couple of years, experts have been locked in a debate about AI’s impact on jobs. Will it create them or destroy them? Will jobs be human-led or AI-led? This binary discussion, as research is revealing, is not asking the right questions.
Two large-scale studies, Stanford’s “WORKBank” (1,500 workers, 844 tasks) and Anthropic’s “Claude Economic Index” (4.1 million chats, 19,000 tasks), show that AI is reshaping work task-by-task, not role-by-role. Fewer than 4% of occupations are close to full automation, yet employees themselves want 46% of individual tasks automated, chiefly repeatable finance, reporting, and data-entry work. Most knowledge workers prefer “equal-partner” copilots over lights-out automation, and real-world usage bears this out: 57% of observed AI interactions are augmentative dialogues, 43% are hands-off delegation. The skills premium is already tilting away from routine analysis toward workflow orchestration, prioritization, and interpersonal influence.
These nuances are important. AI will first shape tasks, not jobs. It is also very likely that very few jobs will fully go away. When we talk about “jobs will be transformed,” this is what it exactly means – many tasks in that job will be done by AI and more time will be spent on other or new tasks.
We need to move on from vague and high-level strategies to detailed approaches such as work graphs at task level. In this article, we will dive into the findings of these 2 studies and then explore a three-pronged playbook.
What Workers Want vs. What AI Can Do: The Stanford “WORKBank” Study
To understand the future of work, we must first understand the work itself. This was the premise of Stanford’s “WORKBank” study, which systematically audited work not from the top down (job titles) but from the bottom up (individual tasks). Surveying over 1,500 U.S. workers across 104 occupations and 844 distinct tasks, researchers built a unique dataset based on a simple but critical question: What parts of your job do you want to hand over to an AI and which ones can it actually do?
What makes this study uniquely powerful is its multi-layered approach. It didn’t just capture worker desire; it cross-referenced it with opinions of 52 leading AI experts who rated the technical feasibility of automating each of those same tasks.
Two Frameworks to Navigate the Future
The Stanford team synthesized their findings into two elegant frameworks:
The Human Agency Scale (HAS): This five-level scale classifies desired human involvement in a task, from H1 (AI performs the task entirely, or “lights-out” automation) to H5 (the task is essentially human and AI has no role). It provides nuanced language for discussing automation, moving beyond the simple “human vs. machine” binary.
The Desire–Capability Matrix: The researchers then plot every role on a matrix. While they use averages of task scores for the 2×2, I believe it is much better to look at the role level aggregate data in Appendix E.4. If we take that data and analyze at role level much clearer Enterprise AI implications emerge. This creates four distinct zones, each with clear strategic implications:
- The Green Zone (Automate): High worker desire, high AI capability. These are no-brainer tasks ripe for full automation.
- The Blue Zone (Innovate): High worker desire, low AI capability. Market opportunities lie here for AI builders addressing problems workers want solved.
- The Yellow Zone (Educate): Low worker desire, high AI capability. Workers underestimate what AI can do, an opportunity for internal education and enablement.
- The Red Zone (Passive): Low worker desire, low AI capability. This is an area where Enterprises should monitor progress but no immediate action.
Key Findings: A Desire for Partnership, Not Replacement
Workers want the drudgery to be automated. The study’s findings dispel myths around one contentious area, that workers inherently do not want AI. A staggering 46% of all tasks were things workers actively wanted to offload, primarily tedious, repetitive work that drains cognitive resources. The top reason cited was ambition: 69% said their goal was to “free my time for high-value work.”
Full automation is not desirable. The desire for AI automation is not a desire for obsolescence. Fear remains, with 28% of workers expressing concerns about job security and the “dehumanizing” of their roles. This is why the ideal interaction model is not replacement but partnership. Across the board, 45% of occupations reported “equal partnership” (H3 on the agency scale) as their ideal state, far preferring a copilot setup to a complete takeover.
Workers consistently ask for more agency than experts say is technically required. This means that executives will have to lead on this path empathetically. Workers want AI but want it less than what is possible.
Perhaps most telling is the emerging “skills inversion.” The premium is rapidly shifting away from routine analytical tasks, the very skills that defined the knowledge worker of the last 20 years, and toward a new set of meta-skills: organizing and prioritizing work, giving guidance, interpersonal consultation, and making decisions under ambiguity. In the agent-led enterprise, your value will be defined less by your ability to do analysis and more by your ability to orchestrate the agents that do.
What People Are Actually Doing: The Anthropic “Claude Economic Index”
If the Stanford study tells us what’s possible and desired, the Anthropic Claude Economic Index tells us what’s actually happening now. By analyzing 4.1 million real-world interactions with its Claude AI model and mapping them to over 19,000 official O*NET tasks, Anthropic has created an unprecedented, real-time snapshot of AI adoption in the wild.
Who Is Adopting and Who Is Not
The data shows AI adoption is not evenly distributed; it has clear hot and cold zones. The “hot” zones are unsurprising: 37% of all usage comes from computer and mathematical occupations (coding, scripting, troubleshooting), followed by 10% from writing and communications (marketing copy, technical documentation). The “cold” zones are roles requiring physical presence: construction, food service, and hands-on healthcare show near-zero engagement.
More revealing is the analysis by “Job Zone,” a classification of roles based on required preparation level. Peak AI usage happens in Job Zone 4. These are roles like software developers, analysts, and marketers that typically require a bachelor’s degree. This group uses AI 50% more than expected, accounting for over half of all analyzed usage. Conversely, usage is lower at the extremes: Job Zone 1 (e.g., baristas) and Job Zone 5 (e.g., physicians, lawyers) both under-index significantly. This tells us AI’s current sweet spot is in structured, analytical knowledge work.
How Are They Using It? Augmentation Still Rules
The study confirms Stanford’s findings on worker preference. A majority of interactions, 57%, are “augmentative,” characterized by iterative dialogue, validation, and learning, a true copilot relationship. Only 43% are fully “automated” or delegated, where users give a prompt and expect a finished product without back-and-forth.
When we drill down into tasks themselves, the pattern becomes even clearer. Dominant use cases are in high-value, complex work: software development and debugging, creating technical documentation, and business process analysis. This is not about automating simple clerical work; it’s about augmenting core functions of the most valuable knowledge workers.
Crucially, the study shows that full job automation is a red herring. Only 4% of occupations see AI touching over 75% of their constituent tasks, and these are typically narrow fields like language instruction and editing. However, 36% of occupations have “highly active pockets” of AI, with technology present in at least a quarter of their tasks. A marketing manager might not use AI for client engagement, but they are heavily using it for market research and strategic planning. This task-level penetration is the metric that matters.
The Executive Playbook: Three Imperatives for the AI Agent Empowered Enterprise
This data is more than academically interesting. It provides a blueprint for an enterprise AI strategy. Here are three specific, actionable imperatives for every senior leader.
1. Targeted Automation and Copilot Opportunities
The approach here should depend on the roles and the tasks. These fall into three zones:
Automate the Obvious (Green Zone): The consensus from both studies is clear. A high share of tasks in finance, accounting, and repetitive data administration are ready for full automation. This is where one should be looking to systematically, at scale, automate low-value tasks.
Deploy Copilots Strategically (Green/Yellow Zone): For functions like business intelligence, compliance, learning & development, and creative marketing, the mandate is augmentation. This doesn’t necessarily mean buying more tools; it means building AI capabilities into existing workflows. Think AI-generated starting-point reports for analysts, AI-powered compliance checklists, or AI-assisted content generation for marketers. The goal is uplift, not replacement.
Educate the Skeptics (Yellow Zone): The Stanford study revealed that many of our most skilled workers, such as engineers, analysts, and managers, underestimate what AI can do. We must investigate if this perception gap exists in our own organization. Is it due to lack of tools? Technical debt? Or cultural fear of being de-skilled? The answer will determine whether we need an enablement campaign (better tools and training) or a perception-shifting campaign (demonstrating value and building trust).
2. Go-To-Market & Product Innovation
Beyond internal efficiencies, this research highlights massive external opportunities for growth (Blue Zone).
Become an “AI Acceleration Partner”: The R&D Opportunity zones from the Stanford study, and underpenetrated areas from Anthropic study highlight industries like Legal, Healthcare, Travel, and E-commerce where either worker desire for AI dramatically outpaces current tech or there is a passive market. These can be areas to build new products and start-ups.
Explore New Product Frontiers: The data also flags specific occupational needs. For instance, both Information Security and Computer Network professionals report a high desire for AI assistance that current tools don’t provide. This is a clear signal for product teams to begin research and discovery. Is there a new security product to be built? A new network management platform powered by agents? The data provides a map to unmet needs.
3. Workforce Transformation & Skill Strategy
This is the most challenging, and most important area. AI’s task-level impact requires a complete overhaul of our talent management philosophy.
Build the “AI Orchestration” Skill Family: Both studies create a clear picture of new premium skills: workflow design, cross-functional orchestration and navigating ambiguity. Enterprises should invest in cultivating these abilities. This means building a new “AI-Orchestration” competency within learning paths and embedding it into career paths and performance reviews. The goal is to train people to excel at directing, validating, and integrating AI capabilities into complex workflows.
Adopt Task-Based Workforce Planning: The high-level headcount budget could become an artifact of the past. Enterprises should think beyond FTEs to modeling “task mixes per role.” This task-based view should drive hiring and redeployment decisions, integrating into budgeting cycles so future workforce choices are driven by the work actually to be done by humans.
Evolve from an Org Chart to a “Work Graph”: The ultimate goal is to move from a static, siloed organizational chart to a dynamic, living “Work Graph.” This is a company-wide map that details tasks, owners, dependencies, and automation states across functions, cutting through silos to optimize for end-to-end value streams. This graph becomes the single source of truth for prioritizing automation projects, identifying skill gaps, redesigning team structures, and even making strategic decisions about which processes to bring back from low-cost locations and which vendor relationships can be supplanted by more efficient AI agents.
The Partnership Imperative
The future of work isn’t about choosing between humans and AI. It’s about architecting their collaboration. The organizations that thrive will be those that move beyond the binary automation debate to focus on intelligent task decomposition, strategic capability development, and thoughtful change management.
The research is unequivocal: workers don’t want to be replaced by AI, but they do want to be freed from the repetitive, low-value tasks that prevent them from doing their best work. Companies that listen to this message and act on it systematically will gain not just operational efficiency, but significant competitive advantage in attracting and retaining top talent.
Perhaps most provocatively, successful organizations should explore bringing fully automatable processes back from low-cost locations into centralized, cloud-native operations supported by AI agents. Simultaneously, they should evaluate external BPO and SaaS relationships, piloting AI substitution where agents can match or exceed vendor service levels and reinvesting the savings in high-agency talent.
The task revolution is already underway. The question isn’t whether AI will reshape work, it’s whether your organization will lead that transformation or be disrupted by it. The choice, for now, remains human.
Shreshth Sharma is a Business Strategy, Operations, and Data executive with 15 years of leadership and execution experience across management consulting (Expert PL at BCG), media and entertainment (VP at Sony Pictures), and technology (Sr Director at Twilio) industries. You can follow him here on LinkedIn.
