AI Is Cannibalising SaaS. I Watched It Happen Three Times in One Afternoon.

I sat down recently to figure out whether an old project of mine was worth turning into a product. A RAG ingestion tool I’d built a while back. By the end of the afternoon I’d talked myself out of three different versions of it.

~

The Ideas

I’d built a RAG agent for a documentation site a while back. Usage had fallen off noticeably. I asked around and the answer kept coming back the same way. People are just asking general AI models now.

Models know public docs well enough these days that a dedicated agent isn’t necessary for most questions anymore.

So, a pivot. Focus on internal docs instead? Company wikis, private knowledge, nothing a public model has seen etc.

Platforms like Microsoft 365 Copilot already do that though -and natively, for anyone with the licence.  It can be grounded in a company’s own systems with minimum deployment effort.

So, pivot again. Forget end users, sell it to developers as an accelerator?  A shortcut so they don’t have to wire up RAG tooling themselves.

Except AI tools can generate that kind of boilerplate on demand now too. Agent frameworks are shipping RAG primitives natively. The accelerator I wanted to sell is becoming a matter of plumbing some tools and SDKs together.

Three ideas in one afternoon, all blown apart.

~

Why This Is Happening

Models have most of the knowledge.

Frameworks and SDKs have the ingestions, retrieval and orchestration built in.

Platforms like Copilot have the integrations, governance, and enterprise trust.

You used to buy separate SaaS tools to plug the gaps between these. That’s changing quickly.

The moats are disappearing.

A lot of SaaS has been selling “I connect this to that” or “I know this thing.” This is what’s getting absorbed fastest.  For example, Jev was just released, a few weeks later, OpenAI rolled out the Decisions API which offered the same capabilities.

~

What Held Up

Three things survived that afternoon.

  1. Data that’s actually private. Customer records, internal decisions, anything that was never public and was never going to be scraped. The model can’t know what it’s never seen.
  2. Governance and compliance. Procurement and audit move slower than capability does. A regulated company isn’t pointing a general tool at legal records just because it’s technically able to handle them. That gap is real, even if it’s shrinking.
  3. The thing that held up best wasn’t a tool, it was teaching and explaining the thing directly. People weren’t paying for RAG they couldn’t get elsewhere. They were paying for someone they already trust explaining it to them.

A model can get better at knowing things. It can’t make someone trust it the way they trust a person they already have a relationship with.

~

What to Consider Instead

Some questions to consider, is <the thing> a capability gap a model can quickly plug, or, a relationship and community that you can add value to?  Would they miss you if you stopped showing up?

Capability gaps close. A model gets better and the gap’s gone. Relationships don’t close the same way and someone trusting you isn’t a feature a model can ship.

~

~

JOIN MY EXCLUSIVE EMAIL LIST
Get the latest content and code from the blog posts!
I respect your privacy. No spam. Ever.

Similar Posts

Leave a Reply