01. EXECUTIVE SUMMARY
The $37 Billion Question
In 2025, companies spent $37 billion on generative AI, a 3.2x increase from the $11.5 billion spent just one year earlier, according to Menlo Ventures. Enterprise AI adoption has reached a fever pitch: a Wharton study found that 82% of workers now use generative AI at least weekly, with 46% using it daily. Worker access to AI tools rose by 50% in a single year, according to Deloitte.
The promise is seductive: give your employees AI, and they will become faster, smarter, more capable. The marketing materials from every AI vendor tell the same story: AI as the great equalizer, the tool that turns every junior analyst into a senior strategist, every average writer into a polished communicator, every novice coder into a productive engineer.
But does it? The academic research tells a far more nuanced, and in some cases, deeply uncomfortable story. This article examines what the evidence actually says about AI and employee expertise, drawing on peer-reviewed studies from Harvard, MIT, Stanford, and other leading institutions. The findings should inform how every organization thinks about AI adoption, training, and the future of human capital.
$37B
Enterprise AI spend 2025
82%
Workers using AI weekly
95%
AI projects fail ROI
11%
Motivation drop with AI
02. The Reality
AI Helps the Weakest Most
One of the most consistent findings across AI research is that generative AI disproportionately benefits lower-performing workers. A landmark study published in Science by MIT researchers Shakked Noy and Whitney Zhang examined ChatGPT’s impact on professional writing tasks. The results were striking: AI reduced task completion time by 40% and improved output quality by 18%. But the gains were not evenly distributed. Workers who started with the weakest writing skills saw the largest improvements, while already-skilled writers saw modest gains at best.
MIT Sloan confirmed this pattern in software development: less-experienced developers showed significantly higher productivity gains from AI coding assistants than their senior counterparts. GitHub Copilot users increased their core coding time by 12.4% while cutting time on non-core tasks but the effect was most pronounced for junior developers.
This sounds like good news. And in some ways, it is. But here is the critical distinction that most AI vendors conveniently omit: making a low performer adequate is not the same as making them an expert. AI narrows the gap between the bottom and the middle. It does not lift the middle to the top. The difference between “competent” and “expert” is not speed or polish. It is judgment, context, and the ability to know what the AI cannot tell you.
"ChatGPT substantially raised average productivity: time taken decreased by 0.8 standard deviations and output quality rose by 0.4 standard deviations. The effect was largest for workers with the weakest initial performance."
03. The Jagged Frontier
Where AI Excels and Where It Fails
The most influential study on AI and knowledge work to date was conducted by researchers at Harvard Business School in collaboration with Boston Consulting Group. The experiment involved 758 BCG consultants (approximately 7% of the firm’s individual-contributor consulting workforce) performing realistic tasks that spanned creativity, analytical thinking, writing, and persuasiveness.
The findings introduced a concept that has since reshaped how researchers think about AI capabilities: the “jagged technological frontier.” For tasks that fell within AI’s capability boundary, the results were remarkable: ChatGPT-4 boosted speed by over 25%, human-rated performance by over 40%, and task completion by over 12%.
But for tasks that fell outside the AI frontier, tasks that required deep contextual judgment, nuanced reasoning about ambiguous situations, or integration of information that the model had not been trained on, AI actually decreased performance. Consultants who relied on AI for these tasks performed worse than those who worked without it. The frontier is “jagged” because there is no clean line between what AI can and cannot do. It excels at some surprisingly complex tasks while failing at others that seem deceptively simple.
The “jagged frontier”: AI capabilities are uneven and unpredictable, excelling in some areas while failing in others
Key Finding
Harvard/BCG Study
The study identified 2 distinct patterns of effective AI use. “Centaurs” strategically divided tasks between themselves and the AI, delegating what AI does well and retaining what requires human judgment. “Cyborgs” integrated their workflow with AI at every step, creating a continuous human-machine loop. Both patterns outperformed workers who either ignored AI entirely or delegated everything to it.
Dell’Acqua et al., “Navigating the Jagged Technological Frontier,” Organization Science, 2026
The implication for organizations is profound: deploying AI without training employees where the frontier lies (and where it does not) is not just inefficient. It is actively dangerous. Employees who trust AI on tasks outside its frontier will produce confidently wrong outputs that look polished and professional but are fundamentally flawed.
As the researchers noted, the focus should move beyond the binary decision of adopting or not adopting AI. Instead, organizations must evaluate the value of different configurations and combinations of humans and AI for various tasks within the knowledge workflow.
04. The Motivation Paradox
More Productive, Less Engaged
Even when AI does improve performance, the psychological costs are rarely discussed. Research published in Harvard Business Review by Liu, Wu, Ruan, Chen, and Xie examined what happens to workers’ motivation after collaborating with generative AI. The findings reveal a hidden trade-off that should concern every leader investing in AI tools.
Across 4 studies, participants who collaborated with AI on one task and then transitioned to a different, unaided task consistently reported a decline in intrinsic motivation of 11% and an increase in boredom of 20%. Workers who completed tasks without AI maintained a relatively steady psychological state.
The mechanism is revealing: AI collaboration reduces workers’ sense of control, the feeling of being the primary agent of their work. When people feel they are not fully in charge of the output, it undermines their connection to the task. Transitioning back to solo work restores autonomy but at the cost of enjoyment. Workers regain their independence but feel “less inspired and challenged.”
"While gen AI collaboration boosts immediate task performance, it can undermine workers' intrinsic motivation and creativity creating a hidden cost that organizations rarely account for."
This is not a minor finding. Intrinsic motivation is the engine of deep expertise. People become experts not because they are told to, but because they are genuinely engaged with the challenge of mastering their craft. If AI tools systematically erode that engagement, even while boosting short-term output, the long-term consequences for workforce capability could be severe.
05. Cognitive Atrophy
The Deskilling Risk
A study from the MIT Media Lab reported that “excessive reliance on AI-driven solutions” may contribute to “cognitive atrophy“, a shrinking of critical thinking abilities. While the study is preliminary, it echoes a growing body of concern from researchers across Harvard, MIT, and Microsoft Research.
Harvard’s Tina Grotzer, a principal research scientist in education at the Graduate School of Education, argues that human minds are “better than Bayesian” in many ways. Our somatic markers enable quick, intuitive leaps. We can detect critical distinctions or exceptions to patterns that a purely statistical approach would average across. While AI can offer analogies, it cannot reason analogically. It cannot make the creative conceptual leaps that drive genuine insight and innovation.
Christopher Dede, a senior research fellow at Harvard’s Graduate School of Education, frames the risk through a memorable metaphor: Athena, the Greek goddess of wisdom, always had an owl on her shoulder. “The owl sits on your shoulder and not the other way around.” When AI does your thinking for you, whether through auto-complete or by writing the first draft that you merely edit, it undercuts both critical thinking and creativity.
Microsoft Research’s own review of overreliance on AI found that users who see AI perform well early on develop “automation bias and complacency,” making significantly more errors due to positive first impressions. The pattern is insidious: the better AI appears to work, the less critically people evaluate its output and the more vulnerable they become to its failures.
06. What AI can do
What AI Can Actually Do Well
None of this means AI is useless, far from it. The research is clear that AI delivers genuine value when applied to the right tasks, with the right expectations, and with appropriate human oversight. The key is understanding exactly where AI adds value and where it creates risk.
01
Accelerating Routine Tasks
Drafting emails, summarizing documents, generating boilerplate code, formatting reports. AI excels at tasks with clear patterns and low ambiguity.
02
Equalizing Access to Knowledge
Junior employees can access institutional knowledge faster. AI acts as a 24/7 mentor for procedural questions and standard practices.
03
First-Draft Generation
Creating starting points for documents, presentations, and analyses that experienced professionals can then refine and improve.
04
Data Processing at Scale
Analyzing large datasets, identifying patterns in structured data, and generating statistical summaries faster than any human analyst.
05
Brainstorming & Ideation
Generating diverse options and alternative perspectives. AI is useful for expanding the solution space before human judgment narrows it.
06
Quality Assurance & Review
Checking code for bugs, reviewing documents for consistency, flagging potential errors, tasks where pattern matching is the core skill.
Stanford’s Human-Centered Artificial Intelligence Institute confirmed that AI agents play a “supportive role in the workplace,” relieving workers of low-value or tedious tasks rather than replacing them. The St. Louis Federal Reserve estimated that workers using generative AI saved 5.4% of their work hours per week, a meaningful but modest gain that translates to approximately a 1.1% increase in overall productivity.
07. What AI cannot do well
What AI Cannot Do and Probably Will Not Soon
Harvard’s Fawwaz Habbal, a senior lecturer at the John A. Paulson School of Engineering and Applied Sciences, puts it bluntly: AI machines “rely on data that has been created by humans, and that data is the same, more or less, across the different AI platforms. When you ask a question to different AI platforms, most of the time their answers are very similar because the database is the same.”
This observation points to a fundamental limitation: AI can tell you how to put things together, but it cannot help you build something that relates to a human context. Machine learning depends on statistical adjustments, whereas humans self-organize life in relation to meaning. The distinction matters enormously for the kinds of decisions that define business success or failure.
01
Strategic Judgment Under Ambiguity
AI cannot navigate situations where the problem itself is poorly defined, where stakeholder interests conflict, or where the right answer depends on organizational context that exists nowhere in the training data.
02
Ethical & Moral Reasoning
AI can process data about ethical frameworks, but it cannot make moral judgments. It lacks the human experience, insight, and ethical intuition that inform decisions about what should be done, not just what can be done.
03
Genuine Innovation
AI recombines existing patterns. It does not create fundamentally new categories of thought. The breakthroughs that transform industries: the iPhone, the theory of relativity, the double helix… require conceptual leaps that statistical pattern matching cannot produce.
04
Relationship & Trust Building
Client relationships, team dynamics, organizational culture, and stakeholder management depend on empathy, emotional intelligence, and the ability to read unspoken signals, capabilities that remain firmly in the human domain.
05
Analogical Reasoning
While AI can offer analogies, it cannot reason analogically. It cannot make the creative conceptual leaps that connect disparate domains and drive genuine insight.
"AI can engage in processes that resemble critical thinking (data analysis, problem-solving, and modeling) but critical thinking requires the human experience, the human insight, and ethics and morality."
08. The 95% Failure Rate
Why Most AI Projects Miss the Mark
MIT researchers studied 300 public AI initiatives, interviewed 150 executives, and surveyed 350 employees to assess what they called the “no hype reality” of enterprise AI. Their conclusion: 95% of enterprise AI projects fail to turn a profit. For context, the failure rate for traditional IT projects is approximately 25%.
Deloitte’s 2025 survey of 1,854 executives confirmed the pattern: despite rapidly rising AI investment, ROI remains elusive. Bain’s executive survey found that while many AI use cases met or exceeded expectations, only 23% of respondents could tie the work directly to measurable business outcomes.
The disconnect is not primarily technical. It is organizational. Companies deploy AI tools without redesigning workflows, without training employees on when to use AI and when not to, and without establishing clear metrics for what success looks like. They treat AI as a plug-in rather than a transformation and the results reflect that approach.
09. Mindset
The Right Way to Think About AI and Expertise
Harvard’s Dan Levy, a senior lecturer at the Kennedy School and co-author of “Teaching Effectively with ChatGPT,” captures the essential distinction: “If a student uses AI to do the work for them, rather than to do the work with them, there’s not going to be much learning.” The same principle applies to every employee in every organization.
The research converges on a framework for thinking about AI and expertise that has 3 dimensions:
01
Strategic Judgment Under Ambiguity
AI cannot navigate situations where the problem itself is poorly defined, where stakeholder interests conflict, or where the right answer depends on organizational context that exists nowhere in the training data.
02
Invest in Frontier Literacy
Every employee who uses AI needs to understand where the jagged frontier lies in their specific domain. This is not a one-time training, the frontier shifts as AI capabilities evolve. Organizations need ongoing programs that help workers identify which tasks are inside the frontier (safe to delegate) and which are outside it (dangerous to delegate).
03
Protect the Expertise Pipeline
If junior employees never struggle with hard problems because AI solves them first, they will never develop the deep expertise that organizations need in their senior leaders. The motivation research makes this explicit: AI collaboration reduces intrinsic motivation by 11%. Over a career, that compounding effect could be devastating.
"At the end of the day, if you think you're in school to produce outputs, then you might be OK with AI helping you produce those outputs. But if you're in school because you want to be learning, remember that the output is just a vehicle through which that learning is going to happen."
10. Framework
A Practical Framework for AI-Augmented Expertise
Based on the research, we propose a framework that organizations can use to deploy AI in ways that genuinely build capability rather than erode it. The framework distinguishes between 4 zones of AI application, each requiring a different approach.
Zone
Task Characteristics
AI Role
Human Role
Automate
Repetitive, rule-based, low ambiguity
Full execution with spot checks
Quality assurance, exception handling
Accelerate
Structured, moderate complexity
First draft, data processing, options generation
Review, refine, apply judgment
Augment
Complex, requires domain expertise
Research assistant, sounding board
Lead thinking, make decisions, own outcomes
Protect
Ambiguous, high-stakes, relationship-dependent
Minimal or none. Risk of degrading performance
Full ownership, deep expertise required
The critical insight is the fourth zone: “Protect.” These are the tasks where AI involvement actively degrades performance, the tasks outside the jagged frontier. Every organization has them, and identifying them is as important as identifying where AI can help. The Harvard/BCG study demonstrated that consultants who used AI on “Protect” zone tasks performed worse than those who worked without it.
11. The Redex Perspective
Human. Intelligent. Measurable.
At Redex, we work with organizations across industries to deploy AI in ways that genuinely build capability rather than create dependency. Our experience has taught us that the difference between successful and failed AI adoption is rarely about the technology. It is about how the technology is integrated into human workflows and decision-making processes.
We believe that AI should make your organization smarter over time, not just faster today. We advise. We build. We do both. We help organizations map their jagged frontier, design AI integration strategies that protect expertise development, and build the platforms and processes that turn AI from a productivity tool into a genuine competitive advantage.
The question is not whether to adopt AI. That debate is over. The question is whether you will adopt it in a way that builds lasting capability or one that creates a fragile dependency on a tool your competitors also have access to.
“AI should make your organization smarter over time, not just faster today. The difference between successful and failed AI adoption is rarely about the technology. It is about how the technology is integrated into human workflows.”
Key Takeaways
- AI disproportionately helps lower-performing workers but making someone adequate is not the same as making them an expert. The gap between competent and expert is judgment, not speed.
- The "jagged technological frontier" means AI excels at some tasks and fails at others and the boundary is unpredictable. Organizations must map this frontier for their specific domain.
- AI collaboration reduces intrinsic motivation by 11% and increases boredom by 20%. Over time, this erosion of engagement threatens the development of deep expertise.
- 95% of enterprise AI projects fail to deliver ROI not because the technology is bad, but because organizations treat AI as a plug-in rather than a transformation.
- The best AI users are "Centaurs" and "Cyborgs". They strategically divide work between human and AI, maintaining active judgment about what to delegate.
- Protect the "Protect Zone". Identify tasks where AI involvement actively degrades performance, and keep humans fully in charge of those decisions.
- Invest in frontier literacy: every AI-using employee needs ongoing training on where AI helps and where it hurts in their specific role.
Ready to build an AI strategy that creates lasting capability?
We help organizations map their jagged frontier, design AI integration strategies that protect expertise development, and build platforms that turn AI into a genuine competitive advantage.
REFERENCES
- 2026. Dell’Acqua, F. et al. “Navigating the Jagged Technological Frontier,” Organization Science.
- 2023. Noy, S. & Zhang, W. “Experimental Evidence on the Productivity Effects of Generative AI,” Science.
- 2025. Liu, Y. et al. “Gen AI Makes People More Productive and Less Motivated,” Harvard Business Review.
- 2024. MIT Sloan. How Generative AI Affects Highly Skilled Workers.
