The argument over return to office misses the more urgent shift. AI agents are taking over the low-stakes, repetitive tasks that once trained early-career employees, while middle managers are stretched thin and, in some firms, being thinned out. The result is a capability gap that compounds over time: fewer on-ramps, fewer coaches, and a pipeline that looks efficient until you ask who will run the place in five years.
What the numbers actually say
Early-career workers are the canaries here. Using payroll data, researchers find that employees aged 22 to 25 in occupations most exposed to AI suffered about a 16 percent relative decline in employment, while experienced workers in the same jobs held steady and workers in less exposed fields grew (Brynjolfsson, Chandar and Chen, 2025). The adjustment has run through hiring and headcount, not wages. The losses also cluster where AI automates work outright rather than augments it, a distinction with direct design implications for leaders.
Employers see a sweeping skill realignment ahead. The World Economic Forum projects that 22 percent of jobs will be disrupted by 2030, with 170 million roles created and 92 million displaced, leaving a net gain of 78 million jobs if firms can match people to new work. Nearly 40 percent of the skills required on the job are expected to change, and 59 percent of workers will need upskilling or reskilling. Employers say the skills gap is the top barrier to business transformation, and while 77 percent plan to upskill in response to AI, 41 percent also plan to reduce headcount as automation spreads (World Economic Forum, 2025). These are expectations rather than observations, but they describe the operating assumptions many leaders now hold.
The tools are arriving fast. In Microsoft’s global survey and telemetry report, 82 percent of leaders say they will use AI agents to expand workforce capacity in the next 12 to 18 months, and 46 percent report that they already use agents to fully automate some workflows or processes. Microsoft introduces a “Work Chart,” a dynamic, outcome-oriented model that replaces static org charts. Workers in “Frontier Firms,” those deploying AI organization-wide with active agent use, are far more likely to say their company is thriving, 71 percent versus 39 percent globally (Microsoft, 2025). Vendor incentives mean these results warrant caution, but the direction of travel is clear.
The apprenticeship gap AI can widen
Most professionals learned by doing work that was safe to learn on: cleaning data, drafting first passes, reconciling lists, building the fifth version of a sales report. That work is exactly what agents excel at. Pair that with a management system under stress, and the ladder loses its first rungs. Managers spend nearly 40 percent of their time on firefighting and administrative work and only 13 percent on developing their people, and more than a third report feeling unprepared for the people side of the job (Deloitte, 2025). The supply of managers is tightening as well. US employers advertised 42 percent fewer middle-management roles at the end of 2024 than in spring 2022, and Deloitte cites a Gartner forecast that by 2026, 20 percent of organizations will use AI to flatten their structure (Deloitte, 2025). Fewer entry jobs, fewer coaches, less time to coach. That is a fragile development model.
This fragility is not a law of AI. It is a design choice. The Stanford evidence suggests that outcomes hinge on whether AI automates whole roles or augments people’s work. The former concentrates losses on the youngest workers. The latter can expand capacity without erasing the experiences that create judgment and tacit knowledge (Brynjolfsson, Chandar and Chen, 2025).
Design learning-grade workflows, not just automations
The fix is to rebuild work so learning is a requirement, not a leftover. Three moves stand out.
-
Re-bundle tasks around outcomes, then reserve deliberate practice for humans. Use a Work Chart logic to map who is responsible for a result, which tasks an agent can handle, and which steps a junior must perform to build fluency before the agent closes the loop. For example, require early-career staff to draft analyses, interpret model outputs, or handle edge cases, then let agents automate the rote wrap-up. This keeps repetition and feedback where it compounds skill, while agents speed the rest (Microsoft, 2025).
-
Appoint managers to orchestrate agents and apprentices together. New hybrid roles are emerging, sometimes called an “agent boss,” responsible for quality, exception handling, and the workflow itself. Make these leaders accountable for human development as well as throughput. Measure two things in parallel: cycle time and skill gain. If output climbs while the share of work done by novices collapses, you are raiding the future to fund the present (Microsoft, 2025).
-
Create time to teach, consciously. Shift administrative load to agents and codify playbooks so managers can move beyond today’s 13 percent share of time spent on people development. Given that 36 percent of managers feel unprepared for the people side of the role, build manager enablement for coaching in an AI context: how to assign stretch work when an agent can finish the task, how to review reasoning, and how to use telemetry for feedback without turning it into surveillance theater (Deloitte, 2025).
Reskilling strategy should match the labor-market signal. Target augmentation. The evidence that job losses concentrate where AI automates suggests that investments should steer toward roles and workflows where humans remain in the loop and where practice builds transferable skill. That lines up with employers’ own plans to reorient to AI-created opportunities, but only if curricula and rotations are tethered to real outcomes, not generic courses. Given that nearly 40 percent of in-demand skills will change by 2030 and 59 percent of workers will need reskilling or upskilling, firms that stage skills sprints tied to concrete deliverables will move faster than those that rely on catalog learning alone (World Economic Forum, 2025).
A different way to count the return
Boards often demand a clean line from AI to cost. They should also ask for a clean line to capability. Microsoft reports that workers in Frontier Firms, which deploy agents at scale, are far likelier to report thriving performance. The mechanism is not magic. These firms redesign work around outcomes and instrument it, which makes it easier to see where humans add value and where agents can safely take over. That visibility is what lets leaders protect and even expand the moments where people practice doing the hard parts of the job (Microsoft, 2025).
The macro picture backs a longer view. The World Economic Forum projects a net gain of 78 million jobs by 2030 if companies can match talent to new roles. That outcome will not arrive by accident. It will be built by firms that treat AI deployment as talent-system design, not just automation. Erik Brynjolfsson has long argued that the gains from new technologies are largest when organizations re-invent how work is done. The newest evidence shows why the reinvention must include the first rung and the managers who maintain it (Brynjolfsson, Chandar and Chen, 2025; World Economic Forum, 2025).
The practical implication is clear. If you automate the apprentice work and shed the coaches without rebuilding how people learn, you will get faster and more fragile at the same time. If you design learning-grade workflows, appoint leaders to run them, and measure skill as rigorously as you measure output, you will get the two things AI makes possible for human organizations: more capacity today and a stronger bench tomorrow.
Sources
- World Economic Forum (2025). The Future of Jobs Report 2025. Projects the scale of disruption, the share of skills that will change, and employers’ plans for upskilling and headcount, framing the reskilling imperative. https://www.weforum.org/publications/the-future-of-jobs-report-2025/
- Deloitte (2025). 2025 Global Human Capital Trends. Shows managers’ time use and preparedness, the drop in middle-management postings, and a forecast of AI-driven flattening, highlighting the coaching bottleneck. https://www.deloitte.com/us/en/insights/topics/talent/human-capital-trends/2025.html
- Microsoft (2025). Work Trend Index 2025 — The Year the Frontier Firm Is Born. Reports leaders’ adoption of AI agents, the Work Chart concept, and higher thriving rates in Frontier Firms, illuminating how outcome-centric workflows enable both capacity and learning. https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born
- Brynjolfsson, Chandar and Chen (2025). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of AI. Finds early-career employment declines in AI-exposed jobs and shows impacts run through employment and concentrate where automation dominates, guiding augmentation-first design. https://digitaleconomy.stanford.edu/publication/canaries-in-the-coal-mine-six-facts-about-the-recent-employment-effects-of-artificial-intelligence/