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

Over the past two decades, I cultivated a career in the business operations of startups, encompassing roles such as sales, marketing, customer success, and operations. However, this all changed this year when my responsibilities as CEO expanded to encompass the entire company. To manage this newfound scope, I began utilizing Claude and later Claude Code.

Realizing the potential of AI, I conceived an AI chief of staff, equipping it with durable skills to gather intelligence across the business, identify connections, and delegate tasks that only I could handle. Within the first week of implementation, the AI agent managed around half of the workload of a full-time chief of staff, handling meeting preparations, generating company updates, and drafting strategy memos.

Within a few months, it became an indispensable daily tool, costing no more than $25 in daily tokens, which is less than 5% of the expenses associated with a full-time employee. However, this journey was not without its challenges. Agents, despite their capabilities, have inherent limitations. Early on, I encountered a situation where I requested an account summary prior to a customer call.

The agent, drawing from various sources like Slack, support tickets, product telemetry, CRM notes, and email, produced a comprehensive response. Yet, upon closer inspection, it became apparent that the source of truth for the customer's revenue was housed in an unconnected spreadsheet. The agent had erroneously generated an inaccurate estimate.

This incident underscored the importance of building specific skills into AI agents. I now implement a Model Context Protocol (MCP) check at the onset of each session, identifying any missing data sources before the agent provides its insights. By establishing connections to all pertinent sources, verifying these connections, and incorporating a skill to execute the check at the beginning of each session, the agent ensures that it operates with accurate and complete data.

Another crucial lesson is the agent's ability to extract value from disparate systems, rather than relying on visual dashboards within individual SaaS platforms. By directly accessing the data and aggregating it across various systems—CRM, support queues, product telemetry, Slack, internal documents, and meeting notes—the agent generates a more comprehensive summary than manual methods.

This capability significantly enhances the understanding of team activities and problem-solving processes. However, the effectiveness of the agent is contingent upon a tailored setup. Generic setups tend to yield mediocre results. Investing in specific tools and contextual information unique to your business is paramount. For instance, a founder seeking progress updates on the road map will receive superior results when provided with specific context documents, Jira MCPs for ticket statuses, standup notes, and access to relevant Slack channels compared to generic road map slides from the previous quarter.

Lastly, building an agent is not a task that can be outsourced. While access to advanced models and tools is readily available, the true value lies in the continuous interaction, prompts, and corrections made by the user. The more one engages with the agent, the more robust its memory and skills become. This personalized effort allows for the replacement of a high-level executive at a fraction of the cost, with the agent costing only about $5,000 annually in tokens, compared to the potential $200,000 salary of an executive.

The journey from my experiences serves as a blueprint for other CEOs. Start immediately, even without a perfect setup. Begin with simple tasks, such as having the agent generate weekly team updates using access to relevant systems like goals, metrics, and ongoing team activities. Specificity is key—assign the agent discrete tasks, like drafting job descriptions or analyzing recent opportunities.

This initial step will initiate the building of context and memory, paving the way for significant efficiency gains and cost savings.