I built a 21-role AI workforce. The hardest part was management

Microsoft's 2025 Work Trend Index for Singapore revealed that 56% of leaders were using AI agents for full automation, while 46% expected their teams to build multi-agent systems. The author of this account built an internal AI office with 21 defined roles across various departments, expecting the hard part to be technical. However, the real challenges lay in managerial aspects such as ownership, review, decision-making, and when a human should step in.

The author discovered that defining roles by what the model could do led to ownership confusion, as a single agent could research, write, publish, and evaluate the result. To address this, the author started defining roles based on what responsibilities they should own. For instance, a writing role writes, while a publishing role prepares distribution. This distinction clarified who was responsible for each job, making it easier to attribute errors.

Another mistake was allowing a single agent to create, review, and approve important outputs. This led to inefficient processes, as the agent tested its own work and then declared it ready without independent challenge. To rectify this, the author created three separate responsibilities: builder, reviewer, and acceptor. The builder produces the work, a separate reviewer tests or reviews it, and final acceptance of high-impact, client-facing, or irreversible actions still requires human oversight.

The author also realized that treating autonomy as a maturity score was incorrect. Instead, the focus should be on the potential consequences of wrong actions. Singapore's updated Model AI Governance Framework for Agentic AI uses a risk-based approach, categorizing actions by severity, reversibility, and human oversight feasibility.

This risk-based approach influenced the author's workflow design, with low-risk, reversible actions having wide freedom for the agent, while high-risk actions with limited reversibility may be blocked entirely.

Lastly, the author emphasized the importance of managing an agent's access to tools, databases, files, and external services. An agent is not just a prompt but has broader capabilities that require careful monitoring. The author now maintains an inventory of agents, tools, and their capabilities, aligning with Singapore's agentic AI framework which recommends bounding agents' powers upfront.