The work-design week: AI is becoming a capacity-planning problem
This week's signals point to a shift from adding AI to jobs toward redesigning work around human judgment, agent capacity, and measurable safeguards.
The signal: capacity now includes agents
Atlassian is hiring a director of capacity planning to build the frameworks that help leaders decide which work should go to people and which should go to AI agents. The role also covers the tools, systems, and agent libraries that make this capacity visible, while business leaders and managers still decide how teams are assembled. 1
That is a more consequential change than adding another assistant to the software stack. Atlassian has put its people team and internal IT organization under one leader, and says roughly 90% of employees use AI daily. The next move is from individual experimentation to team workflows. The operating question is shifting from "How many people do we need?" to "What capacity do we need, and which parts come from people or agents?" 1
Two other signals from the week
The labor market is sorting tasks, not simply removing jobs. The U.S. Bureau of Labor Statistics projects 2024-34 growth of 15.8% for software developers, adding more than 267,000 jobs, while customer service representatives are projected to decline 5.5%, or 153,700 jobs. These are occupation-level projections, not a forecast that AI alone will cause either outcome. They still give operators a useful planning baseline: demand is moving toward technical, analytical, and security work while some high-volume service work faces sustained pressure. 2
Hiring safeguards are becoming an operating requirement. A Princeton and University of Chicago study placed ChatGPT, Claude, and Gemini in a simulated hiring game covering 20 jobs. Although every fictional candidate had the same underlying chance of success, the models increasingly sorted groups into different occupations after receiving early feedback. Human participants scored 0.84 on the study's segregation scale; the models scored about 65% higher, and OpenAI's o3 scored 1.83, close to the theoretical maximum. The experiment does not reproduce real recruiting, where feedback is slower and less complete, but it is a direct warning against treating model output as neutral evidence about candidate quality. 3
What to do this week
- Map capacity by workflow. Pick one recurring process and list its inputs, decisions, handoffs, and outputs. Mark where an agent can execute, where a person must approve, and where the work needs a human relationship or judgment.
- Change the unit of workforce planning. Review roles in terms of skills and outcomes, not only job titles. BLS projections suggest that the useful question is often which tasks are growing inside a role, not whether the role survives unchanged. 2
- Put a named owner on AI decisions. For recruiting and performance workflows, record the model, data, decision point, human reviewer, and appeal route. If nobody can explain why a recommendation was made, the workflow is not ready for scale.
- Train managers for work design. The practical skills are workflow mapping, quality control, data literacy, and ethical oversight. HRD's guidance this week makes the same point: redesign work around capabilities, outcomes, and decision responsibility rather than freezing today's job descriptions. 4
The near-term advantage will go to teams that can describe their work precisely enough to decide what should be automated, what should be learned, and what must remain accountable to a person.
関連コンテンツ
- ログインするとコメントできます。