Was having a discussion of SaaSocalypse today and I was arguing that there is more domain knowledge data embodied in the design of legacy software than people realize, which broad general purpose AIs will miss unless specifically trained on the domain tacit knowledge. Joel Spolsky argued this point pre-AI:

In a way the implicit thesis behind the SaaS crash is a version of Brooks’ “plan to throw one away” at industry-generation scale. As though entire SaaS industry were just version 1 of “CivSoft.” As Spolsky argues, you probably shouldn’t do this. Brooks was kinda wrong.

You’ll vibecode the 20% that does 80% of the job and then realize all the cost at scale is in the final 20% which takes 80% of the effort. Which is what the incumbent will have in its offering. Accommodation for the “reality has a surprising amount of detail” problem.

It’s not a permanent advantage but it means the disruption will take longer than people think.

“Can the incumbent figure out AI before the AI-native disruptor learns all the corner cases?”

The real play might be to offer the SaaS companies ability to train models on current infrastructure proprietary source code and usage data in order to write the next version with AI. A kind of upgrade industry almost like y2k industry. Upgrade from software 1.0 to software 2.0. And probably reimplement in Rust while you’re at it, work in the refactors people have been itching to do, pay off all the technical debt etc.

Unless there’s a fundamentally new idea for the category waiting in the wings, just rewriting an improved version with AI is probably something the incumbent can do best, not AI-native challengers.