Over the last year, I kept hearing the same phrase in AI conversations:

Sometimes it came from investors. Sometimes from operators. Sometimes from founders themselves.

A startup would demo something impressive. People would get excited for a few minutes. Then someone would eventually say: “Yeah… but OpenAI or Anthropic will probably ship this.” Conversation over.

And honestly, sometimes that assessment was correct. Some products really were thin surfaces sitting directly on top of foundation models with very little defensibility underneath.

But what bothered me was that the conversation itself felt shallow. Because occasionally I would look deeper at one of those companies and realize — they actually did have an edge. Not always an obvious one. Sometimes they did not even understand it themselves yet.

Some had workflow gravity. Some had access advantages. Some had embedded distribution. Some had proprietary behavioral data. Some had orchestration hidden underneath the UI. Some had memory accumulating quietly beneath the surface.

The market had vocabulary for product-market fit. But it did not yet have good vocabulary for structural position in the AI era.

Why a supply chain, of all things

At some point, I stopped thinking about AI as “just software” and started thinking about it more like a supply chain. Oddly enough, the mental model that clarified it for me was gold.

Before somebody wears a gold ring, there is an entire chain underneath it: mining, refining, transport, verification, crafting, distribution, retail, and eventually memory about the customer itself. The visible experience is only the final layer.

That idea stayed in my head for months and eventually became part of why I called this framework The Supply Chain of Intelligence™. I'll write separately about the full analogy — it ended up being one of the clearest ways to explain how AI value actually moves through the stack.

The companies that compressed — and the ones that adapted

I kept watching companies respond very differently to the rise of foundation models. Some compressed almost overnight. Others adapted surprisingly well.

Apollo was one example that made me think deeply. At one point, Apollo had a broad set of workflow features: prospecting, messaging, outbound, CRM-like behaviors. But instead of trying to fight the model companies head-on, they increasingly leaned into becoming a trusted data and access layer for the AI ecosystem itself.

The Claude partnership direction was especially interesting. Rather than forcing users into a giant standalone interface, Apollo became useful as structured business intelligence directly inside the AI workflow. They were not trying to out-model the model companies. They were positioning themselves where the models still needed them — where trust, permissions, freshness, enterprise relationships, and proprietary business data still mattered. That was not weakness. That was structural positioning.

Then I looked at companies like Sierra. Everyone called them “agent companies.” But underneath the branding, they were clearly building orchestration, workflow control, enterprise integrations, runtime systems, access layers. The value was not just the conversational surface. The value was increasingly underneath it.

Years earlier, Jasper exploded because the AI surface layer suddenly became valuable. Then ChatGPT arrived and compressed huge parts of that layer almost overnight. At the same time, Grammarly survived far better than many people expected. Why? Because Grammarly was never only a writing prompt wrapper. It already had integrations, embedded workflows, cross-surface presence, habitual usage, plugins, accumulated behavioral context, and distribution embedded deeply into the writing ecosystem itself.

Most AI discussions are at the wrong layer

The more I observed these patterns, the more I realized something important: most AI discussions were happening at the wrong layer of the stack.

People were talking about prompts, models, copilots, agents, interfaces. But the real strategic questions were deeper:

Which layer do you actually own? Which layer can compress you? Which layer compounds over time? Which layer is rented from somebody else? Which layer survives when the foundation model companies move upward?

That eventually became the foundation for what I now call The Supply Chain of Intelligence™ — a framework that maps where AI value is created, captured, compressed, defended, and accumulated across the stack. 10 layers. 50 sublayers. 4 structural laws.

The four structural laws

Law I — Intelligence Commoditizes Downward. Wrappers become features. Anything that exists only at the surface layer eventually gets compressed by the model layer beneath it. But that does not mean applications disappear — it means structurally thin applications disappear.

Law II — Value Accrues at Bottlenecks. Durable value forms where scarcity exists, not where hype exists. Increasingly, the strongest moats sit around proprietary data, workflow ownership, trust, access, orchestration, compliance, and memory.

Law III — Surface Captures Attention; Chain Captures Power. The AI industry massively over-focuses on visible intelligence — interfaces, generation quality, conversation UX, demos. The strongest companies often own deeper layers underneath: integrations, workflow systems, operational embedding, runtime orchestration, accumulated memory, behavioral context. The visible layer gets attention. The deeper chain retains leverage.

Law IV — Memory Is the Final Moat. Most AI systems optimize for generation. But over time, defensibility increasingly comes from accumulation — what the system remembers about the user, the workflow, the organization, and the operating context. That changes how AI product leaders should think entirely.

That question changes roadmaps.

Decoding the vague words

The framework also helped me realize how vague a lot of AI language had become. AI-native. Agentic. Copilot. Assistant. Wrapper. Those words often hide more than they explain.

Take the word “agent.” Most people talk about agents as if they are a category. Structurally, they are usually packaging across multiple layers: execution, orchestration, surface, and sometimes memory.

Without execution, it is often just a chatbot. Without orchestration, it is a workflow script. Without memory, it is frequently a demo instead of a system. The framework forces a more structural conversation.

Same patterns across every vertical

One of the most useful parts of building this has been applying it across industries: healthcare, finance, enterprise SaaS, legal, education, developer tools, infrastructure, vertical AI. Different verticals. Same structural patterns.

The framework consistently exposes compression risk, dependency layers, hidden bottlenecks, moat locations, migration paths, and structural weaknesses. More importantly, it creates a common language — because vague language creates weak strategy.

If teams only say “AI-native” or “agentic” without understanding the underlying layers, they miss the harder strategic questions: Which layers do we actually own? Which layers are rented? What happens if the model layer ships this for free? Which parts compound? Which parts decay?

That is why I built The Supply Chain of Intelligence™. Not as another AI buzzword framework — but as an attempt to create better structural language for how AI businesses actually evolve.

My hope is that over time it becomes useful the same way JTBD became useful: not as something people admire, but as a language founders, PMs, and investors naturally think in.

Because the next generation of AI winners probably will not be determined only by who generates intelligence best. It will be determined by who structurally owns the deepest parts of the chain.

The full framework — 10 layers, 50 sublayers, the four laws, the case studies, the live market map, and the downloadable posters — is free at supplychainofai.com. No signup. Take what is useful. Cite it where it helps.