Defensible SaaS business ideas in the AI era

“A lot of existing SaaS companies and new startups being vibe-coded should not exist.” — Andrej Karpathy

Middleware is no longer a viable business. I think new founders are committing mistakes in building thin middleware over easy “formatting” tasks that modern AI agents are able to commoditize and perform directly with a single prompt in ChatGPT or Claude.

Many new startup ideas I see tend to be a thin wrapper around an LLM. These are weak.

Weak product:

input data → AI → formatted output

Stronger product:

input data from workflow → human feedback → verified (graded) outcome → better workflow system

Durable products in the post-AI era usually own at least one of:

1. proprietary data in an internal system of record
2. user feedback loops (a sort of data)
3. outcome verification (akin to Lean or compilers)
4. distribution / traction
5. trust or proof in regulated and slow-moving industries

That said, most startups tend to not have list items (4) and (5) above. Early stage startups are unlikely to possess any of the above. The question is how do you use your early users and actual user-experience to automatically generate (1) - (3), such that your system is a virtuous flywheel that gets “better” in proportion to its userbase in a way that a new competitor can’t copy, when the cost of copying has reduced to all time lows due to LLMs and agentic AI?

My opinion: I think you still need to own a database with proprietary data. Every user interaction should build upon that internal database, allowing you improve your offering’s value over time. Importantly, while your users and visitors (e.g. AI crawlers) product can access your page, they only see shallow “views” of the database rather than the whole entirety, so as to make distilling the essence of your product or reverse engineering it more difficult. If possible, your product should be in highly regulated or slow moving fields (e.g. healthcare, finance).

In summary, strong products in the post-AI era are feedback engines that compound through human judgment, proprietary context, and verified outcomes.