A startup CEO has published a detailed account of turning Claude Code, Anthropic's coding-focused AI tool, into a personal chief of staff. The system handles meeting prep, company-wide updates, and strategy memos, and it costs roughly $25 a day to run.
That is a striking number next to what a real chief of staff costs. Salary trackers put average US pay for the role between $150,000 and $220,000 a year, and a 2025 industry salary report placed the figure at $165,573, up more than 7% from the year before.
The piece making this possible is a connector standard called the Model Context Protocol, which Anthropic introduced in late 2024. Think of it as a universal plug that lets an AI tool link directly to the software a company already uses, its sales database, support tickets, or team chat, instead of a person copying information between screens by hand.
That direct access is exactly where the story needs a warning label. The CEO describes asking the agent for a customer summary before a call, and the tool pulled data from five different systems but missed one spreadsheet holding the real revenue number. Instead of saying it did not know, the agent guessed, and the guess sounded just as confident as everything else in the report.
This is not a one-off glitch. It echoes a well-documented pattern: an airline was once ordered by a tribunal to honor a discount its support chatbot invented on the spot, because the tool answered a question it had no real basis to answer. Any tool built on a language model will do this unless someone forces it to check its own sources first, which is exactly the fix the CEO built: a health check that runs before every session and flags missing connections before the agent says anything else.
The wider trend backs up why this matters now. Most companies are already testing AI agents in some form, according to recent executive surveys, yet analysts at Gartner expect more than 40% of agentic AI projects to be shut down by 2027 because the payoff never showed up or the risk controls were never built.
That gap between quick pilots and agents built with real guardrails is the whole story here. The savings in this case are real, but they came from months of correcting the tool, wiring it into the right systems, and building checks for its blind spots, not from installing an app. Anyone hoping to copy this needs to budget for that unglamorous setup work, not just the $25 a day.
The lesson for any business is not "hire less." It is that senior-level coordination work, the kind that used to require a dedicated person watching every system, can now be automated cheaply, but only if someone puts in real time teaching the tool where the company's actual truth lives, and only if it is built to admit what it does not know.