The 3-Layer AI Stack Killing Traditional SaaS, and Where the Moat Is

"SaaS is dead" is overblown, but the threat to the business model is real, and the sell-off in SaaS stocks says the market sees it too. Strip a SaaS product down and it is a UI plus a system of record: you operate the interface to manipulate the database. Everything defensible is moving into a third layer that sits between them, and inside that layer only one thing holds up long term: proprietary data that nobody can replicate without licensing it from you.

Credit where due: the starting point is "Context is King" from Evan at The Leverage, a newsletter I genuinely like and am not affiliated with. His argument: a new context layer, institutional knowledge and process logic, is emerging between the UI and the database. That work used to be done by people. The sales rep loaded up the CRM before the call and carried the context in their head; the SaaS existed to store it. Now the layer itself becomes software.

From context layer to value layer

I extended the idea, because context is only one piece of that middle layer, and "context" already means something narrow in AI. I call the whole thing the value layer. I do not love the name either; I could not come up with anything better.

The stack then reads: point-solution applications on top, which might be a classic SaaS UI, a generated-on-the-fly UI, a chat interface, or an agent like Claude Code. The system of record at the bottom, unchanged. Models themselves trending toward commodities, like electricity, with the honest caveat that they are utilities built on other utilities, and all of it has to keep working out for that to hold.

The three tiers of the value layer

Context: data and knowledge. Everything that gets fed into the model: skills as markdown files or Python scripts, MCPs, proprietary datasets, in-house domain expertise, external knowledge, RAG if you like. This is the highest-defensibility tier.

Prompting: instructions and behavior. System prompts, configurations, prompt libraries, skill instructions. Your internal process abstracted into a workflow file: this is how we do outreach, this is how we qualify. Domain-specific by construction, and medium defensible: it encodes an opinion, and opinions can be imitated.

Programmatic: event-driven logic. Hooks, dynamic context injection, subagents, system reminders, conditional routing. The plumbing that makes an LLM remind itself to do things and produce better deliverables. Least defensible, because it is visible engineering anyone can rebuild.

The strategic implication: do not compete with Claude Code or Codex on building the best harness. They will almost always have the best harness unless your use case is genuinely peculiar. Extend their harness with the layers above instead, and put your effort where the defensibility gradient points.

Why proprietary data is the only real moat

The question that decides whether you have a business: what data do you have that only you have, or that is expensive enough to obtain that OpenAI or Anthropic will not bother replicating it, and would rather license it from you?

You can build that dataset two ways. From existing operations: a service business already runs a process and sits on the outcome data from every engagement. Or deliberately, the way OpenAI scraped theirs into existence, which is less defensible but still adds friction against copycats. Defensibility comes in layers.

Then close the loop: once the dataset serves users, product usage should refine it, or spin up an adjacent dataset with synergies. That flywheel is what compounds. Wiring the data to the LLM is the trivial part; a skill or an MCP takes five minutes to stand up. The data underneath is the asset.

The value layer is portable

The property that makes this worth building: the value layer is an LLM extension. Any UI, agent or product that can use your LLM benefits from it. The interfaces on top can churn every six months, chat today, generated UI tomorrow, and the layer keeps its value. Building a narrow SaaS might still work for the next half year to a year; I do not think it is defensible beyond that.

This thesis is the foundation of the Value Layer work we do for clients, where the case studies show what these layers look like built out in practice. And if you think I am talking out of my ass, tell me; this is how I currently think about it, not a law of nature.