Our New AI VP of Finance Closes the Deal, Sends the Invoice, and Chases the Cash. It Took 4 Deals to Train It.

Our finance team went on vacation during SaaStr AI Annual. Our busiest time of the year.

That’s the whole origin story. Collections started slipping, sponsors and vendors weren’t getting billed, and the work that has to happen after a deal closes stopped happening, at the time of the year when time matters most.

It’s was time to bring in an agent ;). Our first instinct was to spin up a new agent for an “AI VP Finance”. Amelia didn’t. She instead built finance into 10K, the agent that was already our AI VP of Marketing. That decision turned out to matter more than the automation, and I’ll get to why.

Now 10K is our AI VP Revenue, and it’s even better than before.

What we built

One deal, start to finish, with nobody touching it:

T+60 seconds. Contract gets signed in PandaDoc. 10K has the document, has read it, and has flipped the deal to Closed Won in Salesforce, stamped with today’s date.

Minutes later. It scans who the contract went to, finds the signers who aren’t contacts on that account, and appends them. It creates the invoice in bill.com with the correct payment terms, including splits. It sends that invoice to whoever is named as the AP contact on the contract, not whoever we happened to be negotiating with.

The days after. It answers the customer’s questions about that invoice directly, back and forth, from our AP inbox. Real customers. They don’t know they’re talking to an agent.

Before due, on due, after due. It runs the reminder ladder without being asked.

7 days past due. It escalates to a human.

Month end. It calculates the AE’s commission off the deal, the payment terms, and the date the cash actually landed. We didn’t scope this one. The agent proposed it.

Continuously. It tells us what we can spend on ads next month, off revenue that was collected rather than revenue that was forecast.

That’s sales ops, AR, collections, sales comp, and a piece of FP&A. One agent, running on tools we already paid for, with no new system of record.

What it replaced

This is how it used to work here, and how it still works at most companies.

A deal closes. You wait for the AE to flip it to Closed Won in Salesforce so the automation fires. Except the AE doesn’t, because the deal is closed and they’re going to go have a sushi dinner and some champagne. Nobody flips a stage field on the night they close a deal. So somebody chases the AE, then flips the field, then sends the bill, then works out the payment terms, then remembers three weeks later that nobody collected.

The gap between a deal getting signed and an invoice going out is one of the least examined cash drains in B2B. It’s a coordination problem, and coordination is something agents handle well.

The 4-deal training curve

Finance is unforgiving, so this didn’t go live on faith.

Amelia tested the whole flow end to end herself first. Then, on the first three real deals with real customers, she ran it manually alongside the agent, one step at a time, approving each one.

Her prompt every single time: tell me what you plan to do before you do it.

That’s still how she works with any agent on anything sensitive. She hands it the contract and asks it to walk through what it’s about to do. Approves step one. Goes and checks Salesforce herself. Are the contacts there? Then asks for step two.

Deal one. It read the contract but missed the split payment terms and generated a single invoice for the full amount on a large deal. She caught it and corrected it.

Deal two. It did the same thing again, and the fix that mattered wasn’t correcting the invoice. It was telling the agent to build the rule into its process for every contract going forward. Agents will take a correction as a fact about that one deal rather than a fact about the job, so you have to spell out the second part.

Deal three. The customer didn’t exist in bill.com yet, which tripped it up. So they walked the branch together: what do you do if we’re already connected to them, and what do you do if they’re brand new?

Deal four. Fully autonomous and correct.

During testing it produced duplicate invoices and sent things to the wrong people, and it would have thrown the books off if any of it had gone out the door. Budget for that. If your contracts vary more than ours, budget for more than four deals too.

What still goes wrong (but ‘human on the loop’ addresses it)

One bad invoice since going live. It got the due date wrong for no reason we could reconstruct, and it hasn’t happened again.

Amelia is copied on everything the agent sends, so she saw it, corrected the invoice, and resent it. Cost us a few minutes. Human “on the loop” vs in the loop.

That CC line is what makes the rest of this acceptable. An agent emailing your customers with nobody watching is a different product with a different risk profile, and staying copied isn’t a launch precaution you graduate out of.

The other thing keeping it safe is that the agent stops when it isn’t sure. Yesterday a deal closed and 10K came back with a question before acting. It wanted to automate a step and asked whether the step was right first. You can build those checkpoints in deliberately, and in finance you should.

Why building it into 10K our AI marketing agent mattered

You could build a standalone AI VP of Finance that does everything in that list. If you’re a larger org with real finance controls and separation of duties, you may have to. Without that constraint, putting finance inside the agent that already owns revenue changes what the agent is capable of.

10K already knew our marketing spend, campaign performance, Salesforce pipeline, and event data. Adding finance means it now knows how much we should be spending on ads, and what revenue our marketing brought in this week off collected cash rather than a forecast.

Some people hear that and think it’s too much in one place. What we found is that the agent is only as good as the surface it can see.

10K told Amelia it could take over commission calculations. Its argument was that it already knew the AEs on every deal, the payment terms, and when the cash landed, so there was no reason to buy a separate tool. She gave it the commission rules. Month end got dramatically easier. Nobody scoped that feature or asked for it. It showed up because the data was in one place and the agent noticed.

Even the best finance people used to sit somewhere else in the building. They understood the revenue team’s numbers because they had to, but nothing about that understanding was linked to anything or actionable in the moment. Finance and revenue run as one system now, with the guardrails you want on the finance side. The invoicing automation is the visible piece. The single system underneath it is worth more.

What we didn’t build

The “AI replaces your whole stack” version of this story is wrong.

We did not rebuild bill.com. We did not rebuild PandaDoc. We didn’t touch QuickBooks, which everything still flows into. I don’t want to build e-signature and I definitely don’t want to maintain it. There’s deep infrastructure in those three products and none of it was the bottleneck.

We hooked tools we already pay for into an agent and then used them far more than we ever had. Same spend, very different leverage on it.

Most B2B companies are in the same position. The existing finance stack is fine. What’s missing is something that can operate it without being asked.

Top 3 Takeaways

1. Start where failure is loud. A wrong invoice surfaces immediately, which is what makes invoicing a good first finance workflow. Don’t hand an agent something where mistakes compound silently for a quarter.
2. Run 3 or 4 real transactions manually first. Approve every step, and keep “tell me what you plan to do before you do it” as the standing prompt.
3. Turn every correction into a rule. Fixing the invoice fixes one deal. You have to tell the agent to apply it to every contract from now on, because it won’t make that leap on its own.