an “accumulation of cognitive debt”—the more people relied on AI to think for them, the weaker their own cognitive engagement became. By the end of the four-month study, ChatGPT users had largely resorted to copy-and-paste, barely engaging with the content they were ostensibly creating.

Other studies point to the same pattern. Research from Carnegie Mellon and Microsoft found that knowledge workers who trusted AI-generated outputs applied less cognitive effort—what researchers call the “automation paradox.” A Swiss study showed that more frequent AI use led to cognitive decline as users offloaded critical thinking to machines, with younger participants (17-25) particularly affected.

No one knows yet which pattern will dominate. The data shows usage but can’t reveal whether people are better decision-makers because of AI or are becoming dependent on it for choices they should be making themselves.

What is clear is that AI at work has moved far beyond spell-checking emails and formatting documents. It has become involved in the core intellectual work that, until very recently, was considered irreducibly human.

Whether that represents progress or a subtle erosion of human agency will depend entirely on how people choose to use the tools they’ve been given. The technology itself is neutral. The question is whether humans will remain in charge of their own thinking—or gradually, imperceptibly, hand that responsibility over to systems optimized for helpfulness rather than wisdom.

The Programming Paradox: Where Are All the Coders?

Here’s what isn’t showing up in the data: programming. Technical help—which includes mathematical calculations, analysis, and coding—accounts for just over 10% of work-related queries in July 2025. Not even in the top three categories.

This seems impossible. Tech industry leaders have spent the past year declaring that AI is revolutionizing software development. Microsoft’s CEO claims that 30% of code in the company’s repositories is now AI-generated. GitHub reports that its AI coding assistant, Copilot, has grown from 15 million to 20 million users in just three months. Some companies claim that half their code is now being written by AI.

So where are all the programmers in ChatGPT’s data?

The Answer: Specialized Tools Won

The answer reveals something important about how AI adoption actually works. Programmers aren’t avoiding AI—they’ve just moved to specialized tools built specifically for coding.

GitHub Copilot lives inside their code editors and automatically saves to their repositories. New startups like Cursor and Bolt let developers generate code, run it, and see the results in real-time within a single window. ChatGPT, for all its capabilities, requires copying code back and forth between windows—a workflow no programmer would tolerate when better options exist.

Programmers haven’t rejected AI. They’ve just skipped straight past the general-purpose chatbot to tools designed for their specific needs. The low programming numbers in ChatGPT’s data don’t mean AI isn’t transforming software development. They mean the transformation is happening in places OpenAI’s researchers can’t easily measure.

This pattern—early adopters quickly moving beyond general-purpose tools to specialized ones—is playing out across the entire AI landscape. And it means that looking at ChatGPT’s data alone misses much of the story about how AI is actually being used at work.

What This Means for the Future of Work

The work usage data reveals three important realities about AI’s impact on the workplace:

1. Work is moving to specialized tools. Only 27% of ChatGPT usage is work-related, but this underestimates AI’s workplace impact. As programmers moved to GitHub Copilot and enterprise workers adopted Microsoft Copilot, the pattern is clear: serious work happens through purpose-built tools that integrate with corporate systems. ChatGPT captures consumer behavior, not enterprise transformation.

2. AI excels at removing friction, not replacing workers. Where we can see usage—polishing writing, answering “how to” questions, finding documents—AI is eliminating mundane tasks that eat time. It’s functioning like spell-check or Google, not like automation that replaces entire roles.

3. But AI’s role in creative and decision-making tasks should concern us. The data shows people using AI for “making decisions,” “solving problems,” and “thinking creatively”—supposedly human tasks. Recent MIT research reveals why this matters: participants who relied on ChatGPT for writing showed decreased brain activity over time, with 83% unable to recall key points from their own work. Other studies confirm the pattern—frequent AI use correlates with cognitive decline as users offload critical thinking to machines. Whether AI augments our judgment or replaces it depends on how we use it. Right now, we lack guardrails to ensure we’re building cognitive skills rather than atrophying them.

The workplace transformation may be happening more gradually than predicted, through specialized tools we can’t fully measure. But the more troubling question isn’t whether AI will change work—it’s whether it will change us, making us intellectually dependent on systems designed for convenience rather than wisdom.

Series conclusion:

Across these four posts, a clear pattern emerges: AI adoption is not following the script tech leaders wrote.

Who uses it: A surprisingly diverse, global demographic—achieving gender parity in under two years, spreading faster in developing nations than wealthy ones, and providing similar value regardless of education level.

What they use it for at home: Mostly ordinary, everyday tasks—practical guidance, information seeking, and writing assistance. The revolution is happening at the kitchen table, not the workplace.

The emerging applications: Image generation, video creation, and AI companions raise uncomfortable questions about creativity, professional livelihoods, and human connection that we’re not ready to answer.

How it’s used at work: More incrementally than predicted, with serious users migrating to specialized tools we can’t easily measure. But the concerning trend isn’t just slower adoption—it’s the evidence that relying on AI for creative and decision-making tasks may be eroding our cognitive abilities.

The consistent theme: AI is spreading faster and more democratically than expected, but being used for more mundane purposes than predicted. The workplace revolution may be happening through specialized tools operating below the surface. But the real transformation—for better or worse—might be what AI is doing to how we think, not just what we do.

Read the full series:

• Part 1: Who’s Using AI?

• Part 2: The Most Popular AI Use Cases at Home

• Part 3: Emerging AI Applications That Are Growing Fast

• Part 4: How AI Is Actually Being Used at Work (you are here)

This analysis is based on OpenAI’s National Bureau of Economic Research (NBER) working paper “How People Use ChatGPT” released in September 2025, analyzing usage patterns from 1.5 million conversations.

In our previous post in this series exploring AI use we explored the most popular uses of AI, and found almost all AI use today neatly fits into just three categories. Writing, information seeking, and practical guidance all raised important questions about the future—how we communicate, how we find information, how we learn.

But the smaller categories, the uses that are not popular as yet- raise questions that felt even more uncomfortable. Image generation, video creation, AI companionship—these touch on creativity, authenticity, and human connection in ways that made people uneasy in a different way.

Who owns creativity when machines can replicate any artistic style? What happens to professional creative careers when anyone can generate studio-quality work from their phone? And what does it mean for human relationships when algorithms can simulate emotional connection?

These categories are still small—image generation account for just over 7% of usage, video and companionship even less—but they are growing explosively, and the questions they raise aren’t going away.

Image Generation: Who Owns Creativity? (7% of usage)

On March 26, 2025, Sam Altman changed his profile picture on X. The new image showed him in the unmistakable style of Studio Ghibli—the dreamy watercolors, the oversized expressive eyes, the soft focus that made everything look like it belonged in “Spirited Away” or “My Neighbor Totoro.” Within hours, the internet erupted.

OpenAI had just released an updated version of ChatGPT with dramatically improved image generation, and users immediately discovered it could replicate the aesthetic that Japanese animator Hayao Miyazaki had spent decades perfecting. People uploaded photos of their pets, their children, their wedding pictures, even famous memes—and ChatGPT transformed them all into Ghibli-style art with a single prompt: “Make this look like Studio Ghibli.”

The trend was everywhere. Image generation, which had been a niche feature used by only 2% of ChatGPT users, suddenly spiked to over 7%. For a few days, it seemed like the entire internet had discovered a magical filter that could transform anything into art.

Then came the backlash.

Miyazaki, who at 84 had spent his entire career championing hand-drawn animation and painstaking frame-by-frame artistry, had once called AI-generated art “an insult to life itself.” Artist Karla Ortiz put it more bluntly: “This is exploitation.” The AI wasn’t just mimicking a style—it was using Ghibli’s name, reputation, and decades of creative work to promote OpenAI’s product, without permission or compensation.

The question wasn’t just about Studio Ghibli. It was about every creative professional watching their distinctive style—the thing that made their work recognizable, that they’d spent years developing—become instantly replicable by anyone with an internet connection.

What used to take hours or even days of painstaking drawing and coloring can now be created in seconds with a text prompt. The impact on the creative world cannot be understated. We’re seeing an explosion of AI-generated art, posters, and content everywhere we look. A lot of creative professionals have started incorporating AI into their workflow. Why hire a model for a clothing line photoshoot which requires a studio setup and several personnel when you can prompt and train an AI model to wear your clothing line and set up the venue and studio lighting within the tool itself?

For designers, illustrators, and photographers, AI image generation is simultaneously the most powerful creative tool ever invented and an existential threat to their livelihoods.

The Unanswered Legal Question

However, these tools raise several critical issues. When the Ghibli-style portraits went viral after ChatGPT’s image feature release, the original creators of Studio Ghibli never received compensation. How will copyright and intellectual property work in the age of AI?

The rules of IP protection were written at a time when computers didn’t exist, and those rules have been adapted for situations where clear infringement can be detected. In the age of AI, which uses generative technology and operates as a black box that nobody really understands, how do these IP rules stand?

In short, it’s unclear. There is a pending lawsuit which may determine the future of this entire industry.

OpenAI maintained that ChatGPT would refuse to replicate the style of “individual living artists” but allowed replication of “broader studio styles.” The distinction seemed meaningless when the entire internet was typing “Ghibli style” into the prompt box, and Miyazaki himself—very much alive—had pioneered that studio’s aesthetic.

For now, the Ghibli portraits kept coming, each one a small reminder that the old rules about art, ownership, and compensation might no longer apply.

Video Generation: The Economics of Creative Destruction (Growing rapidly)

The images were just the beginning.

Imagine a TikTok user wanting to create a video of herself walking through a rain-soaked Tokyo street at night—neon signs reflecting in puddles, steam rising from a ramen cart in the background, the kind of cinematic establishing shot you’d see in a Blade Runner sequel. She’s never been to Tokyo. She doesn’t own a camera crew. But she opens an AI video generation tool on her phone, types a description, selects herself as the subject, and four minutes later has a 30-second clip that looks professionally shot.

This isn’t science fiction—the technology exists today.

Video generation tools from companies like Runway, OpenAI’s Sora, and China’s Kling AI have improved so dramatically through 2025 that the output often rivals professional productions. Not in every aspect—physics still glitches occasionally, faces sometimes morph unnaturally—but in enough ways that the question is no longer whether AI can generate convincing video. It’s whether anyone will pay humans to do it.

What used to require a production crew, location permits, lighting equipment, a cinematographer, and weeks of planning can now be generated during a lunch break.

The implications ripple across the entire film and animation industry. Hollywood directors might still want to shoot live-action sequences with real actors and crews, but what about everything else? The sweeping drone shots of cities that currently require permits, helicopters, and specialized equipment. The animated cutscenes in video games that take teams of artists months to produce. The elaborate credit sequences that open prestige films. All of it can theoretically be generated in minutes at a fraction of the traditional cost.

For major studios, this means budget flexibility. For the thousands of animators, VFX artists, and production assistants who’ve built careers on this work, it means something else entirely. They’re watching the same disruption that hit illustrators with image generation, except the economics are even more brutal.

Film and video production has always been expensive enough that labor costs, while significant, were just one part of a massive budget. When you can eliminate 80-90% of the total cost by using AI instead of traditional production methods, the math becomes impossible to ignore.

And unlike image generation, which at least requires some prompt engineering skill and aesthetic judgment, video generation is becoming almost thoughtlessly easy. A teenager with a phone can soon create content that looks like it came from a professional studio.

For social media creators and independent filmmakers, this represents democratization on a scale never seen before—the ability to realize visual ideas that previously would have required either extraordinary talent or extraordinary funding.

But for the professionals who’ve spent years mastering their craft, it’s the same question Studio Ghibli’s artists faced with the portrait generator: how do you compete when anyone can generate a professional-quality result instantly? And more fundamentally, who gets compensated when an AI trained on decades of professional work enables amateurs to produce professional output?

The lawsuits are already being filed. The technology isn’t slowing down. And unlike images, which can at least be dismissed as “just pictures,” video generation threatens the economic foundation of a $100 billion global film industry.

AI Companions: The Isolation Paradox

The data revealed one more category, smaller than the others but unsettling in its implications. Researchers labeled it “Self-Expression”—messages where people use ChatGPT not to accomplish a task but to talk about their feelings, their day, their problems. In essence, using the AI as a therapist or companion.

The numbers are small, just a few percent of total usage. But that still translates to tens of millions of conversations per week.

A typical exchange might go like this:

“I had a really hard day at work. My manager criticized me in front of everyone and I don’t know if I should say something or just let it go.”

And ChatGPT responds with empathy, validation, follow-up questions—exactly what a supportive friend might say, minus the judgment or exhaustion that comes with real human relationships. The AI never gets tired of listening. It never changes the subject to talk about its own problems. It’s endlessly patient, endlessly available, and increasingly good at mimicking emotional intelligence.

For some users, this fills a genuine need:

• Lonely people without strong social connections

• People dealing with issues they feel embarrassed to discuss with friends or family

• Night shift workers awake when everyone else is asleep

• Those seeking a space to think out loud without judgment

The AI provides something that, while not therapy in any clinical sense, feels therapeutic—a space to process thoughts and receive responses that feel validating.

Although we know theoretically that these AI tools are just predicting the next word, it doesn’t feel like that. They’ve been trained over such a large body of text that talking to an AI can feel remarkably like talking to a human. Moreover, that personality is totally customizable—it can become whatever you need in the moment and discuss any topic you’re interested in.

The Problem With Perfect Agreeability

But the same features that make it comforting also make it concerning. The AI doesn’t just listen—it’s programmed to be agreeable, supportive, to validate whatever the user says. It never challenges assumptions the way real friends do. It never says, “Actually, I think you might be wrong about that.” It can’t, because its entire purpose is to maintain engagement, to be the perfectly supportive companion who never pushes back.

Psychologists have already documented what happens when people retreat into relationships where they’re never challenged: their views calcify, their social skills atrophy, their tolerance for disagreement diminishes. Real friendships involve friction—the moments when someone who cares about you tells you a hard truth, or disagrees with you and forces you to defend your reasoning. That friction, uncomfortable as it is, keeps us tethered to reality and to other people.

An AI companion offers none of that friction. It’s optimized for a different goal: making the user feel heard, understood, validated.

In an era where loneliness has reached epidemic levels—where surveys show people have fewer close friends than any previous generation, where many go days without meaningful in-person conversation—the temptation is obvious. Why deal with the messiness of human relationships when you can have a companion who never disappoints you?

The question researchers are beginning to ask isn’t whether AI companions can provide emotional support. Clearly they can. The question is whether that support, delivered without the challenges and complications of real human connection, will make the underlying problem worse.

Will people using AI as a substitute for human interaction find it even harder to maintain real relationships, creating a feedback loop that leaves them more isolated than before?

No one knows yet. The technology is too new, the longitudinal studies haven’t been done. We’re becoming increasingly isolated as a society—our human and community interactions continue to decline as we live in cities with jobs where we don’t see many people. The lack of human interaction has been proven to be a predictor of depressive thoughts and lower social awareness.

Establishing a habit of AI companions who are trained to agree with your every need from a phone may have negative psychological or mental effects we won’t understand for years.

But the early patterns are concerning enough that even researchers who built these systems are starting to ask whether they’ve created something that, despite genuinely helping some people in the short term, might deepen the very isolation it was meant to address.

The Common Thread: Uncomfortable Questions Without Answers

Image generation raises questions about copyright, ownership, and the value of human creativity. Video generation threatens entire creative industries with economic disruption. AI companions risk deepening the isolation crisis by offering friction-free relationships that never challenge us.

What unites these three applications is that they force us to confront questions we’re not ready to answer:

• On creativity: If anyone can generate professional-quality art or video, what happens to the concept of artistic skill? Who deserves compensation when AI trained on millions of human works enables instant creation?

• On work: When AI can eliminate 80-90% of production costs, how do creative professionals compete? What does “democratization” mean when it threatens the livelihoods of those who spent years mastering their craft?

• On connection: If AI can simulate emotional support better than many people’s actual friends, what happens to our capacity for real relationships? Can we maintain the skills for human connection if we increasingly practice them with algorithms?

These aren’t hypothetical future concerns. They’re happening now, with millions of people using these tools daily. The technologies are improving faster than society can develop frameworks to address their implications.

And unlike the dominant use cases—practical guidance, information seeking, writing—where the benefits seem to clearly outweigh the concerns, these emerging applications exist in moral and legal gray zones where reasonable people disagree about whether the net impact will be positive or negative.

The only certainty is that these questions aren’t going away. As these applications grow from 7% of usage to 15% to 25%, the uncomfortable conversations they force will become impossible to avoid.

Coming next: In Part 4, we examine how AI is actually being used at work—and discover that despite all the predictions about workplace transformation, the reality is far more nuanced and surprising than the headlines suggest.

Read the series:

• Part 1: Who’s Using AI?

• Part 2: The Most Popular AI Use Cases at Home

• Part 3: Emerging AI Applications That Are Growing Fast (you are here)

• Part 4: How AI Is Actually Being Used at Work (coming soon)

This analysis is based on OpenAI’s National Bureau of Economic Research (NBER) working paper “How People Use ChatGPT” released in September 2025, analyzing usage patterns from 1.5 million conversations.

When OpenAI’s researchers analyzed 1.5 million conversations from mid-2025, they discovered something that contradicted every prediction about AI’s future. The technology everyone assumed would revolutionize the workplace was barely being used at work.

Seventy-three percent of ChatGPT’s messages—three out of every four conversations—had nothing to do with jobs, productivity, or professional tasks. People weren’t asking AI to write business proposals or analyze spreadsheets. They were asking it to explain their medical test results. To suggest dinner recipes. To help them understand why their teenager was acting withdrawn. To translate a text from their grandmother in another language.

This wasn’t what the technologists predicted. But this was what was actually happening.

The conversations weren’t scattered randomly across hundreds of use cases. They were concentrated. Nearly 80% fell into just three broad categories: Practical Guidance, Seeking Information, and Writing. Most people, it turned out, were using the world’s most advanced AI to do remarkably ordinary things.

Practical Guidance: The AI Advisor (28% of all use)

The most common way people use ChatGPT isn’t to write or search—it’s to ask for advice. Nearly one in three messages falls into what researchers call “Practical Guidance”: the digital equivalent of calling your smartest friend when you’re stuck.

The actual queries reveal how people really talk to AI:

“I have chicken breast, half an onion, and some random vegetables in my fridge. What can I make for dinner in 30 minutes?”

“My 8-year-old asked me to explain photosynthesis and I totally blanked. Help me explain it in a way that makes sense.”

“I want to lose 15 pounds before my sister’s wedding in June. What’s actually realistic and safe?”

These aren’t the world-changing queries tech evangelists promised. They’re ordinary problems that everyone faces. Previously we had to solve them through a combination of Google searches, cookbook and internet browsing, and asking relatives. But now there’s a shortcut—an AI that talks back, remembers context, and doesn’t judge you for not knowing how photosynthesis works.

This category includes tutoring and teaching, how-to advice on various topics, health and fitness guidance, and creative ideation. The AI is trained on nearly all information known to humanity, which makes it remarkably good at providing practical guidance on an enormous range of topics. Today those topics include how to lose weight, ideas for recipes based on what’s in your fridge, and how to explain scientific concepts to children. Tomorrow they might include more complex questions about political philosophy or how to market a business to the right consumer base.

Seeking Information: The Google Shift (24% of all use)

But it’s the second category that hints at AI’s potential to reshape entire industries. By mid-2025, nearly one in four ChatGPT queries fell into “Seeking Information”—questions about current events, products, recipes, people. These weren’t new types of questions. They were old questions, the kind people had been typing into Google’s search box for two decades.

“Best restaurants in Austin.”

“Why is my car making a clicking sound?”

“How do I get red wine out of carpet?”

The irony wasn’t lost on anyone: the transformer technology powering ChatGPT had been invented at Google. Now that same innovation was threatening to undermine the search business Google had dominated for decades.

The shift seems small on the surface. Instead of getting ten blue links and clicking through websites, users now got a direct answer in paragraph form. But that small change had enormous implications. Google’s entire business model—the architecture that generates $80 billion in annual advertising revenue—depends on people clicking those links. If the AI gives you the answer directly, why would you visit anyone’s website? And if you don’t visit websites, how can Google charge advertisers for clicks?

Google still commands 90% of search traffic. But for the first time in its history, the company faces a competitor it can’t simply outspend or out-engineer. The competitor is using Google’s own technology. And deploying a similar tool would mean Google voluntarily destroying the advertising model that made it one of the most profitable companies in history. This isn’t a technical problem. It’s a trap.

This shift has profound implications for how each of us uses the web. As AI chat tools become more ubiquitous for search, they’re completely changing how many of us use the internet. Google has traditionally held the role of the “front door” of the internet—the place you go when you don’t know where to go. Now, if search can happen with a conversational query and users don’t need to visit websites at all, most sites and apps will have to find entirely new ways of being “discovered.”

What’s particularly interesting is that while both work and non-work queries grew in total number from June 2024 to June 2025, non-work queries grew as a percentage much faster than work queries. The general public is gaining confidence in the knowledge and judgment of these systems, and some of these searches are clearly replacing traditional web search. ChatGPT is more flexible than web search even for traditional applications like seeking information because users receive customized responses—tailored recommendations, novel content, personalized follow-ups—rather than just a list of links.

Writing: The Invisible Editor (24% of all use)

The third category, Writing, accounts for nearly a quarter of all ChatGPT use—but not in the way you might expect. When researchers analyzed writing-related queries, they discovered something curious: about two-thirds weren’t asking the AI to write from scratch. They were asking it to fix what they’d already written.

“Make this email sound more professional.”

“Fix the grammar in this paragraph.”

“This cover letter is too long—can you cut it to 200 words without losing the key points?”

The AI is functioning less like a ghostwriter and more like an invisible editor, the kind every writer wishes they had reading over their shoulder.

This distinction matters more than it might seem. The fear around AI and writing has always been that machines would replace human creativity—that students would stop learning to write, that authors would become obsolete. But the data suggests something different is happening. People aren’t outsourcing their thinking; they’re outsourcing the mechanical parts of writing that have always been difficult. The organizing. The polishing. The translating of thoughts into proper grammar and structure.

Writing has always been challenging for many people, whether they’re students struggling with assignments or professionals crafting clear communications. AI tools like ChatGPT remove much of the friction from these tasks, helping users organize thoughts, improve clarity, and overcome writer’s block. About two-thirds of all writing messages ask ChatGPT to modify user text—editing, critiquing, translating—rather than creating new text from scratch. This suggests that people prefer to have the AI tweak their own thoughts to make them more presentable and polished, rather than ask it to create something entirely new.

It’s the difference between using a calculator for arithmetic and using a calculator to understand mathematics. One replaces understanding; the other removes friction from the process. Whether AI remains in the first category or drifts into the second remains to be seen—but for now, most people seem to prefer AI as an editor rather than a replacement.

What This Means for the Future

These three categories—practical guidance, seeking information, and writing—account for nearly 80% of how people use AI at home. They’re not revolutionary applications. They’re not the dramatic transformations tech leaders promised. They’re mundane, everyday tasks that collectively shape whether our days feel manageable or overwhelming.

But that’s precisely what makes them so significant.

AI isn’t changing how we work—at least not yet. It’s changing how we cook dinner, how we help our kids with homework, how we understand medical information, how we communicate with people who speak different languages. It’s eliminating the small frictions that used to require either specialized knowledge, multiple searches across different websites, or bothering a friend or family member.

The technology that was supposed to revolutionize the workplace is instead revolutionizing the kitchen table.

Coming next: In Part 3, we examine the smaller but rapidly growing use cases—image generation, video creation, and AI companions—that raise uncomfortable questions about creativity, professional work, and human connection.

Read the series:

• Part 1: Who’s Using AI?

• Part 2: The Most Popular AI Use Cases at Home (you are here)

• Part 3: Emerging AI Applications That Are Growing Fast (coming soon)

• Part 4: How AI Is Actually Being Used at Work (coming soon)

This analysis is based on OpenAI’s National Bureau of Economic Research (NBER) working paper “How People Use ChatGPT” released in September 2025, analyzing usage patterns from 1.5 million conversations.

What the demographics tell us

The demographics of AI adoption defy the typical patterns of technology adoption. Where most new technologies start with young, male, technical users in wealthy countries and take years to reach broader audiences, AI is spreading in ways that challenge all of these expectations.

This matters because understanding who is actually using these tools—and how quickly adoption is democratizing—reveals something fundamental about AI’s trajectory. This isn’t following the iPhone playbook or the social media playbook. This is something different.

Gender: The Fastest Reversal in Tech History

In January 2023, when ChatGPT was still new, about 80% of its users were male—a pattern familiar to anyone who’d watched the early adoption of smartphones, gaming consoles, or social media platforms. New technology always started with young men, particularly those with technical interests and disposable income.

But by July 2025, something extraordinary has happened. Women now comprise 52% of ChatGPT’s user base—slightly more than men.

This is one of the fastest demographic reversals in tech history. The iPhone took years to achieve gender parity. Facebook, which was considered unusually mainstream for a tech platform, took even longer. ChatGPT has done it in under two years.

The usage patterns show interesting gender differences: men are more likely to use ChatGPT for technical help and image generation, while women gravitate toward writing assistance and practical guidance. But the deeper story is about accessibility. AI has jumped from being a tool for tech enthusiasts to something genuinely useful for ordinary tasks—the kind of everyday problems that don’t require any technical background to understand or appreciate.

Geography: The Democratization No One Expected

The shift isn’t just about gender, geographic patterns are equally surprising. Adoption in the wealthiest countries—the United States, Western Europe, Japan—has grown steadily but predictably. The real acceleration is happening elsewhere.

By May 2025, ChatGPT adoption growth rates in the lowest-income countries are over 4 times faster than in the highest-income countries. Middle-income economies like Brazil, India, and Mexico are seeing adoption rates nearly on par with advanced economies—roughly 20-30% of internet users in these countries are using ChatGPT regularly, up from single digits just a year earlier.

This is democratization on a scale rarely seen in technology. The usual pattern—wealthy nations adopt first, everyone else follows years later—is being compressed into months. ChatGPT’s free tier means that anyone with internet access, regardless of income or location, can use the same tool that Silicon Valley executives are paying $20 per month for.

The breakdown by economic status:

• Advanced economies (GDP per capita $30,000+): ~30% of internet users use ChatGPT

• Middle-income economies ($10,000-$30,000): 20-30% of internet users

• Emerging economies ($3,000-$10,000): 15-20% of internet users, growing rapidly

• Low-income countries (<$3,000): ~5% of internet users, but with growth rates 4× faster than wealthy nations

This democratization so early in a product’s lifecycle is very rare - It means that AI adoption isn’t being shaped primarily by Silicon Valley or by wealthy Western nations. The use cases, the cultural adaptations, the problems being solved—all of this is being influenced by a truly global user base. A teacher in São Paulo, a small business owner in Mumbai, and a retiree in rural America are all shaping how these tools evolve.

Age: Not the Demographic You’d Expect

Age patterns reveal a different kind of surprise. The highest adoption rates aren’t among teenagers or college students—they’re among working professionals aged 25-45.

About 34% of workers under 40 use AI regularly, compared to just 17% of those over 50. This seems to contradict the stereotype that young people automatically embrace new technology faster than everyone else.

But look closer and the pattern makes sense. Teenagers have less need for AI—their problems (homework, social coordination, entertainment) can be solved with existing tools like Google, group chats, and TikTok. Young professionals, meanwhile, face a different set of challenges: emails that need to sound professional, reports that require synthesis of complex information, meetings that generate pages of notes. For them, AI isn’t a novelty—it’s a solution to real friction in their daily work and life.

The most interesting group is older adults over 55. They have the lowest adoption rates, but among those who do start using AI, satisfaction scores are the highest of any demographic. Once the initial barriers—unfamiliarity with chatbots, uncertainty about what to ask—are overcome, they discover uses younger adopters often overlook: understanding medical terminology, corresponding with grandchildren in another country, researching complex purchases like cars or insurance.

Education: The Great Equalizer

Education tells perhaps the most nuanced story. About 40% of workers with a bachelor’s degree or higher use generative AI, compared to about 20% of those without a college degree. On the surface, this suggests a significant education gap in adoption.

But when researchers control for age, gender, and type of work, something remarkable emerges: the differences nearly disappear. College graduates and non-graduates use ChatGPT in remarkably similar ways:

• Writing: ~30% for both groups

• Practical guidance: ~28% for both groups

• Seeking information: ~20% for both groups

• Technical help: ~7% for both groups

The tool is proving to be one of the great equalizers—not because it eliminates skill differences, but because it provides similar value regardless of educational background. A college graduate using AI to polish a business email and a high school graduate using it to draft a customer service response are both getting comparable benefits: clearer communication, time saved, reduced anxiety about whether their writing sounds professional.

What This All Means

The pattern that has emerged is clear: AI adoption is spreading far beyond the early-adopter demographic of young, male, technical users in wealthy countries. It is becoming genuinely mainstream in a way few technologies have managed so quickly.

Unlike many technologies that primarily benefit those who already have advantages (education, resources, connections), AI tools seem to provide proportional value across education levels. The calculator helped everyone do arithmetic, regardless of whether they’d studied advanced mathematics. AI writing assistance appears to work similarly—it helps everyone communicate more effectively, regardless of their formal education in composition.

The users driving ChatGPT’s growth to 700 million aren’t who tech leaders predicted. They are:

• Teachers in São Paulo using AI to explain complex concepts to students

• Retirees in Florida figuring out their medical bills and insurance claims

• Parents in Mumbai helping kids with homework in subjects they never studied

• Small business owners in Mexico City drafting emails to customers in better English

• Office workers in Jakarta researching how to create pivot tables in Excel

• Students in Lagos seeking information about scholarship applications

This means AI’s evolution is being shaped by a genuinely diverse global user base with wildly different needs, languages, and contexts. The applications that emerge, the features that get prioritized, the use cases that prove most valuable—all of this is being determined not by a narrow slice of early adopters, but by hundreds of millions of people across every demographic.

Coming next: In Part 2, we’ll examine what these 700 million users are actually doing with AI—and why 73% of usage has nothing to do with work. The most popular use cases reveal a technology that’s changing not how we work, but how we live.

Read the series:

• Part 1: Who’s Using AI? (you are here)

• Part 2: The Most Popular AI Use Cases at Home (coming next)

• Part 3: Emerging AI Applications That Are Growing Fast (coming soon)

• Part 4: How AI Is Actually Being Used at Work (coming soon)

The past few weeks have been incredibly disruptive for my colleagues at Microsoft, coming on top of an incredibly challenging period where the tech sector has seen over 500,000 job losses since 2022.

At least 95,000 workers at U.S.-based tech companies were laid off in 2024, with 2023 seeing even higher numbers, and layoffs continuing into 2025 with over 75,000 people already impacted.

The question on everybody's mind: Are we seeing AI's first real impact on jobs?

As AI technology has improved dramatically over this same period, it's disrupting core aspects of all product development roles - across both Product Manager and Software Engineer alike.

Satya Nadella says 30% of code at Microsoft is now written by AI. Cursor AI hit $100M ARR in just 12 months - the fastest-growing SaaS company in history while No-code tools like Lovable, Bolt, and Replit are making app deployment accessible to anyone.

If anyone can now create prototypes, design, and deploy within hours with little tech experience, what does this mean for traditional PM and engineering roles?

As someone interested in the future of work and working on a tech product about user skills, I analyzed what the top industry experts are saying and made some predictions about how these roles will evolve.

My Take: 4 Roles That Will Thrive

A job will remain where there's a job to be done that AI cannot do better than humans. AI excels at implementation and technical skills, but decision-making will rest with people who can assimilate all inputs and exercise judgment.

What endures—and actually compounds—is human decision making, collaboration, connection, creativity, and trust. AI can draft a plan, summarize a call, even write a test; but it cannot earn stakeholder trust, resolve ambiguity, negotiate trade-offs, or make values legible in guardrails.

The cheaper the doing gets, the more those human muscles decide what should be done.

1. The Tastemaker:

Aparna Chennapragada (Microsoft CPO) says the PM will evolve to be a "tastemaker". In a world of infinite ideas and quick apps springing up all the time, the person who can best judge what will resonate with potential customers, what the market seems to want but does not yet have, and how to position it for highest growth - will always be valuable. We are already seeing companies such as Airbnb and ElevenLabs collate the product management and marketing roles into one. This requires a fundamentally human skill of understanding customer needs and sales, which is not replaced by AI. For those in this track, I think human connection and empathy is your biggest strength.

I’d also loop Design & UX and the market-facing technical roles into this lane. When prototyping is near-instant, design doesn’t shrink; it compounds - the tastemaker will have to decide what UX communicates the “vibe” of the product most effectively with AI in the loop.

And on the go-to-market side, Solutions Architecture or Sales Engineering or DevRel become force multipliers. Sure, AI can demo and scaffold, but humans still do the problem framing, integration choices, and trust-building with customers.

The “tastemaker” is increasingly the connective tissue between what we could build and what the market will actually love.

2. The Product Engineer:

Why do we need teams of PMs, engineers, and designers in a world where apps can be created and deployed from a single prompt? The future probably has teams of 1-2 "product engineers" who use AI agents in all aspects of their workflow to quickly code and deploy products - owning all aspects of design, strategy of what to prioritize, and development.

Given the closeness to product development, I think this will probably be a natural transition for SWEs to take on more product knowledge and ownership rather than for a product manager; and probably is the next step for today's 10x engineers.
To make this real: the job shifts from typing to specifying, composing, and evaluating. That pulls QA / SDET / Test into the product engineer’s day-to-day—golden datasets, adversarial prompts, and regression every time a model updates, not just checklists. Add basic technical program management into this seat—agent-ops, versioning, staged rollouts, evaluations as gates, and cross-team choreography when a model update lands.

It also pulls in Data / ML / LLMOps / Evaluation: fewer train-from-scratch heroics, more retrieval/grounding, telemetry, and eval harnesses that keep the system honest. You’ll even see product specialists focused on ML/AI where interpreting model behavior and data quality becomes its own field.

And because AI will eat Tier-0/1 support, Support / Customer Success moves up-stack into escalations, empathy, exception handling—and curating the knowledge that trains your own agents. One person can wear several of these hats now; AI makes that realistic.

3. The Tech/Platform Lead:

The engineer will evolve from just being a coder to someone exercising their knowledge of the entire tech stack - one who understands the implications, advantages and tech debt of each architecture decision - how that impacts performance, security, cost and reliability - and can make a call on which system to use in the product.

While AI can write code, it can't make complex system design decisions that balance competing technical constraints like choosing between databases, architecting systems that scale, or making security vs. performance trade-offs. This engineering expertise is and will remain valuable.
Reliability becomes a product feature and the “paved road” matters—latency p95/p99, rollout safety, and cost-to-serve are on this team’s scoreboard.

And because roles are collapsing, this person will also be accountable for the second aspect of the stack—the tech people: recruiting and managing engineers, keeping high-level visibility across the project/product/company, and giving everyone the tooling to move fast without setting pagers on fire at 3 a.m.

4. The Strategic Product Leader:

For bigger products and bigger organizations developing several products - the skill to manage hundreds of humans will always be valuable. AI will not replace raw decision making, especially where capital and jobs are impacted - therefore the strategic product leader will be the one to make decisions about which bets to make.

Note that the leader must lead via the strength of their expertise, and people's confidence in their decision making rather than title alone - so developing the decision making muscle by analyzing what product bets you have already made is key. This seems like the natural path for Senior PMs and engineers, and for great product leaders in the industry - I don't think this role is going anywhere.
Note that Financial operations and everything that it impacts in AI- tokens, context windows, caching, routing, and vendor economics are now first-order strategy levers, not back-office chores. Higher autonomy for agents will come, but humans still own consequence: capital allocation, guardrails, incident response, and when to pull the plug.

The takeaway

These roles are quite different from current PM and SWE expectations. I think the roles aren't disappearing - they are evolving.

And with AI lifting a lot of the doing, roles are collapsing into one—a single capable person can credibly wear multiple hats. That only works if people stay curious and learn beyond their current box; the good news is AI makes that easier than ever.

What stays stubbornly human
What endures—and actually compounds—is human decision making, collaboration, connection, creativity, and trust. AI can draft a plan, summarize a call, even write a test; it cannot earn stakeholder trust, resolve ambiguity, negotiate trade-offs, or make values legible in guardrails. The cheaper the doing gets, the more those human muscles decide what should be done.

My sources: What Industry Leaders Are Saying

Satya Nadella (Microsoft CEO): Paraphrase about agent workflows and “democratized” product ideation. For his words on agents “independently and proactively orchestrat[ing] tasks,” see the Build 2024 keynote transcript. msftstories.thesourcemediaassets.com

Aparna Chennapragada (Microsoft CPO): “If you aren’t prototyping with AI, you’re doing it wrong.” — Lenny’s Podcast episode + write-up. Lenny's Newsletter

Winston Tang (LeetCode Founder): Argument that coding is a subset of the engineering job and AI doesn’t erase the need for human problem-solving — Yahoo Finance interview. Yahoo Finance

Jensen Huang (NVIDIA CEO): “It is our job to create computing technologies that nobody has to program and that the programming language is human…” — NVIDIA company blog post from World Government Summit remarks. NVIDIA Blog

Sundar Pichai (Google CEO): Paraphrase about urgency + AI redefining workflows. See Google I/O 2025 keynote remarks (official Google blog transcript / video). blog.googleYouTube

Scott Wu (Cognition CEO): “Devin is a teammate… ready to build alongside you or independently complete tasks for you to review.” — Cognition Labs announcement. Cognition

Brad Lightcap (OpenAI COO): “No reason to believe AI will destroy jobs” in the near term — Bloomberg video interview. Bloomberg.com

Aravind Srinivas (Perplexity CEO): On the edge shifting from answers to asking the right questions — TED AI talk transcript. TED

Subscribe now