The Door Was Open. The Room Wasn't.
A few weeks ago I said accounting's AI wedge was wide open. Then a company walked through it and shut down. Here's what open doesn't buy you.
You can find a painful problem, build a real product for it, line up hundreds of companies that want to buy, and still run out of road before you reach a single one of them.
A company called FINNX just did exactly that. They built an AI finance employee named Edith and pitched her as “your first finance team member,” aimed at the exception-heavy last mile of accounting: pull the invoices, match them to purchase orders, chase the vendors, close the month, hand the genuinely messy cases to a human. They built her on deterministic, auditable logic instead of a loose LLM agent, on purpose, because a finance team has to know why a number reached the ledger. More than 700 companies joined the waitlist. The team was about five people. This month, after nearly two years, it shut down before it sold to any of them.
A few weeks ago I wrote that accounting just had its Copilot moment, that the bottom layer of the work was collapsing into AI and the wedge for whoever builds the new layer on top was wide open. I stand by every word. I filed taxes for three of my own businesses in fifteen minutes to prove the shift to myself, and I told you that if you understood a profession well enough, that domain proximity was your way in.
FINNX walked through that exact door. And an open door and a winnable room turn out to be different things. The distance between them is where most of these companies are going to die, and FINNX just paid to measure it.
I’m not here to autopsy a company I admire for trying. I’m here because I recognized something in how it ended, and it wasn’t really about accounting.
I Know This Clock
Here’s the thing every founder who’s built something hard will feel reading the FINNX story. To make Edith safe, the team had to connect to accounting systems, bank feeds, ERP tools, approval rules, and every customer’s own tolerances, because a vendor match that’s right for one company is wrong for another. And almost none of it produced value until most of it existed. One wrong journal entry isn’t a bug you patch next sprint, it’s audit risk on someone else’s books.
That’s a back-loaded value curve. You are worth nothing until you’re worth everything. The product only becomes trustworthy near the finish line, and runway runs out before the finish line almost every time.
I know that curve because I’ve stood on it. Every founder has had the quiet 2 a.m. version of the thought: we are doing real work, the demo lands, people want this, and none of it counts yet because the thing isn’t trustworthy enough to charge for. The founder of FINNX put it more honestly than most of us manage to: the company “ran out of runway before she was ready.” That’s not a confession of laziness. That’s a sequencing failure wearing the costume of an effort failure. The team did enormous work. They just arranged it so the payoff sat behind a wall they couldn’t reach in time.
The Bar Moves With the Blast Radius
There’s a second thing I’ve learned, less from building and more from advising businesses on what software to trust. The trust bar isn’t fixed. It scales with how much it costs to be wrong.
When AI writes code for me, I skim the diff and move on. The worst case is a failed test. When something posts to your ledger, a single wrong entry can surface in an audit eighteen months later, with your name on it. Same technology, completely different bar. The closer a product gets to owning a consequential process from end to end, the more it has to prove before it earns a single autonomous decision. That’s not FINNX’s bad luck. That’s the physics of the category. Anyone walking through that open accounting door is signing up for the highest trust bar in knowledge work, and most of the build happens before the customer lets you touch anything that matters.
What I’d Reach For Instead
So if I were building this, the question I’d hold isn’t “what’s the full process Edith should own.” It’s “what’s the thinnest slice that’s both trustworthy and valuable when it’s only twenty percent built.”
Maybe that’s invoice-to-PO matching, for one accounting system, shipped as a suggestion a human approves with one click. Boring. Narrow. But it’s in production in month three instead of month twenty, and it does two things a back-loaded build can’t. It earns trust you can compound, one task type at a time, the way you’d let a sharp intern graduate into autonomy. And it earns revenue that buys the runway to build the rest. Front-load a little value and you get to stay alive long enough to deserve the big version. That’s the whole game in a category where being wrong is expensive.
There’s also a tax I watched FINNX pay that nobody puts on a slide. The founder said explaining how their deterministic approach differed from a generic AI agent had become a full-time job. When you have to educate the market before you can sell to it, your runway is funding their tuition, not your traction. I’d rather let one boring, perfectly reliable feature do the arguing.
The Amendment
It’s easy to write all of this from the cheap seats. I didn’t make payroll for that team or sit in the diligence meeting where the math stopped working. FINNX found a real problem, made a defensible call, and got 700 companies to raise their hands, and still lost. That’s the part worth sitting with, because it’s the part my earlier article skipped.
I told you the door was open. That was true, and it was the easy half. The hard half is that an open door only rewards the people who can survive the walk across the room. If you’re building in a vertical that just cracked open, the wedge isn’t your real problem. Ask where your value lands on the build curve, because if you’re worth nothing until you’re worth everything, your real competitor was never the incumbent. It’s your own runway.
I write about AI, leverage, and building beyond code every week at macrostack.dev. If the distance between a convincing demo and software you’d trust with your books is something you sit with too, subscribe.