Building a billing system is still hard, even with AI

The article argues that building billing systems remains fundamentally difficult despite advances in AI, because the core challenge is not coding but defining what a "correct" system looks like across complex, real-world business scenarios. The author explains that AI can help write code but cannot resolve the conceptual problems of handling edge cases like credit expirations, mid-cycle plan changes, and revenue recognition under accounting standards. Ultimately, billing is described as a consequence of every business decision, requiring deep systems thinking that AI cannot replace.

8 years building billing systems and I’m still wrong about just how hard it actually is. Not “getting better at it hard”, but like really genuinely discovering new layers of complexity I never saw coming. — Back when we started Paid at the end of 2024, I remember people telling me “billing is not hard, it may be tedious but it’s not that’s not that hard”. That person was wrong then I didn’t push them on this too much – and they’re still wrong today. Yeah yeah, sure, AI made the coding part easier but it didn’t make the software engineering easier. Coding the stuff was never the real bottleneck – not in the last 10 or so years anyway. The bottleneck is figuring out what needs to exist in the first place – what needs to be coded to make a consistent, useful system. Consistent and useful doing heavy lifting there. I’m still dealing with this today: Say a customer buys 1,000 credits in January. They use 800. What happens to the remaining 200? Seems straightforward until you layer in each company’s different models nad realities: - upgrade mid-cycle to a plan that includes 500 free credits - credits expire at different dates - a promotional credit pool that takes precedence - they request a refund but want to keep 50 credits - renewal happens but they’re switching from monthly to annual - have three subsidiaries each with their own credit pools that can’t intermingle - … the list goes on So which credits burned first? Which rolled over? Which expired? How do we do revrec under ASC 606? That’s not a coding problem. That’s a conceptual problem. You can’t Claude Code your way out of “What does correct even mean here?”. Not right now anyway. — The bummer is you also can’t validate these decisions until you’re deep in production with real customers, and you experience these very real edge cases you never imagined, that will cost you the contract with that customer. I’m three years in, Solvimon is just over 3 years in too. We have brilliant engineers. We have the AI, and we still run into cases where someone asks “what should happen here?” and it’s not hyper clear. Billing isn’t a feature or a checklist – it’s a consequence of every business decision you’ve ever made about how you charge customers. It’s the very interesting intersection of sales promises, legal, financial compliance, product functionality, and customer expectations both our customer and their customers in-turn For this, you need really excellent systems thinking. The problem didn’t get easier. If you’re building billing infrastructure – whether internal or as a product – and someone tells you “we’ll just use AI to build it, I’m not worried” – they’ve never actually built it. Respect the problem.