Return-to-office fights and another round of cost cutting have managers stretched thinner than ever. The same quarter, boards are told that AI agents will carry more of the workload. The risk is obvious: layering new software labor onto old chains of command. Engagement is already fragile. Only 20 percent of employees globally report being engaged, the lowest level since 2020, and managers have seen a nine-point drop in engagement since 2022, with a step down from 27 to 22 percent in just one year (Gallup, 2025). If AI simply adds tickets to overloaded backlogs, it will compound the problem.
The operating model is the product
Microsoft reports that 82 percent of leaders say they will use digital labor to expand the capacity of their workforce in the next 12 to 18 months, based on a survey of 31,000 knowledge workers in 31 countries combined with Microsoft 365 telemetry (Microsoft WorkLab, 2025). In organizations that have deployed AI agents across the enterprise, what Microsoft calls Frontier Firms, 71 percent of workers say their company is thriving, compared with 39 percent globally. The correlation is encouraging, but it does not specify causation, and Microsoft has a commercial interest in the technology. What it does usefully foreground is a design choice: replace the static org chart with a dynamic work chart that maps outcomes, flows, and accountable roles, including agents. Nearly half of leaders in the same dataset say they already use agents to fully automate workflows or processes, which makes a paper hierarchy even less relevant (Microsoft WorkLab, 2025).
Consulting evidence points in the same direction, with similar caveats. McKinsey estimates that 89 percent of organizations still run on industrial-age structures, 9 percent on digital-age product or platform models, and only 1 percent as decentralized networks. In client work, the firm observes that a human team of two to five people can supervise an agent factory of 50 to 100 specialized agents running an end-to-end process such as customer onboarding or closing the books (McKinsey, 2025). These are not audited experiments. They are a practitioner’s view of what is becoming possible when work is decomposed into outcomes and executed by networks rather than vertical chains.
The common claim is clear. The unit of design is no longer a department. It is a small human crew accountable for outcomes, paired with a changing set of agents that do the work. The coordination mechanism is the work chart, not the org chart.
Self-management, updated for software labor
This is not a break with everything we know about organization design. It is a remix. The research record on self-managing organizations is explicit about the trade. In environments where tasks are known and stable, managerial hierarchy executes reliably. In complex, cross-functional, non-routine work, the same hierarchy tends to slow problem solving, lock in rigidity, and reinforce status gaps that stunt people’s development (Lee and Edmondson, 2017). Self-managing organizations decentralize authority formally and systematically across the firm, and their hallmark is the elimination of the manager-subordinate relationship. That tradition was born for human teams. AI agents expand the logic, because the work now crosses boundaries by default and changes state more quickly than approvals travel.
A human-agent organization borrows two core moves from self-management and applies them to software labor. First, decentralize decisions to the edge where information is richest. Second, make those decisions legible to peers through explicit roles and contracts. Microsoft’s work chart concept provides the scaffold, a living map of outcomes, owners, and interdependencies that can be adjusted as agents evolve and tasks change (Microsoft WorkLab, 2025). McKinsey’s agentic organization frames the same move as a flat network of outcome-aligned teams that exchange tasks and results rather than escalate requests up a chain (McKinsey, 2025). In both accounts, traditional spans and layers matter less than the clarity of who owns what outcome and how it is tested.
The promise is organizational agility without the managerial overhead of routing every decision. The warning from the academic record is equally important. Self-management is not the absence of structure. It is the presence of different, often more explicit, structures that define accountabilities and conflict resolution at the point of work (Lee and Edmondson, 2017). Swapping bosses for bots without that scaffolding will not work.
Safety is the governor
Teams that learn outperform teams that posture. Two decades before AI agents were a management topic, field research with 51 work teams showed that psychological safety, a shared belief that it is safe to take interpersonal risks, was associated with learning behaviors. Learning mediated the relationship between safety and performance, while general team efficacy did not predict learning when safety was controlled (Edmondson, 1999). This is not a vibe metric. It is a work variable that governs whether people speak up about errors, surface weak signals, and update routines.
In a human-agent organization, that mechanism becomes central. Agents change quickly. Prompts, guardrails, and handoffs evolve by the week. If people do not feel safe to say when an agent struggled or a workflow failed, learning stalls, and the system decays. Safety is the condition that lets small human crews critique their own tools in public, not a substitute for performance standards. The lesson carries from Edmondson’s study to the current context because the learning mechanism it identified is technology agnostic and team based (Edmondson, 1999).
How to start a human-agent design
Leaders do not need a grand reset to begin. They do need to stop pouring software into the old mold.
- Publish a work chart for one end-to-end process. Name the outcome, the human owner, the agent roles, and the upstream and downstream contracts. Treat it as a living artifact that changes when the work changes (Microsoft WorkLab, 2025).
- Form a small crew to supervise the process. Two to five people can often manage dozens of agents in practice, according to consulting experience, but treat this as a design hypothesis to test, not a rule (McKinsey, 2025).
- Make the learning loop explicit. Define how issues are surfaced, how changes are proposed, who approves them, and how fast they propagate. Measure the learning behavior you want before you measure throughput. The evidence links safety to learning to performance at the team level (Edmondson, 1999).
- Watch the people metrics. Global engagement is at 20 percent and managers are sliding. Best-practice organizations reach 79 percent manager engagement, which argues that design choices, not macro conditions, drive the gap (Gallup, 2025).
The practical bet is straightforward. If digital labor is going to grow, design around outcomes and learning, not around boxes and arrows. Microsoft’s Frontier Firms report higher rates of perceived thriving among employees, a signal worth chasing even if the causal chain is not proved (Microsoft WorkLab, 2025). McKinsey’s client work suggests small human teams can safely supervise large pools of agents when the work is decomposed and contracts are clear (McKinsey, 2025). The academic record reminds us that flattening authority improves complex problem solving when matched with explicit structures and psychological safety (Lee and Edmondson, 2017; Edmondson, 1999).
Do not add another layer of coordinators to keep the old chart intact. Replace the chart. Name the outcomes. Put a crew on them. Let the agents do the busy work. Then make it safe to say when the system is wrong so it can get better fast.
Sources
- Microsoft WorkLab (2025). Work Trend Index 2025: The Year the Frontier Firm Is Born. Reports that 82% of leaders plan to use digital labor, Frontier Firms show 71% of workers saying the company is thriving vs 39% globally, and introduces the dynamic Work Chart concept; supports the shift from org charts to outcome-based coordination. https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born
- McKinsey (2025). The agentic organization: contours of the next paradigm for the AI era. Consulting view that most firms remain industrial-age, and that small human teams can supervise 50–100 agents in end-to-end processes; supports the feasibility of agent factories and flat networks. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-agentic-organization-contours-of-the-next-paradigm-for-the-ai-era
- Lee and Edmondson (2017). Self-managing organizations: Exploring the limits of less-hierarchical organizing. Synthesizes evidence that hierarchy suits stable, known tasks but impedes complex cross-functional problem solving, and defines SMOs as eliminating the manager–subordinate relationship; grounds the argument for decentralization. https://www.sciencedirect.com/science/article/abs/pii/S0191308517300059
- Edmondson (1999). Psychological Safety and Learning Behavior in Work Teams. Field study linking team psychological safety to learning behaviors, which mediate performance; establishes why safety is a work variable, not a perk, in human–agent teams. https://journals.sagepub.com/doi/10.2307/2666999
- Gallup (2025). State of the Global Workplace 2025. Finds global engagement at 20%, managers’ engagement down nine points since 2022, and best-practice manager engagement at 79%; underscores the cost of layering AI onto fragile human systems. https://www.gallup.com/workplace/659610/state-global-workplace.aspx