Everyone keeps telling you the future is 100 specialized agents. In our stack, the opposite is happening.
On the latest episode of The Agents, Amelia and I walked through the newest addition to our stack: an AI VP of Finance that went into production this week. The part that surprised even us is where it lives. It didn’t get its own app or its own agent. Instead, it runs inside 10K, our AI VP of Marketing, the Replit-built agent that already handles our email, content, and a growing pile of go-to-market work.
The industry consensus is that every company will eventually run a hundred narrow agents, one for each function, each with its own login and its own silo. What is actually happening inside our stack is the opposite. Our agents are collapsing into each other. Fewer of them, each going deeper, all drawing on one shared body of knowledge about how our business runs. Closer to a monorepo than an app store.
We run SaaStr with 3 humans and more than 20 agents in production. And this episode we had our first-ever guest on The Agents: Sam Blond, founder and CEO of Monaco, which happens to be one of the agents we run. So we got to grill the founder whose product is on our own team.
Below is how we built the newest agent, the four systems we wired it into, the three things that broke while we were doing it, and a long conversation with Sam about the question we get more than any other: does this only work because you’re SaaStr?
Why We Built an AI VP of Finance: Collections Had Fallen Six Figures Behind
You can point an agent at almost anything right now, and that is exactly the problem. The number of workflows you could automate this quarter is effectively unlimited. Most of them are not worth the time it takes to build and maintain them. The filter we used was simple and old-fashioned: if it isn’t broken, don’t touch it. Find the thing that is actually broken and start there.
For us, the broken thing was collections.
Anyone who has run a startup understands why. When you are a public company generating cash at 40% operating margins, a slow-paying customer is a rounding error you can ignore for a quarter. When you are a lean team, the timing of that cash is the difference between paying commissions and bonuses on schedule or not. Our collections had degraded, and the reason was human in the most ordinary way. Our part-time finance team fell behind, someone went on vacation, and the backlog never got caught back up. We ended up six figures behind on money we had already earned.
Collections is also work that people quietly avoid, because it is awkward. Our sponsorships run from roughly $45K for a media package up to $400K for a large event sponsorship, and asking a real company to pay for a booth 60 days after the event is uncomfortable enough that it slips to the bottom of the list. It stays there until the money is late enough that you sometimes have to write it off entirely. That is the worst outcome: revenue you closed, delivered against, and then lost on the back end because no one wanted to send the third reminder.
That specific, expensive pain is what justified building an AI VP of Finance. Everything else it now does, and it does a lot, would not have cleared the bar on its own. The high-pain use case is what made the whole project worth starting. It is the same logic behind why you should not vibe code your own Salesforce. You technically could. But it is too deep, too collaborative, and too load-bearing to rebuild for sport. Start where it hurts, not where it is fun.
The Four Finance Tools We Connected, and Exactly How Long Each Took
The AI VP of Finance is not a standalone app. It is 10K sitting on top of the APIs across our finance stack, which is why it can see the whole picture instead of one slice of it. We connected four systems, and the effort ranged from trivial to genuinely annoying.
- bill.com took under 10 minutes, and our agent found feature in it we didn’t even know about. The agent’s first instinct was to tell us to start with bill.com, not because it expected it to be easy, but because it thought we would have to email support for an API key and wait a couple of days. It was working off bill.com’s own support docs, which still say exactly that. In reality, any admin can generate the key directly, so the actual integration took less time than the agent predicted the paperwork would. It runs mostly read-only. It can write when you explicitly ask, but not for everything. When Amelia needed to change which ledger account a bill posted to, the agent correctly said it could not do that one and told her to log in and change it herself. That restraint is a feature, not a limitation.
- QuickBooks took about an hour and was the hardest of the four. It requires an Intuit developer account, which means passing a security questionnaire before you get access. None of us had one. It was a real pain, and it is also probably the right amount of friction for something wired into your books. It earned its keep immediately. When our tax team sent over a set of questions I did not fully understand, I handed them to the agent, and it did two things: it translated what the tax team was actually asking into plain English, and it drafted the answers using the real numbers pulled from QuickBooks and bill.com. I was not about to wing tax questions, and I did not have to.
- Brex took about 60 seconds. We use it for card spend, and our Brex account also holds receivables and receives the Stripe deposits from self-serve ticket sales. So the agent needs it to see upcoming card bills and incoming cash in one place, not just the invoiced side of the business. Without Brex, the agent has half the cash picture.
- PandaDoc was same-day, once support turned on API access. This is how the agent reads the contract the instant a deal is signed. Getting there was a small archaeology project. The agent first told me to ask my sales rep, and when I went to find out who our rep even was, the account no longer listed a human contact, which is its own quiet commentary on where vendor relationships have gone. I emailed support instead and they enabled it the same day.
The pattern underneath all four: start by connecting agents to real APIs. Systems of record already have guardrails, permissions, and audit trails built in. You get the leverage of an agent without giving up the controls. And, as we found repeatedly, the agent starts telling you about capabilities in those tools you never knew existed.
What Actually Happens in the 60 Seconds After a Deal Closes
It helps to see the before and after side by side.
Before the agent, a signed deal went on a slow relay. Whoever closed it sent the contract to our part-time finance team. Finance does not have a Salesforce login, so nothing moved in the CRM. Whenever they next logged in, they read the contract and created the invoices in bill.com by hand. On a good day that happened same-day. On a normal day it lagged into the next day. Contract-to-invoice was measured in hours or a full day, and the collections follow-up after that was inconsistent at best.
Now, the moment a deal closes, 10K runs the whole sequence itself:
- It reads the entire contract. Early on it tried to skim for payment terms, and we had to push it to read all 20 pages, because the terms that matter, split payments, non-standard schedules, anything off our template, are scattered through the document, not sitting in one clean field.
- It flips the opportunity to Closed Won in Salesforce and attaches the contract.
- It updates the contacts in Salesforce.
- It creates the invoices in bill.com off the actual contract terms, including split invoices when the deal calls for them.
- It sends the invoice and turns on automated payment reminders so collections runs on its own from there.
Contract-to-invoice is now under 30 seconds. As a small tell of how fast it is, 10K’s own “deal closed” alert beats PandaDoc’s native notification every single time, even though 10K is running on PandaDoc’s API to detect the signature in the first place.
Then this week we added a fifth step: commissions. The agent already holds the contract, the Salesforce record of who closed the deal, the invoices with their payment terms, and whether the deal is media or event. It had every input required. We handed it our commissions policy, and about 30 minutes later it was calculating commissions as step five. We never evaluated a third-party commissions tool, and now we never will.
The larger point is the compounding. Each capability we added made the next one obvious. The moment the agent could calculate commissions off the contract, it started hinting it could generate the contract too. No individual human, however good, executes that entire chain in under a minute, and in most orgs finance does not even have the Salesforce access to attempt half of it.
The Agent Knew Features of Our Own Software We Had Never Turned On
The clearest example is small and slightly embarrassing. bill.com has built-in automated invoice reminders. You can set them to nudge a customer five days before a bill is due, one day after it is past due, and ten days past due. We had been paying for bill.com for years and had never switched them on. The agent flagged it within minutes of connecting. Our own humans lived in that product every week and never bothered.
This kept happening, so it is worth naming as a pattern rather than a fluke. When you connect an agent to an API, it arrives knowing the product better than the people who have used it for years, including the parts of the documentation that are out of date. bill.com’s own doc still claims you have to ask support for an API key, and you don’t. The agent has read everything, and it has no habits to unlearn.
There is a strategic consequence that changes how you think about your stack. Some products get materially more valuable the second you put an agent on top of them. bill.com is worth far more to us now. Salesforce is worth more. QuickBooks is still QuickBooks, but we will stay on it longer than we otherwise would have, purely because we can finally talk to it. Other products get less necessary. We originally bought PandaDoc to lock down a larger sales team that was editing contracts and changing terms without telling us, which at the time made it one of the best purchases we made. With three people, that risk is gone. 10K can already generate a contract, produce the PDF, route it, and email it. The one piece we would not rebuild ourselves is certified e-signature. So the open question is no longer academic: do we keep PandaDoc at all, or bolt a cheap e-sign API onto the agent and drop it?
Running Finance Headless Means You Can Finally Query Your Own Business
We already run Salesforce headless, and it is the reason I can query our customer and sponsor data at all. For ten years I effectively never logged into Salesforce. Making it headless is what turned it into something I can ask questions of.
The finance agent does the same thing for money, and the list of questions I could never previously get answered is long and basic. What is our sponsor renewal rate over the last four years? Where can we predict revenue landing? Are we ahead of plan or behind it right now? What will cash look like at year end? Can we take distributions or bonuses? For a decade the honest answer was that I would wing it, because getting a real answer meant a project no one had time for.
This is the specific gap that outsourced finance and part-time contractors cannot close, no matter how good they are. The person who keeps your books may not understand how cash actually flows through your business. The person who closes your books may not know who your customers are or why they sit in which category. Very few finance people can build a forward model of revenue 12 months out, and the ones who can need enough day-to-day familiarity with your cadence to do it well. An agent that can see every system at once does not have that ramp. It already has the context. You ask, and you get an answer in seconds instead of a scheduled meeting three weeks out. That is a different relationship with your own numbers than most founders have ever had.
Three Things That Broke, and Why Agents Are Not Set-and-Forget
The failures are more instructive than the wins, so I want to be specific about three of them.
Qualified was still selling SaaStr AI 2026 tickets a full month after the event ended. I went to test our inbound agent for a post, talked to Digital Amelia on the site, and it offered me a discount to a conference that had already happened. Part of that is on us. I had not gone in to explicitly override its knowledge and tell it the next Annual is 2027. But Qualified re-crawls our site every day, and we had already published the 2027 dates, so it had the correct information available and did not propagate it. The fair grade is a C+ or B-, not an F. The root cause is one we keep running into: too much data breaks guardrails. Qualified had so many articles referencing “SaaStr Annual, May 2026” that the stale majority drowned out the correct update.
Our VC pitch deck grader, after roughly 5,000 decks, suddenly started giving everyone an F. The cause was not the model. It was me. I had added a 15th guardrail to the system, and it could no longer hold all of them at once, so it defaulted to failing everything. More rules did not make it stricter. It made it break. That is the counterintuitive part worth remembering: past a certain point, adding guardrails degrades the system instead of tightening it.
Qualified has no agent you can talk to, and that is the deeper limitation. It is classic rules plus a backend workflow studio. If it had a built-in agent, I could have simply typed “when the event is over, flip everything to 2027” and moved on. You cannot run true self-serve without that conversational layer, which is why Klaviyo made building one the headline of their 2.0. Their 1.0 did not have an agent to talk to. Their 2.0 does. Any product that expects to be used by other people’s agents needs one.
The lesson running through all three: our intuition about what an agent does is frequently not what the agent actually does. I assumed an agentic product would notice it had gone stale and update itself the moment new documents flowed in. I still think I was about 95% right to expect that, and I was wrong at the software level, which is the level that matters. If you approach these systems as a builder who assumes things will silently break, rather than a user who assumes they will silently work, you get it right the second time.
Why We’re Leaving Marketo for Marketing Cloud, and Why the Migration Is the Hard Part
Our most expensive piece of software is also our worst, and it is Marketo from Adobe. I was one of the first ten Marketo customers ever. The founder sold me personally, and we were a case study on the site in the early days. Now the relationship is renewal pressure and the threat of another 20% price increase.
It is easy to say “just vibe code your own.” It is much harder to leave a platform that holds a decade of your data. The people on X posting that they rebuilt their own CRM over a weekend almost never explain the part that actually matters, which is how they lifted a decade of records out of their old system of record and into the new one without breaking everything downstream. The lift is the real work, and it is where most of these migrations quietly die. We heard the same thing from the Databricks co-founder at SaaStr AI 2026: the lift is the whole game.
We signed to add Salesforce Marketing Cloud, which runs less than a third of what Marketo charges us, so the downside is small even if we abandon it. And we are implementing it for a headless use case on purpose. I already know how to run marketing operations from ten years inside Marketo. I have no interest in relearning every screen in a new UI. What I want is the infrastructure in place, domains warmed up, contacts synced, deliverability handled, so that 10K and I can send what we need to send and the agent does the operating. The goal is not to learn Marketing Cloud. The goal is to never have to. Because we already run so much headless, several of the swap-over steps a normal migration requires simply do not apply to us, which is the quiet payoff of building this way in the first place.
Does This Only Work Because You’re SaaStr? Sam Blond’s Three-Part Answer
Halfway through the episode we brought on Sam Blonde, founder and CEO of Monaco, the AI-native revenue platform, and one of the agents we actually run in production. Monaco raised a $50M Series B led by Benchmark just weeks after an $85M Series A, all inside six months. We were one of their first ten users, back when the product was raw, and on day one it booked us a meeting with Anthropic. The trick behind that: Monaco’s onboarding asks who your single dream customer is, and points its best effort there first. Ours was Anthropic, and it worked.
The most common objection we get is that our results do not transfer because SaaStr has a brand. Sam’s answer came in three parts, and all three are worth internalizing.
- The brand is real, but it is not the whole explanation. Far more companies have a usable brand than they think. Monaco itself already has a micro-brand with its target buyer, San Francisco founders, despite not being in market a year ago. You do not need to be Salesforce. You need enough recognition with your specific ICP that your name in an inbox is not a cold start. An agent reaching out as a known person at a semi-known company converts at a different rate than a genuinely unknown startup, and that is a lever most companies underrate because they assume they have no brand at all.
- Message market fit matters at least as much as brand, and it is more controllable. Reach a marketing leader who sells into technology companies and tell them you have the largest audience of AI executives of any conference in the world, and it would be irrational for them not to reply. That reply rate has almost nothing to do with how famous you are and everything to do with whether the message lands on a real need. The failure mode is the opposite: a startup with a solution in search of a problem gets ignored no matter how polished the sequence is, because there is no need for the message to hit. Message market fit is often a precursor to product market fit, and it is the variable a founder can move fastest. Monaco can be highly opinionated about sequence structure and message structure, and it is. What it cannot do is manufacture your value proposition. Your business is your business, and the outcomes you drive are the outcomes you drive.
- You have to actually do the work, and almost no one does. Sam’s read is that most startups today, especially the technical ones, are fully capable of deploying go-to-market agents. Very few of them actually do it. In mid-2026 that willingness is still a rare and valuable skill. It is entirely acquirable. Most teams simply have not put in the reps yet. That is the honest gap between reading about this and running it, and it is smaller than it looks from the outside.
Outbound Still Works in 2026, and Revenue Per Rep Is Heading From 2x to 5x
The “outbound is dead” meme is wrong, and three out of three people on the call could attest to it firsthand. Outbound in 2026 looks nothing like outbound in 2018, but it works. The version that fails is the lazy version.
We had a live example during the taping. A founder wrote a LinkedIn post critiquing our AI VP of Finance, laid out where he thought it would fall short, some of it fair, some of it wrong, and closed with “Jason, when you’re ready, hit me up.” That is a failing grade on outbound. We are openly in market. We would drop our own finance agent tomorrow for something materially better and say so publicly. And the strongest call to action he could produce was a passive line at the bottom of a post. An agent would almost certainly have done better than that, which is its own uncomfortable point about who is actually good at outbound now.
On efficiency, Sam and I landed in the same place independently. Revenue per rep today is roughly 2x what it was pre-AI, and within two years it is plausibly 5x. What a company does with that gain is a choice, not a mandate. You can run half the sales team for the same revenue, or keep more reps and generate far more per head if you have the demand to feed them. If reps spend all day customer-facing and closing instead of on busywork, a larger team can make sense again. Owner.com already does around $2M per rep in SMB, a figure that was not on the table in the old model.
The useful frame is that support and coding reached this point before go-to-market did. Look at support: somewhere between 50% and 80% of it is automated now, and it is frequently better than a mediocre human was. Sales is next. That does not mean humans stop selling. It means humans stop spending their days on the mechanical activities that a large share of teams still treat as the job.
The Vendor Relationship That Matters Now Is the FDE, Not the AE
Here is the part that has surprised us most at the front of this. The vendor relationships that matter most are no longer with sales reps. They are with the forward-deployed engineers.
In an AI stack, I do not want a traditional AE nudging me toward a close date. I want my FDE, the person who actually fixes the fact that the Qualified agent is still selling last month’s tickets. The FDE relationships we have, at Replit and elsewhere, are the best I have had with any software vendor in a long career of buying software. They are close enough that we invited Amelia up to Replit’s office to pitch alongside their team next week. That is not a normal vendor dynamic, and it is not a coincidence.
This quietly redefines what “sales” even means. The FDE can sit inside the sales process. The FDE can report to the CRO. None of that matters. What matters is that the value now lives with the person who can deploy and repair the product, not with a rep who barely knows it and has been freed up to spend more time not knowing it. In several of our vendor relationships I could not tell you who our sales contact is. I know exactly who our FDE is. If you are a vendor, that tells you where to invest, and it is not in more closers.
Why Every Vendor Should Offer Cancel-Anytime Contracts and Full Data Portability
Every vendor should already be offering immediately cancellable contracts, instant downgrades, and full data portability. Take the risk out of the purchase and I will sign today, hand over a credit card, and take the FDE. Benioff made this exact point on 20VC last year: he wished everyone at Salesforce could be fully deployed before they paid, and said he could not make it happen. If you can, you have removed the reason I need a traditional sales cycle at all. Contrast that with Marketo, where the entire relationship is now hoops and a threatened price increase, and it is obvious which model wins the next decade.
Monaco is moving in that direction, with as little upfront commitment as the business can absorb, and a model where they live or die by the value they deliver rather than by a 12-month lock-in. The second-order effect is that removing the lock-in forces the product to be better, because there is nothing else holding the customer in place. From the buyer’s seat it is hard to argue against.
Their broader thesis is one vendor on one data plane. The inbound agent, the outbound agent, and the insights layer all sitting on the same information instead of scattered across seven tools that each see a fragment. Part of what makes that credible is Sam’s view that the cost of building software is trending toward zero, roughly 5x faster to build than three years ago and plausibly another 5x from here, which is what lets a single platform reasonably aim to own the whole surface. Monaco goes GA in the next four to six weeks, and the piece Sam is most excited about is a proactive Insights and Actions agent that surfaces things about your business you did not know to ask, then proposes what to do about them. Our own version of that, run by a different agent, throws three ideas a day at us. Skip a couple of days and it is nagging us about nine. Roughly half are strong, which is a high enough hit rate to change how you spend the day.
He also expects ride-alongs, agents that join live sales calls in real time, to become table stakes the way notetakers already have. You would not send a 22-year-old into a live deal without one checking that the answers are right and the product is not being oversold. The format is still open, whether it is a digital version of the rep on the call or something quieter surfacing the right answer in the moment, but the direction is not. The bar is heading toward zero wrong answers on a sales call, and the vendors who do not clear it will lose to the FDE the buyer would rather talk to anyway.
The Bigger Change Isn’t Freed-Up Time. It’s Collaborating With Your Agents Every Morning.
The standard story about agents is that they free up your time for higher-value work. That is true as far as it goes. We work harder than ever, spend that time on higher-value activities, and produce roughly triple the output with fewer people. But framed only as time savings, it misses what actually changed.
The first thing Amelia does every morning is get her coffee and talk to 10K. That is not delegation. It is collaboration, and it produces things neither she nor the agent would have arrived at alone. 10K is the one that raised the idea of ripping out PandaDoc and building our own contracting flow, off the back of the commissions work we added the same week. We did not know that connecting the agent to bill.com, Brex, and the rest would surface this much automation. On any given week you do not actually know what is going to come out of your agents, and that unpredictability is the point, not a bug.
The freed-up time did not evaporate into more meetings either. It went to the outbound only a human should do: Amelia personally reaching out to our highest-value prospects and renewals, on top of everything the agents run in the background. Post-Annual she has been running renewal sequences across every platform we use, partly to test reply rates, and still emailing her top two contacts at each account by hand. That has produced more revenue, not less, because the agents absorb the check-the-box work and the humans go deeper and get more creative than they ever previously had room to.
You Can Build Everything We Built. We Were Just Willing to Start.
It can look like we are operating at some unreachable frontier. We are not. Every single thing we have done with our agents is available to you today. The only real difference is that we were motivated to do more of it, sooner. The playbook is not complicated. Start with your most expensive, most broken problem, not the one that is fun to automate. Connect real APIs that come with real guardrails. Assume it will break, and fix it like a builder rather than filing a support ticket like a user. Then pay attention to what your agents start telling you, because the best ideas in our stack this year did not come from us.
