The Agentic Pyramid: Where the Hell Do You Actually Sit?

I pitched visualAI as four AI product-discovery apps: shopperGPT, studioGPT, cleanerGPT, catalogGPT — polished, 1-click tools merchants and agencies could deploy in minutes. Investors nodded, asked sharp questions, and passed. “App company.” “Wrapper.” “What happens when Shopify or OpenAI ships this natively?”

They weren’t wrong to ask. They were reacting to the story I was telling, not the company I’d actually built. Underneath those four apps sits discoverGPT — a REST API and 40 MCP tools, years of engineering, abstracting product discovery, virtual try-on, catalog quality, and AI-visibility scoring for anyone who needs agentic discovery in their stack. The four apps aren’t the business. They’re the fast, friendly front door to it.

I’d been describing the door and wondering why nobody wanted to buy the house.

That mistake sent me looking for a map — something that could place every kind of company in this industry, including my own, so “app,” “wrapper,” “platform,” and “infrastructure” stopped being marketing words and started being structural ones. I couldn’t find one that held up. So I built it.

The Agentic Pyramid

Picture it top to bottom: narrow and expensive at the top, wide and disposable at the bottom. Defensibility increases as you go up. Volume and interchangeability increase as you go down. Every AI company sits somewhere on this, whether it knows it or not.

1. Foundation Models — OpenAI, Anthropic, Google, xAI. The reasoning substrate everyone else calls by API. Capital-intensive, structurally scarce.
2. Hyperscale Infrastructure — AWS, Google Cloud, Azure, Nvidia. The compute everything above runs on. Hugging Face sits just below, distributing open models to anyone who can’t train their own.
3. Protocols — the layer most people miss, and the one moving fastest. Anthropic’s MCP, Google’s A2A, and in commerce specifically, OpenAI/Stripe’s Agentic Commerce Protocol and Google’s AP2 and Universal Commerce Protocol. UCP has effectively won that fight this year: Amazon, Meta, Microsoft, Salesforce, and Stripe all joined its Tech Council — as members, not competitors. Nobody wanted to own the protocol war. Everybody wanted to be the rails underneath whoever did.
4. Agent Orchestration — LangChain, Salesforce Agentforce, Lovable, and — for engineering specifically — GitHub Copilot and Cursor. Scaffolding for building agents. Some of this layer’s biggest output is more noise for layer 7: Lovable, for instance, is a factory for exactly the kind of thin app that gets built and discarded fastest.
5. Domain Infrastructure — the *-as-a-service layer. Stripe, Visa, Mastercard, and Adyen for payments. Twilio for messaging. discoverGPT for retail discovery. Deep, hard-to-replicate backend capability built once and exposed to whoever needs it — your own app, someone else’s SaaS product, or an agent that doesn’t exist yet.
6. SaaS / Applications — Shopify, Salesforce, Adobe, Clay. The software people actually log into. shopperGPT, studioGPT, cleanerGPT, and catalogGPT live here too, as fast-deployment shells over discoverGPT — the same way Stripe’s Checkout and Billing sit over Stripe’s raw API.
7. Agentic Wrappers, Bots, Point Agents — the base. Thin, prompt-based, easily cloned. Thousands of nearly identical products calling the same model API with a different UI. This is where “AI company” perception goes to get commoditized — and it’s exactly what killed my pitch every time someone glanced at four apps and stopped looking.

Three things this map makes obvious

A handful of companies — Google chief among them — occupy almost every layer at once: model, cloud, protocol, orchestration, and app. That’s not a move available to the rest of us. If you’re not a hyperscaler, the winning strategy is to pick one layer, go deep enough that a giant won’t bother commoditizing you, and defend it hard. Trying to look like you play at every layer is how you end up looking like you play at none.

The industry has already encoded this pyramid into its own plumbing, whether it says so or not. Read Stripe’s or Google’s documentation for ACP and UCP and you’ll find an explicit split between an “agent/app layer” — what the buyer sees — and a “merchant/business infrastructure layer” — the backend a business builds once and distributes to any agent. That’s the SaaS/DaaS line. I didn’t invent it for a pitch deck. It’s the actual architecture the protocols assume.

And one more thing worth saying plainly: how much money a company has raised is not the same axis as where it sits on this pyramid. This year has already produced eye-popping valuations for “zero-employee, agents-run-everything” companies sitting at the very base — layer 7, all orchestration, no proprietary backend. That’s a bet on narrative and velocity, and it can absolutely pay off. It’s just a different bet than the one you’re making building layer 5 infrastructure, and it’s worth knowing which one you’re placing before you tell your story to an investor, a partner, or yourself.

The lesson

If you’re building real infrastructure — a data asset, a model, a pipeline that took years and real capital — and the market keeps calling you an app or a wrapper, don’t just blame the market. Check whether your own story is pointing at the door instead of the house. The apps you ship for speed of adoption aren’t a confession that you’re an app company. They’re a distribution strategy, same as Stripe’s Checkout button. But you have to say so, explicitly, before someone else decides what layer you’re in for you.

I wish I’d drawn this pyramid sooner. I’m publishing it now so the next founder, investor, or partner sizing up a company in this space has a map instead of a vibe.

Where does your own company actually sit?