There is a divide opening up between the companies that are meaningfully compounding their AI advantage and the companies that are running impressive-looking pilots. I have come to think the divide is captured by a single frame — and once you see it, the pattern is difficult to unsee.
The winners are treating AI as a plumbing layer. The strugglers are treating it as a product layer.
The distinction matters, and it is worth spelling out.
What “AI as a product layer” looks like
The product-layer approach positions AI as something the customer experiences. Chat interfaces. Copilots. Content generation surfaces. Assistants that recommend, summarize, or reply. There is nothing structurally wrong with these — some of them will become large, durable businesses — but they share a critical property: the customer is the AI’s point of contact, and the burden of interpreting the AI’s output is exported to the customer.
That is not, in most operational businesses, where AI’s biggest advantage sits.
The product-layer companies typically follow a predictable arc. They ship a demo. The demo impresses. Enterprise customers pilot. The pilots go well in curated conditions and fail in messy ones. Renewals stall. The team spends the next twelve months trying to make the demo work in production, competing for attention with a dozen other companies doing the same thing.
Meanwhile, the operating model of the company is unchanged. The margins are unchanged. The unit economics look the same on the other side of the pilot cycle as they did before it.
What “AI as a plumbing layer” looks like
The plumbing-layer approach puts AI inside the operating model. It doesn’t touch the customer directly. Instead, it eliminates invisible cost centers that live below the surface — the ones that erode margin, slow down decisions, and require humans to do work that no human should be spending time on.
Some concrete examples of what plumbing looks like:
Reply classification. In cold outbound programs, the largest operational cost isn’t sending — it is separating hot responses from unsubscribe-shaped responses, auto-reply-shaped responses, and out-of-office-shaped responses. Legacy tools bucket them all together, and hot leads die in the noise. An AI classifier trained on a real production reply corpus can separate them at a level of accuracy no rule engine ever will.
Copy allocation. The traditional workflow — a monthly meeting where a small group of humans debates which subject line and body variant is winning — is a candidate for total elimination. AI can re-weight variants continuously against live open and reply signals with no human in the loop.
Data reconciliation. Every service business runs on the invisible tax of manual data reconciliation between systems. AI handles this at scale, without complaint, and without the errors that manual reconciliation invariably produces.
None of these are customer-facing. None of them will show up in a marketing deck. They will, however, show up in operating margins.
The economic signal
The companies I am watching most carefully — and, in the case of one specific company in a portfolio I help operate, the transformation I am seeing play out in real time — are the ones running historically services-shaped operating models onto AI plumbing.
The old model, a manual services business with human-heavy delivery, generates around twenty-five percent operating margins on a good day. The AI-plumbed version of the same business, executing the same customer-facing service, migrates over time toward the kind of margin profile that SaaS businesses have been running.
That is not a small transition. It is not a demo. It does not require a chat interface. It is a fundamental change to the operating model of a business, produced by the deliberate application of AI to the parts of the operation that were previously invisible cost centers.
The companies executing that transition are the ones I would bet on.
The one place a human still belongs
There is one asymmetry worth naming.
AI is wrong most expensively at the initial hot-lead handoff — the moment where a real prospect, with real intent, first touches a human on the seller’s team. A false negative there is not a training-data issue. It is a lost customer relationship, and no amount of model tuning after the fact will recover it.
The plumbing-layer discipline is to run AI unattended wherever the cost of being wrong is small, and to keep a human on the loop wherever the cost of being wrong is a customer relationship. That is not a compromise. That is a design principle.
The framing that follows
For anyone leading a company, the question worth asking is not “where can we add an AI interface to what we already ship.”
The question is “where in our operations is a cost center that only exists because we have not yet automated the interpretation of ambiguous signals?”
Almost every operating business has one. Most have several. The teams that find them, and that build the AI plumbing to eliminate them, are compounding an advantage that the demo-shipping teams will not catch.
That is the divide, and it is widening.
Ronald Hoplamazian is the Managing Member of Pluribus Capital LLC and serves as Chief Revenue Officer at SalesDesk, a portfolio company. He previously spent 13+ years at GE Capital across board oversight and institutional lending roles.