Why I Treat AI as Infrastructure, Not a Feature

<p>There's a pattern I keep noticing when people talk about "adding AI" to a product: they treat it like a feature. A chatbot in the corner. A summarize button. Something you can point to in a demo and say "look, AI." I think that's the wrong way to think about it, and it's a mistake I'm actively trying not to make across the projects I run — AIREP, Find a Sign, and Nova itself.</p><p>My view is that AI works best as infrastructure, not decoration. A database isn't a feature of your app — it's the thing everything else is built on top of. I think AI capability needs to sit at that same layer. Not "here's an AI feature," but "this system reasons, classifies, and automates as a baseline, and the AI parts are invisible because they're just how the system works."</p><p>Take an ERP system like AIREP. The obvious, lazy move is to bolt a chat assistant onto the dashboard so users can "ask questions about their data." That's a feature. It's fine, but it's not differentiation — every SaaS product has one of those now, and most of them are mediocre wrappers around a general-purpose model with some context stuffed in. The more interesting question is whether the system itself gets smarter about branch-scoped data patterns, anomaly detection, document processing, workflow suggestions — things that happen whether or not a user ever opens a chat window. That's AI as a competency, not a checkbox.</p><p>Same logic applies to a marketplace like Find a Sign. The tempting AI feature is some kind of "smart search" gimmick. But the actual leverage point, in my opinion, is using AI to keep the discovery process honest — surfacing suppliers based on fit and quality signals rather than who paid for placement. That's not a feature you screenshot. It's a design principle that happens to be implemented with AI underneath. I've been fairly vocal about disliking pay-to-rank models, and I think that distrust extends naturally into how I think about AI: if the AI layer is just another lever for manipulation — nudging results toward whoever integrates best with your sales funnel — it's not actually serving the customer. The infrastructure framing keeps that honest, because infrastructure doesn't play favorites; it just works the same way for everyone underneath the product.</p><p>Nova is the clearest version of this for me personally. It's not a product with an AI feature bolted on — it's a multi-agent system I use as the backend service layer for things like my own site. The agents aren't there to impress anyone in a demo. They're there to do orchestration, specialist tasks, language work, pipeline execution — the unglamorous stuff that makes a system actually useful day to day. When it's working well, you don't notice the AI part at all. You notice that things got done.</p><p>I think this distinction matters more than it sounds like it should, because it changes what you build first. If AI is a feature, you build the demo-able bit — the chat window, the summary button — and worry about the plumbing later. If AI is infrastructure, you build the plumbing first: the orchestration, the data access patterns, the agent coordination, the boring reliability work. The flashy stuff comes later and is often optional.</p><p>It's slower to build this way, and it's much less photogenic. Nobody screenshots an orchestration layer. But I'd rather spend the next stretch of work making the infrastructure solid — in AIREP, in Nova, in how Find a Sign surfaces suppliers — than ship something that looks like AI and behaves like a toy. The goal isn't to have AI. It's to have systems that are quietly better because of it.</p>