Workers Are Drawing the Line on AI — and Managers Are on It

What workers keep

Mercor, working with researchers from Stanford, the O*NET Resource Center, and the National Bureau of Economic Research, has published a study of what workers want to hand to AI and what they want to keep. The researchers surveyed more than 200 participants from the original TaskNet study about AI's impact on their jobs, collecting 9,000 task ratings. TaskNet is built on the structure of O*NET, the US Department of Labor's map of occupational tasks. The headline finding is a split: workers prefer to retain tasks that are relational or high-stakes, and to hand over clerical work. m:…9214

The numbers give the split shape. Of the 9,000 task ratings, 14 percent required no human-AI collaboration, while 27 percent required essential collaboration, the researchers found. Lawyers rated 38 percent of their tasks as requiring essential human-AI collaboration, compared with 19 percent for software developers. And 78 percent of US knowledge workers want humans leading high-risk decisions, while 3 percent would let AI make them alone. Workers in more exposed occupations are more likely to report significant job changes due to AI in the past year, the study found, yet ER physicians report lower levels of job change and worry than their AI exposure scores would suggest. m:…9214

What workers would do with the time matters as much as what they hand over. The study found 38 percent of respondents would do more of the tasks they enjoy with time saved by AI, 22 percent would do the same work at higher volume, and 18 percent would move into new kinds of work. That is not a workforce asking to do less. It is a workforce asking to spend its hours differently, and the split between enjoying the work, doing more of it, and doing new work is close enough that no single answer dominates. m:…9214

The manager question

Amplify Partners asks the harder version of the question: why should managers exist at all, if AI keeps getting cheaper at producing work? The firm's answer is that AI reduces the cost of producing work faster than the cost of verifying it, and that firms exist because market transactions have costs, an argument associated with Ronald Coase, with management persisting because firms must solve integration costs, not just production costs. Where integration problems can be expressed computationally, AI allows greater autonomy; where they cannot, judgment and leadership remain bottlenecks, the post argues. Integration regimes can be categorized by formalization and authority concentration, it says. m:…8298

The post notes that AI has not yet significantly impacted productivity statistics, and that AI agents can potentially reduce transaction costs by automating information discovery and contract drafting. The history it leans on is old: the Pennsylvania Railroad developed the line-and-staff organization in 1857, and Alfred Chandler argued that railroads necessitated the creation of the managerial firm. The point is that management is an answer to a coordination problem, and coordination is the part AI has not yet made cheap. m:…8298

HR is where the line gets tested

Serval shows what the line looks like inside one function. The company says AI agents in HR carry out multi-step tasks through connected applications and answer employees' questions, going beyond traditional chatbots, which primarily help employees find information by retrieving answers from an HR knowledge base. Serval lists five use cases: talent acquisition agents that coordinate recruiting tasks such as scheduling interviews; onboarding agents that initiate account setup after HR approves the hire; HR help desk agents that escalate cases needing specialist attention; workforce planning agents that compare approved headcount with open positions to find capacity gaps; and performance management agents that flag overdue manager feedback and suggest training. m:…5327

The governance warning comes with the pitch. Serval says IT needs to know which systems the agent will access, the changes it can make, and who takes responsibility when something goes wrong, and that AI agents reduce the manual work that falls between departments. The company says it tracks AI resolution rate to check how many requests its agents complete without human intervention, and offers Serval Simulations so IT teams can test AI help desk agents on realistic scenarios without touching real systems. Serval also says it raised $150 million at a $3 billion valuation. The Every Media conference adds a cultural coda: Dan Shipper argues that human perspective is the currency of the AI era because AI models struggle to prioritize insights, and Katy George predicts that by 2027, the primary human role will be teaching AI to perform tasks. Workers are drawing the line themselves, and the line keeps moving toward judgment, teaching, and the work only a person can own. m:…5327 m:…7771