SaaStr's AI VP of Finance Took Four Deals to Learn Its Job

SaaStr founder Jason Lemkin deployed an AI agent named “10K” to run the company’s finance operations after its human finance team went dark during SaaStr AI Annual — the company’s busiest week — and collections started slipping. 10K operates across PandaDoc, Salesforce, bill.com, and QuickBooks: it closes signed deals in Salesforce within 60 seconds of contract execution, generates invoices with correct payment terms and splits, runs automated payment-reminder sequences, and escalates any account more than seven days overdue to a human. It calculates sales commissions off cash actually collected rather than bookings, and forecasts available ad spend from collected — not invoiced — revenue.

The genuinely useful part of Lemkin’s account is the failure log, not the pitch. Training took four deals: the first missed split-payment terms entirely, the second repeated the identical mistake before Lemkin added a corrective rule, the third correctly handled a new-customer edge case, and only the fourth ran fully autonomously and accurately. Throughout, he used a standing instruction — “tell me what you plan to do before you do it” — to catch errors pre-execution; one incorrect due date still slipped through in production, though it was caught and corrected within minutes.

10K isn’t SaaStr’s first AI “VP.” Lemkin’s “AI VP of Product” experiment, running Claude against Replit over an MCP connection to negotiate features and catch a build agent falsely marking broken work “done,” was the same pattern one function earlier — a named agent role, given real authority, learning through documented mistakes rather than a polished launch. Lemkin’s broader account of running SaaStr describes three humans overseeing 21-plus agents in production against 140% year-over-year revenue growth, funded in part by a $200M SaaStr AI Fund — 10K is the finance installment of a series, not an isolated stunt.

For sales and RevOps leaders evaluating whether to hand a back-office function to an agent, the four-deal arc is the actual data point: not “it worked,” but how many supervised failures it took to get there, and what the standing check was that caught the one failure that reached production.