In the agent era: are you the orchestrator, or the thing being orchestrated?

For decades, software organized itself into systems of record (where data lives), engagement (the interface you work in), and action (getting things done). You switched between apps — CRM here, email there, calendar over there — and the interface was where the value and the customer relationship lived. AI agents scramble that. An agent can become a single entry point that you talk to, which then reaches into all those systems behind the scenes. The engagement layer — the interface — is suddenly at risk of being stripped away and unified by something new.

That opens the highest-stakes contest of the AI era: the battle for the workflow entry point, with two extremes. At one end, the orchestrator — the agent the user talks to directly, holding the intent and the routing decisions, capturing the customer relationship. At the other, the backend pipe — software called invisibly behind an agent, losing its interface, brand, and pricing power, at risk of commoditizing to compete only on price. These are two ends of a spectrum, not a clean binary: in reality there are many middle positions — domain agents, embedded agents, and high-value layers that don’t own the end interface yet still control critical data, compliance, or execution rights.

What decides where a company lands? The same three moat sources this whole book has hammered, now in AI’s context. Data — whoever holds the most proprietary, hardest-to-obtain data (the system of record) holds the source of an agent’s intelligence, and stays unbypassable even if the interface is taken. Distribution — whoever is already the entry point users open every day is best placed to become the unified agent (Microsoft putting Copilot into Office and Windows is exactly this). Trust — when an agent must act autonomously on your behalf (send email, spend money, change data), trust becomes the gate, and incumbents who’ve earned deep enterprise trust have a natural edge.

Notice that none of the three deciders — data, distribution, trust — is a feature you can ship in a sprint. They’re accumulated over years, which is why the entry-point battle so often favors whoever already holds them, and why an AI-native newcomer with a brilliant agent but no proprietary data, no installed distribution, and no earned trust is more fragile than its demos suggest. The interface can be rebuilt overnight; the three walls cannot.

So the analytical maxim for AI-era SaaS isn’t “how good are its features?” It’s positional: in a world of agents, does this company have the ammunition — data, distribution, trust — to be the orchestrator or at least an unbypassable data source? Or is its value dangerously deposited in an interface that an agent can strip away? The entry-point battle is, at bottom, a battle of moats — and the shallow-moat, interface-heavy tool is the one that gets pipelined first and most thoroughly.

This is Chapter 18 of “Compounding in the Cloud.”