Software Architecture Insights

The matrix works unchanged. The definitions underneath it don’t. Row 42 of your risk matrix warns against the AI assistant issuing a refund against the wrong order. Nobody has ever seen it happen. The capability is real, it sits under a dollar threshold, and the team scored it Low Likelihood without much debate. Then they got to severity and the room went quiet, because the money is gone, the customer hears about it from their bank, and there is no undo. Low Likelihood and High Severity....

AI-Native Architecture, part four of four: the component that changes on someone else’s schedule. There was no deploy. Nobody merged anything. The last change to that service went out five weeks ago, and it was a logging tweak. And yet classification accuracy on the intake queue has been sliding for nine days. This morning somebody noticed the assistant has started answering a whole category of question in a different tone. Shorter. More hedged. Technically fine, and different enough that a...

AI-Native Architecture, part three of four: your model is only as good as what you feed it. A customer asks the support assistant whether they can return an opened item. The assistant says yes. Within sixty days, no receipt needed. Clear, polite, correctly formatted, and completely wrong. The company changed that policy in June. A content editor updated the help article, which is exactly what they were supposed to do. Nobody told them an AI system reads that article. Nobody told them because...

AI-Native Architecture, part two of four: every inference has a price and a latency. The feature demoed beautifully back in March. A support assistant that reads the customer history, checks the order, and drafts a reply. Everyone in the room agreed it should ship. It shipped in June. By August somebody in finance is asking why one line on the cloud bill is growing faster than the company is. Nobody did anything wrong. The feature works exactly as designed. That’s exactly the problem. The...

AI-Native Architecture, part one of four: correct is now a distribution. The regression suite has been green for six weeks. Every build passes. The dashboard has settled into the kind of calm that makes people stop looking at it. Support has been collecting tickets that whole time. The summarizer keeps getting things wrong. Nothing garbled, nothing that looks like a crash. Just answers that are fluent, specific, and false, in ways somebody only caught by reading the source document. Those two...

Criticality is one decision. Ownership is the other. Most organizations have made neither. It’s 3:07 in the morning and a phone is going off. The alert says a payment-related service is throwing errors for about a third of its requests. The engineer who acknowledges the page has never touched that service. They open the runbook, which was last edited by someone who left the company in March. They page the team they believe owns it. That team replies that it owns the client library, not the...

A plain language breakdown of what the regulation means for software teams building or deploying AI, even if you aren't in Europe. Your product manager has just heard about the EU AI Act. They send you a message: "We use AI in our product. Do we need to comply with this?" Maybe your legal team forwards a three page summary they don't fully understand. Maybe a customer in Germany asks whether your system is "EU AI Act compliant." Here's the honest answer: it depends on what your AI system...

The role of evolvability in modern application development. Your team adopted AI coding assistants four months ago. Output is up, and everyone is happy about it. Then last week someone shipped a change to the billing service that quietly broke an assumption three other services were making, and nobody caught it for nine days. Nothing about that change was careless. The system simply moved faster than anyone’s ability to remember what it was holding together. For most of the last decade, when...

Three terms that get used interchangeably, and that's a problem for practitioners. The VP of Engineering sends an email with the subject line "AI safety concerns." What exactly does that mean? The engineering team thinks it's about hallucination rates and failure modes. Legal thinks it's about regulatory exposure risks. The product ethics lead is certain it's about the recent bias tracking metrics. They all show up to the same meeting. All with different agendas. They all are wrong. Everyone...