We recently filed patents on two of our products, ERPClaw and TailTest. I want to write about that honestly, because most companies either say nothing about their patents or wave them around like trophies, and both are a little dishonest. A patent is not a trophy. It is a bet, with lawyers, and the only interesting question is why you chose to make it on one thing and not another.
A small studio has no business filing patents on reflex. They are slow and expensive, and for most software the protection is thinner than founders imagine. So our default is to not bother. The two times we did bother are therefore the interesting cases, because they tell you what we actually think is novel rather than merely good.
The filter: novel, not just nice
Plenty of what we build is good. Very little of it is novel in the sense that matters for a patent, which is a specific mechanism that is genuinely non-obvious and that we would mind a competitor copying wholesale. Most software is a smart arrangement of known parts, and there is no shame in that. The filter I used was simple. Would I be annoyed in a deep, structural way if a well-funded competitor lifted this exact idea? For almost everything, the honest answer is no, because the moat is execution and operation, not the idea. Twice, the answer was yes.
ERPClaw: making an ERP answer back
A traditional ERP is a vast, rigid system that makes you learn its language. ERPClaw inverts that. You ask it a question the way you would ask a competent colleague, and it does the unglamorous work of translating your intent into the rigid world underneath, then translating the answer back. The novel part is not the chatbot. Chatbots are a commodity. The part worth protecting is the specific mechanism that maps loose human intent onto a strict enterprise data model reliably enough to trust with real operations. That mapping is hard, it is non-obvious, and it is the thing I would mind seeing copied.
We did not patent the idea of an idea. We patented the specific, awkward mechanism that makes the idea survive contact with a real system.
TailTest: testing that thinks like a skeptic
TailTest came out of an old frustration. Testing tools check the behavior you remembered to specify, which means they are blind to exactly the failures you did not think of, which are the ones that hurt. The mechanism we filed on is about generating and prioritizing tests the way a suspicious senior engineer would, going after the cases a tired human would skip. As AI writes more of our code, this matters more, not less. When a model produces code at speed, the scarce thing is judgment about what could go wrong, and that is precisely what we tried to bottle.
What a patent actually buys a startup
Honestly, less than founders hope and more than cynics claim. It does not stop a determined competitor; it raises their cost and slows them down. It does not build a business; a patent on a product nobody wants is an expensive certificate. What it does give a small company is a small, specific edge in a world tilted toward whoever has more capital. We did not file to look impressive. We filed on two mechanisms we run in our own production every day, which is the only kind of innovation I trust, because we build and run the AI we recommend.
If there is a lesson, it is that the patentable thing and the impressive thing are rarely the same thing. The impressive thing is usually the whole product. The patentable thing is one stubborn, specific mechanism buried inside it, the part you bled over and would hate to see lifted. Find that, and you know what you actually invented.