SaaStr Built an AI VP of Finance. It Took Four Deals to Trust It.

SaaStr didn’t buy new finance software to automate its deal-to-cash pipeline — the same company that ran its Q1 2026 sales operation with 1.25 humans and 20+ AI agents has now pushed the same playbook into finance. It repurposed an existing AI agent, gave it access to the tools already in use, and let it fail on three real deals before it worked.

What Happened

SaaStr’s AI agent — internally called 10K, originally built as an AI VP of Marketing — was handed accounts receivable, collections, and commission calculations without any new system purchases. It connects directly to PandaDoc for e-signatures, bill.com for invoicing, Salesforce for deal tracking, and QuickBooks for the books. Once a deal closes in Salesforce, 10K tracks it within 60 seconds, generates the invoice with correct payment terms and splits, and starts a collections workflow if payment lags.

It didn’t work on the first try. Deal one and two missed split-payment terms, forcing the team to build explicit rules for how multi-party contracts should be invoiced. Deal three surfaced gaps in the customer database that needed branching logic to handle. Only on deal four did the agent run the full sequence — invoice, terms, commission calculation — without a human catching an error.

Beyond invoicing, 10K also handles customer invoice inquiries, runs commission calculations based on payment terms and actual cash-receipt dates rather than booking dates, and generates budget recommendations tied to revenue that has actually landed in the bank rather than revenue that’s merely been signed. That distinction — cash collected versus cash promised — is the same discipline a controller enforces manually; the agent just runs it as a standing rule instead of a monthly reconciliation exercise.

Why This Matters for Operators

The standing instruction the team gave the agent during testing was simple: “Tell me what you plan to do before you do it.” That’s the operating model worth copying — not full autonomy from day one, but a supervised ramp where the agent proposes actions and a human approves them until the error rate drops to zero. Four deals is a strikingly small number of training examples for a function (invoicing, collections, commission math) that most finance teams assume needs a dedicated tool purchase and a multi-month implementation.

Even post-launch, the system isn’t unsupervised. One invoice went out with an incorrect due date; a team member caught it via email CC and corrected it within minutes. SaaStr describes this as “human on the loop” rather than full automation — the agent acts, but a person is positioned to intercept mistakes before they compound.

This matters because a growing category of dedicated AI billing and AR platforms — the kind that promise contract-to-cash automation out of the box — are pitching themselves as the fix for exactly this workflow. SaaStr’s version is a reminder that the underlying capability isn’t gated behind a new SaaS purchase; it’s gated behind tool access and a willingness to supervise the agent through its early, error-prone deals. A dedicated platform may still be the faster path for teams without an existing capable agent to redeploy — but the capability itself is now general-purpose, not proprietary to any one finance-AI vendor.

Source: SaaStr

Content created with AI assistance and reviewed for accuracy.