AI's Impact on Knowledge Worker Productivity

Is AI delivering real productivity gains? What's the ROI so far? Hot takes abound, but data have been scarce. Noam Segal and I took it upon ourselves to find out what’s actually happening on the ground by running one of the largest independent, in-depth surveys on how AI is affecting productivity for tech workers (1,750 respondents). We surveyed product managers, engineers, designers, founders, and others about how they’re using AI at work. tl;dr: AI is overdelivering. 1. 55% of respondents say AI has exceeded their expectations, and almost 70% say it’s improved the quality of their work. 2. More than half of respondents said AI is saving them at least half a day per week on their most important tasks. We’ve never seen a tool deliver a productivity boost like this before. 3. Founders are getting the most out of AI. Half (49%) report that AI saves them over 6 hours per week, dramatically higher than for any other role. Close to half (45%) also feel that the quality of their work is “much better” thanks to AI. 4. Designers are seeing the fewest benefits. Only 45% report a positive ROI (compared with 78% of founders), and 31% report that AI has fallen below expectations, triple the rate among founders. 5. Engineers have accepted AI as a coding partner and now want it to handle the more boring (but necessary) work of building products: documentation, code review, and writing tests. 6. n8n is currently dominating the agent landscape, though actual adoption of agentic platforms in 2025 has been slow. 7. A whopping 92.4% of respondents report at least one significant downsides to using AI tools. There’s definitely room for improvement. Here's the full report: https://lnkd.in/gR5G88yA Inside: - What exactly AI is doing for people, function by function? - Where are the biggest opportunities for AI startups? - Which AI tools have product-market fit? - The downsides of AI productivity - Bonus: The state of agentic AI: promise outpaces practice - What this all means - Appendix: Who took this survey

A new study from Stanford University and Carnegie Mellon University offers an important signal about the future of knowledge work. The paper "How Do AI Agents Do Human Work? Comparing AI and Human Workflows Across Diverse Occupations" evaluates how well AI agents can execute realistic, multi-step computer-based tasks. These were not short prompts, but full workflows: data interpretation, analysis, tool use, code generation, visualization, and written reporting. Link to the paper in the comments. Key results: ⚡ Agents successfully completed tasks covering approximately 72% of computer-using occupations ⚡ On comparable work, agents were roughly 88% faster ⚡ Their cost of execution was roughly 90–96% lower ⚡ Agent workflows exhibited ~83% structural similarity to human workflows, preserving ~99.8% of step order ⚡ Agents consistently chose to execute tasks programmatically, writing code even for activities humans typically do in UI (e.g., slide layout, formatting, design) In practice, agents behaved more like autonomous junior engineers than chat interfaces: installing packages, retrying errors, and generating multi-artifact outputs (analysis, charts, narrative). A key weakness was also noted: when inputs failed (e.g., a document could not be read), agents sometimes sourced alternative data without flagging the substitution. This underscores the need for strong verification, provenance, and oversight. Separately, market activity is emerging that reflects this shift. Mercor, a fast-growing San Francisco startup now valued at $10B, has built a labor platform focused on training AI systems. It recruits highly skilled professionals, including PhDs, lawyers, and financial analysts , and compensates them up to $200/hour to teach frontier models from OpenAI, Meta, and others. In Mercor’s model, humans are not performing the work directly. Instead, they provide judgment, nuance, and domain expertise so that AI systems can perform knowledge work more effectively over time. The company has interviewed more than 100,000 white-collar workers and reports revenue run-rate of roughly $500M. Link to the article in the comments. Taken together, the research and the market response suggest a meaningful shift: 💡 Humans increasingly provide judgment, supervision, training, and narrative context. 💡 AI systems increasingly perform the execution steps within knowledge workflows. This has implications across all sectors. We may all need to consider how work is decomposed, how domain knowledge is captured, and how verification frameworks ensure trust and compliance. A useful question for leaders may be: How should we prepare for a world in which AI agents perform much of the execution, and humans focus on design, oversight, and decision-making? This appears to be where the frontier is heading.

AI Won’t Just Boost Productivity. It Will Flatten Your Org Chart. Everyone believes AI simply boosts productivity. They’re missing the bigger picture. Generative AI isn’t just making tasks faster—it’s fundamentally redefining what tasks are essential and who performs them. They’ll argue AI can’t replace core human capabilities like leadership, creativity, and collaboration. Maybe they had a point—until tools like GitHub Copilot entered the scene and proved otherwise: as demonstrated in recent research by Harvard Business School (Hoffmann et al., 2025) 🔴 Traditional Knowledge Work: • Loaded with constant project management distractions • Often bogged down by collaborative friction and coordination delays • Primarily focused on established routines and known tasks (exploitation) • Dominated by hierarchical structures and top-tier talent acting as gatekeepers • Reliant heavily on frequent, time-consuming meetings and manual oversight 🟢 Generative AI-Driven Work: • Shifts attention decisively toward high-value, core creative and strategic tasks • Eliminates much of the collaborative friction, dramatically enhancing independent, focused productivity • Drives substantial exploration, experimentation, and innovation, fostering continuous growth • Democratizes contribution, significantly boosting lower-ability workers’ effectiveness • Empowers talent at all levels, reducing dependency on a few critical gatekeepers Think about it: GitHub Copilot alone increased coding activity by 12.4%, significantly reduced project management overhead by nearly 25%, and encouraged teams to explore new, innovative projects. These findings are detailed in the working paper “Generative AI and the Nature of Work” by Hoffmann, Boysel, Nagle, Peng, and Xu (2025), which provides extensive empirical evidence supporting these transformative impacts. This transformation isn’t incremental, it’s revolutionary. It’s like Slack, but instead of improving communication, it virtually removes the need for it altogether by allowing individuals to work autonomously yet effectively.

The report titled "Estimating AI productivity gains from Claude conversations" by Anthropic, released in November 2025, provides valuable insights into the impact of AI on productivity. Some highlights: AI is applied to substantial work: The median task handled with Claude would take ~1.4 hours without AI, indicating use on meaningful professional tasks rather than trivial micro-work . Time savings are large but uneven: Median estimated savings are ~80–84%, concentrated in reading, synthesis, and writing tasks; tasks requiring physical presence or quick expert judgment see much smaller gains. Higher-wage roles benefit more: Management, legal, and analytical occupations both use AI on longer tasks and capture higher economic value from time saved, amplifying productivity effects. Productivity gains are highly concentrated: Software developers, managers, marketers, customer service reps, and teachers account for most of the estimated economy-wide impact, while sectors like construction, restaurants, and in-person healthcare see little effect. Acceleration creates bottlenecks: Tasks that AI does not speed up, such as supervision, travel, or enforcement, become the dominant constraints within jobs, limiting overall productivity gains. The 1.8% productivity estimate is an upper bound: It assumes universal adoption, static workflows, and no time spent on validation, likely overstating near-term gains even if long-term effects could be larger. Measurement is the key contribution: The report’s main advance is a scalable method for tracking AI productivity using real usage data, enabling longitudinal analysis as tasks, models, and adoption evolve.

Generative AI is often framed as either skill-biased or deskilling. Evidence has been limited on what happens when people with very different educational backgrounds perform the same task with equal access to the tool. A recent randomized experiment with 1,174 adults asked participants to solve a workplace-style business problem with or without an AI assistant. The baseline gap was large: higher-education participants outperformed lower-education participants by 0.548 sd. With AI access, the difference fell to 0.139 sd -- put differently, roughly three quarters of the productivity gap disappeared during the task. Both groups improved, but gains were larger for less-educated participants. The mechanism is not that skills vanished. Instead, the tool substituted for specific cognitive inputs such as structuring an argument, organizing information, and producing written output. Once AI access was removed, performance differences widened again. This pattern suggests a different margin of technological change than earlier waves of automation. AI compresses effective productivity on a given task while leaving underlying human capital differences intact: • Education increasingly affects how people guide, edit, and evaluate AI rather than whether they can complete the task at all • Access and workplace policy may matter more for inequality than raw technical capability • Job design, not just automation, will determine labor market outcomes The distributional effects of AI will be decided inside organizations, through task allocation and permissions, rather than solely through technological capability. #AI #LaborEconomics #Productivity #Inequality #FutureOfWork

Does generative AI actually make knowledge workers more productive? That’s the billions-dollar question. And this now-classic RCT offers one of the best answers we have. It’s easy to find newer studies, but I keep coming back to this one for its experimental rigor and real-world relevance. And even though the capabilities of LLMs have progressed significantly since this study was run, its findings hold. The core insight of a jagged AI frontier remains highly relevant to today's models: LLMs perform exceptionally well on some tasks but poorly on others, even when those tasks seem similarly complex. Researchers tested GPT‑4 with >750 professionals, assigning them real-world business tasks like writing a press release, brainstorming product ideas, analyzing data, and responding to frustrated clients. Some participants had access to GPT‑4; others didn't. On tasks inside the jagged frontier (like idea generation, summarizing, editing), GPT‑4 was a clear win. Users completed 12% more tasks, worked 25% faster, and produced output rated over 40% higher in quality. The biggest gains came from those in the bottom half of baseline performance. But on tasks outside the frontier, AI often made things worse. It produced polished, confident responses that sounded impressive, but missed key contextual cues, leading users to make incorrect decisions more often than those who worked without it. It sounded right, even when it wasn’t. Using GPT‑4 dramatically improved performance when producing a fairly straightforward press release, for example, but reduced accuracy when diagnosing a business problem from a mix of qualitative interviews and financial data. So does AI make us more productive? Yes, sometimes dramatically. But not for everyone, and not for everything. GPT‑4 excels at structured, language-based tasks. It struggles when success depends on subtle judgment, synthesis of mixed data, or reading between the lines. In those cases, it can produce confident, polished output that’s just plain wrong. As we integrate AI into more of our work, it’s worth asking whether we are automating the right parts of our work (and therefore our lives). Are the tasks AI struggles with the ones we should be rushing to offload, or perhaps the ones we might want to save for ourselves? After all, "high productivity" is only a win if it moves us toward work that matters. Otherwise, we risk simply doing more of the parts we never meant to prioritize in the first place. This is part of the Friday Findings series from Nuance Behavior, curated research at the intersection of minds and machines. — Fabrizio Dell'Acqua, Edward McFowland III, Ethan Mollick, Dr. H /Hila Lifshitz, Kellogg, Saran Rajendran, Lisa Krayer, PhD, François Candelon & Karim Lakhani (2023). Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality. HBR Working Paper 24-013.

Where are the productivity gains from AI? 🤔 The technology has been advancing rapidly but when do we actually see its impact in the real economy? For the UK, this turning point might be happening right now. 🇬🇧 I was excited to read a new report by Snowflake, based on research conducted by YouGov among 500 senior decision-makers across major UK organisations: ➡️ 23% report that AI is already delivering productivity improvements at scale ➡️ Another 45% see gains emerging in specific use cases ➡️ Just 1% planning to reduce AI spending This is a big deal. We may finally be moving from AI promise to measurable productivity reality 📈 What is slowing AI productivity gains down? The report also highlights something equally important: The main bottlenecks are not technological. Only 19% of organisations cite technology as a key barrier. Instead, the real challenges are poor data quality, organisational silos and, most crucially, a shortage of skilled talent. This strongly aligns with findings from our www.skillscale.org research group at the Oxford Internet Institute, University of Oxford. Firms are desperately looking for AI talent and workers with AI skills 💷 earn ~23% higher wages, 🏡 are 3x more likely to enjoy job perks like remote work, 📩 have higher chances of landing a job in the first place. What should we learn from this? AI is not just a technology story, it’s a skills story. If we want to sustain and scale these productivity gains, the priority is clear: Invest in people, not just in tools. You can find the Snowflake/YouGov report here: https://lnkd.in/e73t_MRd Curious to hear your thoughts—are you already seeing productivity gains from AI in your organisation? 👇

Productivity goes up when you add a second AI tool. It goes up again with a third, but slower. After three, it drops. That's from a BCG study of 1,488 workers, published in HBR. As companies roll out more multi-agent systems, people end up toggling between more tools. Contrary to the promise of freeing time for meaningful work, juggling becomes the work itself. The researchers call the cognitive cost "AI brain fry": 33% more decision fatigue, 39% higher intent to leave. I'd been noticing the same dynamic around me, so it's good to see it named and measured. More tools follow the same logic as any scaling decision: returns diminish, then reverse. Go deep on a limited number. We already have an AI force on the technology side, and within the people function, we've done a lot of experimenting and filtering already. A formal AI scouting team feels like a logical next step: people who understand the domain processes and knowledge, connected to our tech team on one side and to the external landscape on the other. Their job: filter what's coming, test what fits, and protect the rest of the function from adopting everything at once. Scale doesn't mean more tools. It means more from fewer. https://lnkd.in/eBuZcWxg #FutureOfWork #AIStrategy #TalentStrategy

We just ran Guild's first-ever AI-specific productivity survey. 233 learners. And the results sit in stark contrast to what the rest of the industry is reporting. Companies spent 93% of their AI budgets on tooling last year and only 7% on the people using it. No wonder adoption metrics are lagging: -NBER surveyed 6,000 executives — 90% of firms reported no AI impact on productivity. -MIT Media Lab found 95% of organizations saw zero measurable return on GenAI investments. - The San Francisco Fed compared it to replacing a steam engine with an electric motor but leaving the factory floor unchanged. The tools are there. The human capability isn't. Our survey told a very different story: -90% of Guild AI learners use AI tools at least a few times a week at work. -84% report at least one measurable workplace outcome. -73% say their AI education helped them create results at work — not "someday" results, but changes to how they work right now. The difference isn't the tool. It's whether someone actually taught them how to think with it. If your AI strategy is a software rollout without an education strategy behind it, you're not deploying AI. You're just buying licenses. What percentage of your AI budget is going to your people?

From Gallup's State of Global Workplace 2026 released today: Disengaged employees, negative emotions and distracted managers the big obstacles to AI productivity for firms. Link to the report below: "The technology works. Large language models can draft legal contracts, write code and synthesize research at speeds no human team can match. But those gains are not showing up in the bottom line. A recent MIT study found that despite roughly $40 billion in enterprise investment, 95% of organizations have seen zero measurable impact on profits.1 An NBER survey of nearly 6,000 global executives reports that 89% see no effect on labor productivity. In Gallup's own data, only 12% of employees in AI-implemented organizations strongly agree that AI has transformed how work gets done in their organization. So, if the technology isn't the problem, what is? Gallup's data points to an answer the corporate world has largely ignored: the manager. In organizations investing in AI, the strongest predictor of employee adoption, aside from technical integration, is whether their direct manager actively champions it. Even the most sophisticated neural network cannot overcome an indifferent team leader. OpenAI would likely agree. In its 2025 enterprise report, the company states: "The primary constraints for organizations are no longer model performance or tooling, but rather organizational readiness and implementation." The relationship between realized technological gains and great management is not new. A decade ago, researchers at Stanford, Harvard Business School and MIT found that differences in management practices accounted for about 30% of the variation in total factor productivity, the most common measure of the impact of technology on productivity. For decades, organizations worldwide have struggled to manage people effectively. Now, the financial stakes are far higher. Winning the AI revolution will depend not just on the technology you deploy but also on how well you lead the people using it. This report establishes a global baseline for management effectiveness in the AI era. " Full Report: https://lnkd.in/gwa6_Tsr?