Skip to main content
Research Feature

How AI Reshapes the Tasks Behind Every Knowledge Job

A data-grounded visualization of how AI adoption progressively replaces tasks within jobs — and what happens when a job crosses the threshold where it no longer exists in its current form.

Brad Anderson
Brad Anderson
CTO / Founder · Fruition · Published July 3, 2026 · Updated July 4, 2026

AI isn't replacing jobs. It's replacing tasks within jobs — one at a time, year by year, until the job as currently defined no longer makes sense as a standalone role. Then the person shifts to something new.

This sounds abstract until you watch it happen. So that's what this article does. Below are two interactive charts. The first shows a single job broken into its tasks, with AI capability filling each one over time. The second shows all 16 knowledge worker jobs on one timeline, so you can see which roles cross the transformation threshold first.

The model is grounded in three frameworks: Rogers' diffusion theory Diffusion of Innovations (1962) Everett Rogers' foundational work showing that new technologies spread through a population in five adopter categories: innovators (2.5%), early adopters (13.5%), early majority (34%), late majority (34%), and laggards (16%).
Read the article →
(how technologies spread), the Bass model A New Product Growth Model (1969) Frank Bass' mathematical model of how new products get adopted, using two parameters: p (innovation coefficient, avg 0.03) and q (imitation coefficient, avg 0.38). The model produces the S-shaped adoption curve that matches real-world diffusion patterns.
Read the article →
(the math behind the spread), and the task-based theory of automation The Skill Content of Recent Technological Change (2003) Autor, Levy, and Murnane's breakthrough paper showing that computers substitute for workers on routine tasks (following explicit rules) and complement workers on non-routine tasks (problem-solving and communication). The foundation of task-based automation theory.
Read the article →
(why AI replaces tasks, not jobs). Every claim is backed by research you can hover over to read.

How to use this page. Start with the first chart below — pick a job from the dropdown, drag the year slider, and watch AI fill the task bars. Then scroll to the second chart to see all 16 jobs compared. The narrative between the charts explains what you're looking at and why it matters.

First: how fast is AI spreading?

Before we look at jobs, we need to understand the adoption curve. The dark chart below shows enterprise AI adoption from 2018 to 2035, modeled with the Bass diffusion equation A New Product Growth Model (1969) Frank Bass' mathematical model of how new products get adopted, using two parameters: p (innovation coefficient, avg 0.03) and q (imitation coefficient, avg 0.38). The model produces the S-shaped adoption curve that matches real-world diffusion patterns.
Read the article →
. The colored bands are Rogers' five adopter categories Diffusion of Innovations (1962) Everett Rogers' foundational work showing that new technologies spread through a population in five adopter categories: innovators (2.5%), early adopters (13.5%), early majority (34%), late majority (34%), and laggards (16%).
Read the article →
: innovators, early adopters, early majority, late majority, and laggards.

The Stanford AI Index (2025) AI Index Report (2025) Stanford HAI's annual report tracking AI adoption, investment, and capability. Key data: 78% of organizations use AI in 2024 (up from 55% in 2023), $33.9B in generative AI investment, and a 280× drop in inference costs from 2022 to 2024.
Read the report →
reports that 78% of organizations were using AI in 2024, up from 55% in 2023. That puts us past the chasm — the dangerous gap between early adopters and the early majority where most technologies die. AI didn't die. It crossed.

The blue dot on the curve shows where we are today. Everything to the left is observed data; everything to the right is projection.

Enterprise AI adoption (Bass model, p=0.07, q=0.45)

Innovators (2.5%) Early Adopters (13.5%) Early Majority (34%) Late Majority (34%) Laggards (16%)

Here's why this matters for jobs: the adoption curve is the engine that drives task automation. When AI adoption is at 15% (2020), only the most obvious routine tasks are affected. When it hits 78% (2024), entire task categories start moving from human to machine. And as it pushes past 90% into the late majority and laggards, even complex tasks that seemed permanently human begin to shift.

The chart you're about to see takes this adoption curve and applies it to the specific tasks that make up a single job. Pick a job. Drag the slider. Watch what happens.

Now: one job, broken into its tasks

The chart below shows a single job decomposed into 8–9 core tasks, drawn from O*NET occupational data O*NET OnLine The U.S. Department of Labor's occupational database, providing detailed task decompositions for 900+ occupations. Used in this visualization to break each job into its constituent tasks with time-allocation weights.
Read the article →
. Each bar is a task. The blue portion is what humans still do; the lighter blue is what AI augments; the dark portion is what AI handles on its own.

Try it. Pick "Software Developer" from the dropdown. The slider is set to 2026. You'll see that writing boilerplate code is 98% AI — nobody writes code from scratch anymore. But requirements and stakeholder communication is still entirely human. The job overall is 80% automated, which means it has already crossed the transformation threshold. The developer's role has shifted from writing code to reviewing AI output and talking to stakeholders.

Now switch to "Paralegal" or "Copywriter" and drag the slider forward to 2030. Watch how fast the bars fill. Then try "Sales Representative" — notice how much stays blue, because relationship-building and negotiation are still fundamentally human. And look at the bottom of each chart: the green bars marked ✦ are new tasks that didn't exist before AI. Drag the slider back to 2022 — they show "Did not exist yet" as a faded placeholder. Drag forward to 2026 and they appear, growing as AI creates the need for human work like AI orchestration, AI content governance, and AI bias auditing.

2020202520302035

Software Developer

AI is augmenting routine tasks

14%
tasks AI handles
Human-performed tasks AI-augmented tasks AI-automated tasks New human tasks (created by AI) Task didn't exist yet

What the threshold means. When the total automation percentage (the big number above the chart) crosses 50%, the job has transformed. This doesn't mean the person is unemployed — it means their role has shifted. The Acemoglu-Restrepo framework Automation and New Tasks (2019) Acemoglu and Restrepo extend their framework, showing that automation displaces workers but the creation of new tasks reinstates labor. Historical evidence suggests productivity growth from automation benefits workers only when new task creation keeps pace.
Read the article →
calls this the balance between displacement (tasks moving to AI) and reinstatement (new tasks emerging for humans). The outlined bars at the bottom of the chart — "AI orchestration," "AI content governance," "AI bias auditing" — are the reinstatement effect in action: new work that didn't exist before AI created the need for it.

The 2026 baseline. Each task has a 2026 baseline — the AI capability level for that task right now, grounded in what AI tools actually do today, not theory. "Writing boilerplate code" is at 98% because Copilot, Cursor, and Claude handle code generation. "Architecture and system design" is at 80% because AI assists heavily but doesn't own the decisions. "Requirements and stakeholder communication" is near 0% because AI barely touches it. The key insight: writing code is 98% AI does not mean the software developer job is 98% automated. Coding is only 15% of the job's task weight. The other 85% — architecture, requirements, communication, judgment — determines the overall number.

The big picture: all 16 jobs on one timeline

Now step back. The chart below shows every knowledge worker job at once, from 2020 to 2035. Each line is a job. The dashed line at 50% is the transformation threshold. Lines that cross it — and thick out — have shifted to a new task profile. Lines that stay below it are still mostly human.

Try it. Hit "Animate All" and watch the lines climb. The first jobs to cross are the ones weighted heavily toward routine information work — copywriter, paralegal, accountant, customer support. The last to cross are the ones weighted toward judgment and relationships — sales, project manager, data scientist.

The ranked table below the chart updates with the year. It shows every job sorted by how much AI handles at that point in time. The ⚠ marker means the job has crossed the transformation threshold.

Year: 2026
Engineering & Development Marketing & Creative Data & Analysis Legal & Compliance Operations & Admin
50% job transformation threshold Current year position

Ranked by AI automation at 2026

What this tells us. At 2026, most jobs have already crossed or are near the threshold. The Eloundou et al. GPTs are GPTs (2023) Eloundou, Manning, Mishkin, and Rock assessed every O*NET occupation against LLM capabilities. Found that 80% of the U.S. workforce could have ≥10% of tasks affected, 19% face ≥50% task impact, and 47-56% of all tasks could be completed faster with LLM-powered software. Published in Science, 2024.
Read the article →
research predicted this: 80% of the workforce has at least 10% of tasks affected, and 19% face 50% or more. The McKinsey (2023) The Economic Potential of Generative AI (2023) McKinsey estimates that generative AI could automate work activities absorbing 60-70% of employee time today, force 12 million U.S. workers to switch occupations by 2030, and add $2.6-4.4 trillion annually to the global economy.
Read the article →
report estimated that 60–70% of employee time could be automated — and 12 million U.S. workers may need to switch jobs by 2030.

But here's the critical nuance: crossing the threshold doesn't mean the job disappears. Brynjolfsson's customer support study Generative AI at Work (2023) Brynjolfsson, Li, and Raymond studied a staggered rollout of a generative AI assistant to 5,179 customer support agents. Found a 14-15% average productivity increase, with 34% improvement for novices and minimal impact on experienced workers. Published in QJE, 2025.
Read the article →
showed that AI made agents 14% more productive on average and 34% more productive for novices — but didn't eliminate a single agent. The job transformed: agents spent less time typing responses and more time handling complex escalations and de-escalation. That's the reinstatement effect in practice.

The WEF Future of Jobs Report (2025) Future of Jobs Report (2025) The World Economic Forum's survey of 1,000+ employers across 22 industries and 55 economies. Projects 170 million new roles created and 92 million displaced by 2030, with 86% of businesses reporting AI is transforming their operations.
Read the article →
puts numbers on this: 170 million new roles created, 92 million displaced, for a net gain of 78 million. The question isn't whether AI eliminates jobs. It doesn't. The question is whether the new tasks it creates are ones that humans can step into quickly enough. That's the central economic question of the 2020s.

In Practice

How Fruition is living this cycle right now

The displacement → reinstatement cycle isn't theoretical for us. We've watched tasks move from labor to capital (AI), and we've watched new tasks emerge that didn't exist before — which are now becoming the core of the job.

The Acemoglu-Restrepo framework Automation and New Tasks (2019) Acemoglu and Restrepo extend their framework, showing that automation displaces workers but the creation of new tasks reinstates labor. Historical evidence suggests productivity growth from automation benefits workers only when new task creation keeps pace.
Read the article →
describes a cycle: automation displaces human tasks (labor → capital), but it also creates new tasks that only humans can do (capital → labor). Those new tasks eventually become routine, get automated, and the cycle repeats. It's not a line going one direction. It's a loop.

We've been inside this loop for three years. Here's what it looks like at Fruition, role by role — the old tasks that moved to AI, the new tasks that emerged for humans, and where the cycle is starting to repeat.

Role Old task → AI (labor → capital) New human task (capital → labor) Cycle status
Software Developer Writing code from scratch → Copilot, Cursor, Claude now generate 98% of boilerplate AI output review, architecture decisions with AI assistance, eval engineering, AI orchestration Already shifted
WebOps / Support Manual ticket triage, hosting checks, deployment steps → FCP automates 95% of routine ops tickets AI monitoring governance, anomaly review, scope management, client strategy Already shifted
Marketing Writing first drafts, keyword research, social scheduling → AI generates content and optimizes SEO Generative Engine Optimization (GEO), brand voice enforcement, AI prompt strategy, creative direction on AI output Already shifted
QA / Testing Writing test scripts, running regression suites → AI generates tests and executes them AI test output validation, exploratory edge-case testing, test strategy for AI-generated code Shifting now
Design Asset creation, resizing, layout variations → AI generates visuals and components AI asset curation, creative direction on AI output, design system governance Shifting now

The cycle is already repeating. Some of the new tasks that emerged in 2023–2024 are starting to shift toward capital too. AI output review — the task developers shifted into when code generation was automated — is itself being augmented by AI code review tools. The developer's role is shifting again: from reviewing AI output to making judgment calls about which AI suggestions to accept, which to reject, and why. That's a higher-level task. And when AI gets good enough at that, a new task will emerge above it.

This is what Acemoglu and Restrepo (2018) The Race Between Man and Machine (2018) Acemoglu and Restrepo's model of automation's dual effects: displacement (tasks move from human to machine) and reinstatement (new tasks emerge for humans). The net effect on wages depends on which effect dominates.
Read the article →
mean by the "race between man and machine." It's not a race with a finish line. It's a cycle where each lap creates new work that didn't exist before — as long as the organization is structured to capture it. The risk isn't that AI takes your job. The risk is that your organization can't create new tasks fast enough to replace the ones AI took.

This is the problem our product strategy is built to solve. If the risk is not creating new tasks fast enough, then the solution is building systems that find, surface, and orchestrate new work automatically. That's the thread connecting our platform roadmap. Unroo™ centralizes task discovery and work orchestration — so when AI frees up capacity, the next task is already queued, not lost in someone's inbox. FCP's "what to work on next" (WTWON) engine scans the infrastructure for emerging issues and creates tasks before they become incidents. Our audit tools continuously scan for ADA gaps, security vulnerabilities, and performance regressions — generating new tasks that didn't exist before the scan found them. Each product is a different lens on the same underlying principle: the organization that wins is the one that can detect new work faster than AI can eliminate old work.

How we turned labor into capital. Three years ago, our developers spent 15% of their time writing boilerplate code. That time is now near zero — the code is generated by AI. But that 15% didn't disappear. It moved to AI output review, architecture decisions, and stakeholder communication — tasks that are higher-value, harder to automate, and billable at the same rate. We turned the cost of writing code (labor) into the cost of running AI (capital), and freed the labor to do work that wasn't possible before. The same pattern holds for WebOps (FCP automated the ticket queue, freeing our team for client strategy) and marketing (AI generates the first draft, freeing our team for GEO and brand direction).

The net effect: our team is doing higher-level work than three years ago, serving the same number of clients with the same headcount, and the work itself is more interesting. That's the reinstatement effect working as intended.

And it generates net gains for the client. When we automate a task that used to take 4 hours of human time down to 15 minutes of AI time, the client doesn't pay for 4 hours anymore — they pay for the 15 minutes of compute plus the 45 minutes of human review. The same deliverable costs less. But the more important shift is what happens with the 3 hours we saved: we spend it on work that wasn't affordable before — deeper QA, performance optimization, accessibility audits, proactive security scanning, GEO strategy. The client gets a better product for the same budget because AI made the routine work cheap enough that the high-value work fits within the same retainer.

This is the Brynjolfsson productivity finding Generative AI at Work (2023) Brynjolfsson, Li, and Raymond studied a staggered rollout of a generative AI assistant to 5,179 customer support agents. Found a 14-15% average productivity increase, with 34% improvement for novices and minimal impact on experienced workers. Published in QJE, 2025.
Read the article →
at scale: AI made customer support agents 14% more productive, with the biggest gains for novices. At Fruition, the gains are larger — because we're not just augmenting a single task, we're restructuring the entire task mix around what AI does well and what humans do well. The client's dollar buys more outcome, not more hours. That's the difference between using AI as a tool and using it as an organizational design principle.

The products we built to capture the cycle

We didn't just adopt AI tools — we built platforms that institutionalize the labor → capital shift. Each one takes a category of human labor and turns it into automated infrastructure, then frees the human to do the work the infrastructure can't.

FCP
Fruition Control Plane — infrastructure automation
Replaces: manual ticket triage, deployment steps, backup verification, security patching, hosting checks — the routine ops work that used to fill a WebOps specialist's day. FCP automates 95% of the ticket queue, with AI-powered root cause analysis for the 5% that needs human attention. The human shifts to monitoring governance, client strategy, and scope management. See FCP →
FlyFruition
Airport operations data platform
Replaces: manual flight data aggregation, TSA wait time reporting, parking status updates, weather integration — work that airport operations teams used to do by hand across disconnected systems. FlyFruition aggregates, normalizes, and delivers it all through a single API and digital signage system. The human shifts to operational decision-making and passenger experience design. Live at Denver International Airport (65M+ annual passengers). See FlyFruition →
FruGPT
Enterprise document intelligence
Replaces: manual document review, knowledge base searches, compliance reading, contract summarization — the reading and retrieval work that paralegals, compliance officers, and analysts spend hours on. FruGPT processes millions of documents and delivers instant, accurate retrieval across the entire library. The human shifts to synthesis, judgment, and decisions that require context the AI doesn't have. See FruGPT →
PDF ADA
Automated PDF accessibility remediation
Replaces: manual PDF tagging, alt-text writing, structural remediation — hours of painstaking work per document for accessibility teams. The AI engine does structural fixes in under 2 seconds and adds semantic alt-text via vision models. The human shifts to compliance governance and document strategy. Production API at ada.fru.io. See PDF ADA →
Shield
WAF + infrastructure security scanning
Replaces: manual security monitoring, WAF rule tuning, vulnerability scanning, threat response — the 24/7 vigilance work that used to require a dedicated security engineer. Shield WAF sits in front of the origin with managed rules, geo-blocking, rate limiting, and DDoS protection. Infrastructure scanning runs continuously. The human shifts to security architecture and incident investigation. See Shield →
Steldris™
Modern WordPress replacement
Replaces: the clunky 20-year-old WordPress editing experience and its accumulated baggage — plugin conflicts, security holes, performance overhead, and accessibility gaps. Steldris™ delivers performant sites with security and accessibility built in, not bolted on. Sites can be static or dynamic — with form handling, real-time features, API integrations, and whatever else the site needs — without the weight of a legacy CMS. The human shifts from plugin maintenance and security patching to content strategy and user experience.
Unroo™
Knowledge-to-AI work orchestration
Replaces: scattered tickets across Jira, email, Slack, and spreadsheets — the coordination overhead that eats billable hours on every client engagement. Unroo™ orchestrates knowledge work across teams, AI agents, and tools — unifying project tracking, task management, time logging, and client communication into one platform that syncs with the systems already in use. The human shifts from chasing status updates and reconciling timesheets to actual delivery work and client relationship building. See Unroo™ →

The labor → capital → labor cycle

Labor
Human performs task
Capital
AI automates task
New Labor
New task emerges for humans

Each cycle, the new task is higher-level than the one it replaced. The work doesn't disappear — it moves up.

The Research

The studies behind the numbers

Every parameter in this model is grounded in peer-reviewed research. Hover any citation to read a summary and link to the original.

Eloundou et al. (2023) — “GPTs are GPTs” GPTs are GPTs (2023) Eloundou, Manning, Mishkin, and Rock assessed every O*NET occupation against LLM capabilities. Found that 80% of the U.S. workforce could have ≥10% of tasks affected, 19% face ≥50% task impact, and 47-56% of all tasks could be completed faster with LLM-powered software. Published in Science, 2024.
Read the article →

Researchers at OpenAI and the University of Pennsylvania assessed every O*NET occupation against LLM capabilities. Key findings:

  • 80% of the U.S. workforce could have at least 10% of their work tasks affected by LLMs
  • 19% of workers may see at least 50% of their tasks impacted
  • With LLM-powered software (not just raw models), 47–56% of all worker tasks could be completed significantly faster at the same quality
  • Higher-income jobs face greater exposure — the opposite of previous automation waves

McKinsey — “The Economic Potential of Generative AI” (2023) The Economic Potential of Generative AI (2023) McKinsey estimates that generative AI could automate work activities absorbing 60-70% of employee time today, force 12 million U.S. workers to switch occupations by 2030, and add $2.6-4.4 trillion annually to the global economy.
Read the article →

  • Generative AI and existing technologies could automate work activities absorbing 60–70% of employees' time today — up from a previous estimate of 50%
  • 12 million U.S. workers may need to transition to new occupations by 2030
  • Potential productivity boost of 0.5–3.4% annually through 2040, with generative AI contributing 0.1–0.6 percentage points

Brynjolfsson, Li & Raymond (2023) — “Generative AI at Work” Generative AI at Work (2023) Brynjolfsson, Li, and Raymond studied a staggered rollout of a generative AI assistant to 5,179 customer support agents. Found a 14-15% average productivity increase, with 34% improvement for novices and minimal impact on experienced workers. Published in QJE, 2025.
Read the article →

A staggered rollout of a generative AI assistant to 5,179 customer support agents produced the first large-scale real-world productivity data:

  • Average productivity increase of 14–15% (issues resolved per hour)
  • 34% improvement for novice and low-skilled workers
  • Minimal impact on highly experienced workers — AI mostly compresses the learning curve
  • AI assistance reduced average handle time and improved customer satisfaction scores

WEF Future of Jobs Report (2025) Future of Jobs Report (2025) The World Economic Forum's survey of 1,000+ employers across 22 industries and 55 economies. Projects 170 million new roles created and 92 million displaced by 2030, with 86% of businesses reporting AI is transforming their operations.
Read the article →

  • 170 million new roles created globally by 2030
  • 92 million existing jobs displaced — a net gain of 78 million
  • 86% of businesses report AI is transforming their operations
  • 22% of today's jobs will be affected by macrotrends including AI

Stanford AI Index (2025) AI Index Report (2025) Stanford HAI's annual report tracking AI adoption, investment, and capability. Key data: 78% of organizations use AI in 2024 (up from 55% in 2023), $33.9B in generative AI investment, and a 280× drop in inference costs from 2022 to 2024.
Read the report →

  • 78% of organizations reported using AI in 2024, up from 55% in 2023
  • Private investment in generative AI reached $33.9 billion globally, an 18.7% increase from 2023
  • AI inference costs dropped 280× from 2022 to 2024
Methodology

How the model works

Each job is decomposed into 8–10 core tasks drawn from O*NET occupational data O*NET OnLine The U.S. Department of Labor's occupational database, providing detailed task decompositions for 900+ occupations. Used in this visualization to break each job into its constituent tasks with time-allocation weights.
Read the article →
. Each task has an AI susceptibility score (0–1) based on the Autor, Levy, and Murnane The Skill Content of Recent Technological Change (2003) Autor, Levy, and Murnane's breakthrough paper showing that computers substitute for workers on routine tasks (following explicit rules) and complement workers on non-routine tasks (problem-solving and communication). The foundation of task-based automation theory.
Read the article →
routine-vs-nonroutine framework and the Eloundou et al. GPTs are GPTs (2023) Eloundou, Manning, Mishkin, and Rock assessed every O*NET occupation against LLM capabilities. Found that 80% of the U.S. workforce could have ≥10% of tasks affected, 19% face ≥50% task impact, and 47-56% of all tasks could be completed faster with LLM-powered software. Published in Science, 2024.
Read the article →
LLM exposure rubric.

Each task also carries a 2026 baseline — the AI capability level for that task as of 2026, grounded in real-world observation. For example, "writing boilerplate code" is at 98% because in practice, no developer writes code from scratch anymore — Copilot, Cursor, and Claude handle code generation. But that doesn't mean the software developer job is 98% automated, because coding is only 15% of the job's task weight. The 2026 baseline lets the model distinguish between "this task is already done by AI" and "this job is already done by AI" — the key insight from the task-based automation literature.

AI capability for each task is modeled as a logistic curve anchored at the 2026 baseline. The adoption curve that drives task activation comes from the Bass model A New Product Growth Model (1969) Frank Bass' mathematical model of how new products get adopted, using two parameters: p (innovation coefficient, avg 0.03) and q (imitation coefficient, avg 0.38). The model produces the S-shaped adoption curve that matches real-world diffusion patterns.
Read the article →
with parameters calibrated to Stanford AI Index adoption data AI Index Report (2025) Stanford HAI's annual report tracking AI adoption, investment, and capability. Key data: 78% of organizations use AI in 2024 (up from 55% in 2023), $33.9B in generative AI investment, and a 280× drop in inference costs from 2022 to 2024.
Read the report →
: p = 0.07 and q = 0.45. The model's adoption curve matches observed data: ~15% in 2020, ~55% in 2023, ~78% in 2024.

The total automation percentage at any point in time is the weighted average of AI capability across all tasks, weighted by the time each task typically consumes. When this percentage crosses 50%, the visualization shows the job transforming: the human's task profile shifts to new, higher-level work that the AI has made feasible or necessary.

This is a simplified model for illustration, not a prediction. Real task displacement is heterogeneous across industries, firms, and individuals. The model does not account for regulation, cost, organizational resistance, or the fact that many tasks are more complex than they appear. The research cited above provides the empirical grounding; the visualization is an interpretation.

Brad Anderson

Brad Anderson

CTO / Founder, Fruition

Brad founded Fruition in 2003 and has spent decades leveraging technology — from the .com boom to blockchain to AI — to solve problems that matter for enterprises, airports, and government agencies. He holds a B.A. in Economics from the University of Colorado at Boulder, an M.S. in Technology Management from the University of Denver, and a J.D. from the University of Denver Sturm College of Law.

References

  1. 1. Rogers, E. M. (1962). Diffusion of Innovations. Free Press. (5th edition, 2003.) Diffusion of Innovations (1962) Everett Rogers' foundational work showing that new technologies spread through a population in five adopter categories: innovators (2.5%), early adopters (13.5%), early majority (34%), late majority (34%), and laggards (16%).
    Read the article →
  2. 2. Bass, F. M. (1969). A new product growth for model consumer durables. Management Science, 15(5), 215–227. A New Product Growth Model (1969) Frank Bass' mathematical model of how new products get adopted, using two parameters: p (innovation coefficient, avg 0.03) and q (imitation coefficient, avg 0.38). The model produces the S-shaped adoption curve that matches real-world diffusion patterns.
    Read the article →
  3. 3. Autor, D. H., Levy, F., & Murnane, R. J. (2003). The skill content of recent technological change. QJE, 118(4), 1279–1333. The Skill Content of Recent Technological Change (2003) Autor, Levy, and Murnane's breakthrough paper showing that computers substitute for workers on routine tasks (following explicit rules) and complement workers on non-routine tasks (problem-solving and communication). The foundation of task-based automation theory.
    Read the article →
  4. 4. Acemoglu, D. & Restrepo, P. (2018). The race between man and machine. AER, 108(6), 1488–1542. The Race Between Man and Machine (2018) Acemoglu and Restrepo's model of automation's dual effects: displacement (tasks move from human to machine) and reinstatement (new tasks emerge for humans). The net effect on wages depends on which effect dominates.
    Read the article →
  5. 5. Acemoglu, D. & Restrepo, P. (2019). Automation and new tasks. JEP, 33(2), 3–30. Automation and New Tasks (2019) Acemoglu and Restrepo extend their framework, showing that automation displaces workers but the creation of new tasks reinstates labor. Historical evidence suggests productivity growth from automation benefits workers only when new task creation keeps pace.
    Read the article →
  6. 6. Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023/2024). GPTs are GPTs. Science, 384(6698), 986–989. GPTs are GPTs (2023) Eloundou, Manning, Mishkin, and Rock assessed every O*NET occupation against LLM capabilities. Found that 80% of the U.S. workforce could have ≥10% of tasks affected, 19% face ≥50% task impact, and 47-56% of all tasks could be completed faster with LLM-powered software. Published in Science, 2024.
    Read the article →
  7. 7. Brynjolfsson, E., Li, D., & Raymond, L. (2023/2025). Generative AI at work. QJE, 140(2), 889–942. Generative AI at Work (2023) Brynjolfsson, Li, and Raymond studied a staggered rollout of a generative AI assistant to 5,179 customer support agents. Found a 14-15% average productivity increase, with 34% improvement for novices and minimal impact on experienced workers. Published in QJE, 2025.
    Read the article →
  8. 8. McKinsey Digital. (2023). The economic potential of generative AI: The next productivity frontier. The Economic Potential of Generative AI (2023) McKinsey estimates that generative AI could automate work activities absorbing 60-70% of employee time today, force 12 million U.S. workers to switch occupations by 2030, and add $2.6-4.4 trillion annually to the global economy.
    Read the article →
  9. 9. World Economic Forum. (2025). The Future of Jobs Report 2025. Future of Jobs Report (2025) The World Economic Forum's survey of 1,000+ employers across 22 industries and 55 economies. Projects 170 million new roles created and 92 million displaced by 2030, with 86% of businesses reporting AI is transforming their operations.
    Read the article →
  10. 10. Stanford HAI. (2025). Artificial Intelligence Index Report 2025. AI Index Report (2025) Stanford HAI's annual report tracking AI adoption, investment, and capability. Key data: 78% of organizations use AI in 2024 (up from 55% in 2023), $33.9B in generative AI investment, and a 280× drop in inference costs from 2022 to 2024.
    Read the report →
  11. 11. O*NET OnLine. U.S. Department of Labor, Employment and Training Administration. O*NET OnLine The U.S. Department of Labor's occupational database, providing detailed task decompositions for 900+ occupations. Used in this visualization to break each job into its constituent tasks with time-allocation weights.
    Read the article →