For most of the last decade, serious back-office automation was a capital expense, not a software decision. It required an engineering team, a year of integration work, and a budget that only an institution could justify. Those economics held firm for long enough that small companies stopped asking the question, and a generation of founders simply assumed the answer was no. Dylan Pulver, an engineering leader and founder who has spent the last ten years in financial technology (fintech) and automation, argues that the assumption is now out of date, and expensively so. “Ten years ago, if you wanted software that ran your back office without a person in the loop, you needed a bank’s budget, a team of engineers, and a year of integration work before anything moved,” he says. “But today, a 10-person company can run the same thing. I know because I build it.”
The Banks Were Never Buying Intelligence
The most useful thing Pulver does is correct a widespread misreading of what enterprise automation costs. Leaders tend to assume the price tag reflected sophistication, some proprietary cleverness available only at scale. It did not. “What the banks were paying for was never intelligence. It was plumbing,” he says. “Every system had to talk to every other system, and someone had to build that connection and fix it when a vendor changed a login screen. That’s where the money went.” The spend went into connective tissue and the permanent maintenance burden that connective tissue creates. A vendor redesigns an interface and the pipe breaks. Somebody gets paid to fix it. That cycle, repeated across dozens of systems, is where the budget disappeared.
Small companies never opted out of that cost. They paid it differently. “Small companies couldn’t afford that plumbing, so people became the plumbing,” Pulver says. “The export on Monday, the paste into the other system, the invoice checked by hand because two totals never matched up.” That line deserves to sit with any executive who believes their firm has avoided integration spend. It has not. The expense was simply moved off the software line and onto headcount, where it is harder to see, never depreciates, and scales linearly with volume. A manual reconciliation performed every Monday is an integration project financed in salary rather than capital, and unlike software, it gets more expensive each year.
What An Agent Does That A Pipe Cannot
The structural shift Pulver identifies is that an AI agent does not need the pipe at all. It operates the system the way a person does, through the same screens and the same logins, which removes the entire integration layer from the cost equation. “An AI agent can log into a system and work it the way a person would. No integration project needed, no year of waiting,” he says. The consequence is that the economics invert. Automation stops being a project with a long payback horizon and becomes something closer to a weekly operating decision. The barrier was never the intelligence. It was the twelve months of plumbing that had to be laid before anything could move.
The timeline compression is the part worth planning around. Pulver has watched the effect directly: “I’ve watched a process that took a week of someone’s month drop to under an hour of review, and it was built in days, not quarters. That’s the bank’s automation at a small company price.” Days rather than quarters changes how a leadership team should think about risk. When a build takes a year, every automation decision carries the weight of a bet, which is why they get escalated, debated, and deferred. When it takes days, the cost of being wrong collapses, and the rational move is to try something, measure it, and move on. Firms that keep applying year-long governance to a days-long build will lose ground to competitors who simply start.
Write The Job Down And Count The Hours
For a small company, getting started doesn’t require specialist technical skills or a data science team. As Pulver puts it: “What it takes is not a data science team, it’s a written description of the job. Write it down the way you would explain it to a new hire, including the exceptions, because that document is the automation.” The exceptions clause carries most of the weight. Every back-office process has a documented version and a real version, and the gap between them lives in the heads of whoever has done the work longest. The edge cases, the vendor that’s known to cause problems, the invoice that always arrives in the wrong format. Teams that skip that part produce automation that works until the first irregular Tuesday.
That framing also exposes why most firms stall. Writing the job down feels ordinary and tedious, yet it reveals how much institutional knowledge was never captured anywhere, which is precisely why it is the real work. Pulver’s advice on where to begin is deliberately unambitious. “Then start with the most boring hour on the list and count. Hours in, hours out.” Not the highest-value process, not the one with executive visibility. The dullest one, with a number attached, so the result is arguable rather than anecdotal. That reframes the entire conversation for a leadership team. “So the question for a small company is no longer whether you can afford automation,” he says. “Instead, it’s which hours you want back first.” Affordability was the old constraint. Attention is the new one, and attention is rationed by whoever decides to sit down and write the first document.
Follow Dylan Pulver on LinkedIn and at dylanpulver.com for more insights on AI agents, back-office automation, and fintech engineering.