The conversation around AI agents has moved quickly from demos to organisational design. Microsoft’s 2025 Work Trend Index for Singapore reported that 56 per cent of Singapore leaders were already using agents to fully automate workstreams or business processes, while 46 per cent expected their teams to build multi-agent systems.
I understand the appeal. I built an internal AI office with 21 defined roles across strategy, finance, marketing, sales, customer success, delivery, engineering, design, quality assurance, security, research and data.
I expected the hard part to be technical. It was not. The harder questions were managerial: Who owns the work? Who reviews it? What can an agent decide on its own? When does a human step in?
Most of my 21 roles are real and used in live work today, but most orchestration is still interactive. That distinction matters because the biggest lesson was not how to remove humans. It was how to design responsibility around machines.
The first mistake: Capability is not ownership
My earliest agents were too broad. Help with marketing or help with development sounded reasonable because modern models can do many things.
The problem was not output quality. It was ownership.
If one agent can research, write, publish and evaluate the result, who is actually responsible for the job? If something is wrong, where does the failure sit?
I eventually stopped defining roles by what the model could do and started defining them by what the role should own.
A writing role writes. A publishing role prepares distribution. A research role scouts for information but does not install what it discovers. A builder builds, but does not declare its own work production-ready.
That sounds obvious in a human company. With AI, it is easy to forget because the same underlying model may be capable of doing all of those jobs.
Also Read: When a slot opens, let the AI agent act – within limits
The second mistake: Letting one agent close the loop
I also learned that an agent should not be allowed to create, review and approve the same important output.
When a builder tests its own work and then declares it ready, the process can look efficient while removing the independent challenge that catches weak assumptions.
So I separated three responsibilities for higher-impact work: builder, reviewer and acceptor.
The builder produces the work. A separate role tests or reviews it. Final acceptance of important client-facing, production or irreversible actions still comes back to a human.
Not every task needs three layers. Drafting an internal note does not need the same controls as deploying software or sending something to a client.
The point is proportionality. The higher the impact of being wrong, the more important independent review becomes.
The third mistake: Treating autonomy as a maturity score
At first, I saw autonomy as progress. If an agent could do more without asking me, the system felt more advanced.
I no longer think that way.
The better question is: what happens if this action is wrong?
Singapore’s updated Model AI Governance Framework for Agentic AI uses a similar risk-based approach. One case study in the framework tiers actions by severity, reversibility and the feasibility of human oversight. Low-risk and reversible actions can be automated. Moderate-risk actions require human approval. High-risk actions with limited reversibility may be blocked entirely.
That logic changed how I design workflows.
An agent drafting an internal summary can have wide freedom. An agent changing permissions, deploying code, committing money or sending an external communication should face a much higher bar.
More autonomy is not always better. Appropriate autonomy is better.
The fourth mistake: Managing the agent but not its access
An agent is not just a prompt.
It may have access to tools, databases, files, connectors, scheduled jobs and external services. As my system grew, I realised the real management problem was not simply how many agents do I have. It was what can each of them touch.
That pushed me to maintain an inventory of the agents, tools and capabilities around them.
This aligns with Singapore’s agentic AI framework, which recommends bounding agents’ powers upfront, including their autonomy and access to tools and data.
For founders, this is easy to overlook because experimentation moves quickly. A useful prototype can become part of an operating process before anyone has paused to document what permissions accumulated around it.
I now treat that inventory as basic operating hygiene, not administrative paperwork.
Also Read: From KYC to KYA: how AI agents are reshaping payment risk
The most important lesson: Humans did not disappear
The fully unattended version of my AI office is still a work in progress.
Some workflows run automatically. Many still begin with a human instruction. Important outward actions still need approval.
I once saw that as a sign the system was unfinished. Now I think the better goal is not zero humans. It is humans only where judgement, accountability or relationships genuinely matter.
The AI roles can absorb repeatable work, prepare options, check outputs, monitor systems and move tasks forward. The human role becomes narrower but more important: direction, trade-offs, relationships and final accountability.
That is also why I am cautious with the term AI employee. An employee is not valuable because they can perform many tasks. They are valuable because there is clarity around what they own, what authority they have and how their work fits with everyone else.
What I would tell another founder
Before adding another agent, I would ask four questions:
- What job does this agent actually own?
- What systems and data can it access?
- Who independently checks important work?
- Which actions can it take without a human, and why?
If those answers are unclear, adding more agents does not create an AI workforce.
It creates more capability without more management.
That was the biggest change in my thinking. I started by trying to build better agents. I ended up redesigning the organisation around them.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post I built a 21-role AI workforce. The hardest part was management appeared first on e27.