Building My AI Chief of Staff: Lessons from Harvard’s One-week AI Intensive
In this month’s update, I provide a look into my one-week intensive course at Harvard, studying Claude Cowork as a “coding” system, and learning how to think of AI as an active set of resources, operating as a compliment to my work. I’ll share what I walked away with and thoughts on implications for the market.
When Mark Zuckerberg publicly announced he was building his own personal AI agent, it caught the business world’s attention for a particular reason. Here is a man with every conceivable firm resource at his disposal — thousands of engineers, billion-dollar infrastructure, and direct access to the world’s most advanced AI research. Yet rather than simply delegating the task, Zuckerberg reportedly spends five to ten hours per week personally overseeing and shaping his own AI agent. The signal was clear: building your AI is not a task you outsource. It requires your voice, your judgment, your context. That story was the catalyst that drew me to Harvard’s 1-week Intensive.
The ‘OpenClaw Strategy Intensive’, held at the Harvard Data Science Initiative was a practitioner-oriented program designed to close the gap between AI curiosity and AI capability. The course introduced a pivotal conceptual shift: AI has moved beyond the “copilot” model — where the agent’s verb is answer — into the era of the “colleague,” where the verb is act. Participants learned not merely to prompt an AI but to architect autonomous agents that take initiative, execute workflows, and adapt over time.
The centerpiece of the program was a hands-on build: each participant left with a functioning AI Chief of Staff. Mine operates on a daily schedule, entirely on its own. Every morning it goes directly to my email, processes my inbox, and delivers a personalized summary in my specific format — prioritized, categorized, and actionable. It also reviews my calendar for the day ahead and prepares me for each meeting by pulling together notes, relevant context, and web-sourced background on each person I’m meeting with.
What once consumed the first thirty to sixty minutes of my morning now arrives before I pour my coffee. The agent is built on four components — Data, Skills, Connectors, and Schedule — and runs autonomously through Claude Cowork. It does not wait to be asked. ACT (not just answer)
This brings me back to Zuckerberg. The deeper lesson of his example is not about technology — it is about ownership. An AI agent reflects the person who built it: their priorities, their communication style, their definition of what matters. No engineer, however talented, can hand you that. Research presented during the course found that high performers redesign their workflows 2.8 times more frequently than laggards, precisely because they remain close to the machine. The most effective AI twin is not the most powerful one; it is the most personal one. Zuckerberg’s hours spent shaping his agent are not inefficiency — they are the investment that makes it genuinely his.
I left Harvard with something more than a working agent. I left with a framework for thinking about AI, not as a tool to be handed off, but as a capability to be actively shaped. The question is no longer whether autonomous AI belongs in your workflow. The question is whether you are close enough to it to make it truly yours.
TAKEAWAYS:
1. Humans will need to be involved. As AI moves into “Agentic” phase, it will actually require more human involvement to be effective. This pushed the timeline on AI’s true impact and benefits felt from my perspective.
a. Because there’s still Level 3 Complex tasks that can’t be taught quickly enough or with enough precision around unique situations to make sense to offload to an agent
b. Even an effective AI agent must be initially trained… and then monitored, as business elements change, so should the agent.
2. Anyone can build this. You don’t need to know how to code. Claude CoWork, which is a specific platform that they launched (different from Claude Code) allows you to type out instructions in plain English, in your own voice, and then takes that and interprets it back to the computer.
3. It’s cheap. It’s basically an additional employee for the $20/mo Claude cost
4. But your tech setup matters. Some work setups control or limit
5. Governance remains a question. During the course they shared that Europe in general has rules and regulations around using these tools, and so European businesses don’t have access to this tool. That impacts investment outlook.
6. Governance remains a question #2. Further, who owns your AI agent (your Chief of Staff) if you leave a company? Under today’s standard employee contract, your company owns anything you produce while working on company time or company equipment, making it pretty clear that in today’s regulatory environment your firm owns your agent.
7. It’s coming, whether we like it or not. Meta & Klarna. I sat in class with a gentlemen from Meta who shared that their policy is that everyone in the company will build their own AI Chief of Staff, like Mark is. This is non-voluntary. To be clear, it’s mandatory for employees. But even as firms push further, Klarna offers an early example of firms rolling out AI without strong process. After laying off an entire department of service professionals, they had to hire them back when the AI customer service failed to deliver suitable results.
The opinions voiced in this material are for general information only and are not intended to provide specific advice or recommendations for any individual.