A CEO's $25-a-day agent filled the one gap in its data with a confident guess

The summary depended on six sources and five of them were connected. Slack, support tickets, product telemetry, CRM notes and email were all wired in ^[7]; the customer's revenue number, the figure that would decide how the call went, sat in a spreadsheet that was not ^[8][4]. An agent short one of six inputs hands back a whole answer rather than five-sixths of one with a labelled hole, delivered in the same register as the answers that happen to be right ^[8], and the author says he nearly took that into the meeting ^[9].

A day of the agent runs no more than $25 in tokens, which he calls less than 5% of a full-time hire ^[5][6], so the hire he is comparing against costs upward of $500 for the same day ^[1], or roughly $125,000 a year against $6,250 in tokens if you annualize both over 250 business days ^[2]. Set that beside his other figure, roughly half a chief of staff's workload in the first week ^[4], and the work absorbed came in under a tenth per unit of the human cost ^[3].

What is described here is narrower than headcount substitution: a read layer across systems the company already pays for. He used to work out of a bookmarks bar full of SaaS dashboards, and the agent goes to the data instead, pulling one narrative out of CRM, support queue, telemetry, Slack, internal docs and meeting notes ^[12][13]. What the piece reports is replacing the functions of a chief of staff, not a role left unfilled or a person let go ^[16]. His half-a-workload number describes week one ^[4]; the months-later passage gives the daily habit and the token cost ^[5].

For any recurring answer, two axes sort what to hand over next. Either every source it depends on is connected or at least one lives off-system. And either the output stays inside the team or it goes to a customer. Internal prep with full coverage is the safe corner to start in. That account summary sat in the opposite one, partial coverage and an external audience, where a guess becomes something said out loud to a paying customer. What he added is an ordering: a Model Context Protocol health check runs first in each session and names the missing sources before the agent says anything about what it knows ^[11]. His two-founders comparison argues the same from the other direction, since the identical question about roadmap progress gets a better answer from a quarterly goals doc, a Jira connection, standup notes and access to the engineering Slack channel rather than from a better model ^[15].

On his own numbers the tokens are the cheap input; the expensive one is the inventory of sources behind each recurring question, and whether the agent reads that list back before it starts talking.