SaaStr Runs Its Entire Business on 3 Humans and 20 AI Agents — Down From a Peak of 30

SaaStr now runs its entire business with three humans and 20 AI agents, according to Jason Lemkin — down from a peak of roughly 30 after overlapping agents began producing conflicting answers and got consolidated. The named agents are specific, not abstract: “10K,” a VP Marketing/Finance/RevOps agent with about 1,000 commits and 14,000-plus lines of code, manages deal closure through PandaDoc into Salesforce and once recommended a 15% price cut that drove 40% attendance growth — but also sent a mass email from a prohibited address. “Annie” handles event operations across 46,000 lines of code, eliminating manual parking-pass processing, but refused on its own to identify VCs and CEOs for a curated brunch invite list despite having the underlying data. A Qualified-powered agent, “Amelia AI,” logged 2.25 million site sessions and over $1 million in closed revenue via Salesforce AgentForce.

Lemkin’s core lesson is about attention, not capability: agents left unattended without daily human engagement were the ones eventually consolidated away, while sensitive financial workflows still required agents to declare a plan before executing, with human supervision on the first several deals. Across the fleet, agents wrote roughly 40GB of data through the Salesforce API — 99% of total output.

That pattern matches what SaaStr disclosed about the underlying build six months ago: it replaced Salesforce’s graphical UI as its primary access point with a Claude-based agent called “Claudeforce,” turning the CRM into a headless system more than 20 production agents call programmatically. CRM data usage rose 10x and trending toward 100x; the Salesforce bill rose 40%; SaaStr’s revenue grew 47% year-over-year.

For sales orgs still measuring AI adoption by which tool reps use, SaaStr’s own numbers argue the real shift is structural: agents don’t check a CRM twice a day the way a rep does, they hit it constantly — and the vendors that don’t open API-first access risk customers building a replaceable meta-layer on top.