How I built an AI chief of staff for $25 a day

A CEO built an AI chief of staff using Claude Code that handles roughly half the workload of a full-time chief of staff within its first week and costs no more than $25 in tokens per day, less than 5% of a full-time hire. The executive said the agent runs meeting prep, gathers company updates, and drafts strategy memos, but nearly delivered an inaccurate account summary before a customer call because it filled a data gap with a confident guess instead of flagging it. The fix was adding a Model Context Protocol (MCP) health check at the start of each session to flag missing data sources before the agent reports what it knows.

For 20 years, I built my career on the business side of startups: sales, marketing https://www.fastcompany.com/section/marketing , customer success, operations—areas where I never needed to learn coding or build something myself. That changed this year. My CEO role grew beyond my domain of expertise to span the company’s full operations, and I needed a way to keep up. After using Claude for almost a year, I started with Claude Code and realized I could replace the traditional functions of a human chief of staff with an AI https://www.fastcompany.com/section/artificial-intelligence solution. I built my own agent, turning memory and previous sessions into durable skills that help me gather intelligence across the business, connect dots I couldn’t see, and free me up for the decisions only I could make. Within the first week, it was already handling roughly half the workload of a full-time chief of staff, running meeting prep, gathering company updates, and drafting strategy memos. A few months later, I now use it daily, and it costs no more than $25 in tokens on a given day. That’s less than 5% of what a full-time hire https://www.fastcompany.com/section/hiring would cost. It’s taken quite a bit of iteration to get there. Agents have limits, and I learned that the hard way. Early on, I asked for an account summary ahead of a customer call. I asked it to use Slack, support tickets, product telemetry, CRM notes, and email for context. What came back looked complete and sounded certain. But it wasn’t. The source of truth for the customer’s revenue lived in an unconnected spreadsheet, and instead of flagging the gap, the agent filled it with an inaccurate guess. I nearly walked into that call with the wrong picture of the account. Here’s what I’ve learned overall: An AI agent will rarely tell you when it’s missing something. If a data source goes down or a permission expires, an agent will keep answering anyway, and a confident wrong answer is far more dangerous than a visible error. The fix isn’t a smarter model; it’s building the right skills into the agent. I now have the agent conduct a Model Context Protocol MCP health check at the start of each session, before it does anything else. It flags any missing data sources before it tells me what it knows. To do this, first set up all relevant MCP connections to your sources of truth, check the connection in Claude Code, and then build a skill to repeat that check at the start of sessions anytime you make a prompt or enter the agent’s command. Until I built this agent, I lived in many different SaaS dashboards. My bookmarks bar is a big collection of icons pointing to the various systems where important data lives and gets aggregated into views. But an agent doesn’t need visual dashboards. It goes straight to the data and aggregates it across systems, in whatever display you specify. For the customer health check I mentioned above, data comes from our CRM, support queue, product telemetry, Slack, internal docs, and meeting notes. The summary I get when I use this skill is more comprehensive than anything I’ve pulled manually because of the agent’s ability to pull a coherent narrative across data sets and disparate systems. The LLM’s intelligence cuts across domains, and it can summarize data that previously needed executives to reach down into their teams for progress and project-level information. Set up right, it can cut days of back-and-forth to get to the very specific detail you need, and it dramatically increases your grasp of the specific work that’s being done or problems you need to solve across the team. The fastest path to mediocre results is to shortcut specific iteration and try to use generic skills or an off-the-shelf setup. An agent is only as good as its tools and context, and you need to invest in the specific tools your business uses and the context that exists only in your communication, documents, and the thinking that goes into your skills and prompts. This encodes your edge cases, your rules, and the judgment specific to how your company actually runs. Two founders both ask the same question, “What is the progress on our road map this week?” The founder who gives it a specific context doc on quarterly goals and road map, a Jira MCP to pull the latest ticket status, notes from last week’s standup, and access to the Engineering Slack channel will get a far better result than the one who points at the board slides from last quarter and asks for an update. If you’re trying to build your own agent, you’ll have access to the latest and best models, and you’ll have tools that will help you get initial value quickly. But you will be what really makes this work. The more you use the tools, the stronger the memory and skills become. That value can’t be replicated by another team member or outsourced to a point solution. It requires your interactions, prompts, and corrections over time. If you take that time, you can replace a $200,000-level employee with roughly $5,000 in annual token spend. The management overhead is different—it’s more intense and creative than just having 1:1s and taking notes. But it’s about the same effort, and the results are much more consistent. If you take anything from my experience, start right now. Do not wait to use these tools or set up the perfect structure. Your interaction with Claude or ChatGPT is already helping build the context and memory that will enable you to get value from a chief of staff agent. You can start small. Ask for an agent that writes your weekly team update, and give it access to the systems that the update depends on: your goals, your metrics, the actual work happening across teams. You can be specific. Give it a discrete task you would have done yourself, like writing a job description for a new hire or pulling supporting data on a problem or opportunity you recently surfaced. You’ll get a double win as you’ll iterate on the skills and build more context while you complete the new work. Our job as leaders is to make the best possible decisions with the available information. What’s available just got exponentially better, and it’s our responsibility to learn how to use it. The CEOs who succeed in the intelligence era will master this skill, and the time to get started is right now.