SaaStr built an AI agent that handles the entire post-close workflow: contract to Salesforce flip, invoice generation, collections, and sales comp calculation. The agent runs autonomously with a human copied on emails, not approving each step.
The build came out of necessity. Finance team was out during SaaStr AI Annual, their busiest period. Collections slipped, sponsors were not getting billed, and the gap between signed deal and invoice was costing them cash.
The workflow: Contract signs in PandaDoc. Agent reads it, flips the deal to Closed Won in Salesforce within 60 seconds, creates the invoice in bill.com with correct payment terms and splits, sends it to the AP contact listed in the contract. Then it runs the reminder ladder, answers customer questions from the AR inbox, escalates at 7 days past due. Also calculates AE commission based on actual cash landing, not forecast.
Training took 4 real deals. Deal one: missed split payment terms. Deal two: made the same mistake until told to build the rule into every contract going forward. Deal three: did not know what to do when customer was not in bill.com yet. Deal four: fully autonomous.
One bad invoice since going live. Got the due date wrong, no clear reason. Amelia caught it because she is copied on everything the agent sends. That is the safety mechanism: human on the loop, not in the loop. Agent also stops when it is not sure and asks before acting on edge cases.
The agent lives inside their existing AI VP of Marketing, now rebranded as AI VP Revenue. They did not spin up a separate finance agent. That architectural decision matters because it keeps context across the deal lifecycle: marketing to close to cash to comp.
This is post-sale workflow automation, not a new rev ops platform. It runs on tools they already pay for: PandaDoc, Salesforce, bill.com. No new system of record.
Worth noting: this works because their contracts follow patterns. More variance means more training deals. Budget for duplicates, wrong recipients, and invoice errors during testing. If you let those hit customers before a human checks them, you are running a different risk profile.