MIT Technology Review: Competitive advantage in enterprise AI comes not from the model, but from the operational layer

In corporate AI, long-term advantage comes not from the model itself, but from control over the layer through which intelligence reaches real business processes. This is where decisions are made about what data the system sees, who confirms disputed decisions, how human corrections are factored in, and whether individual successful responses become sustainable operational practice.

In a piece for MIT Technology Review, this layer is described as a bundle of workflow software, data collection, feedback loops, and governance rules situated between the model and real work. Public discussion still revolves around comparisons of GPT, Gemini, and other foundation models, but for large companies this is no longer the main question.

If intelligence is invoked via API as a one-off service, it can be very powerful, but remains loosely coupled with the daily operational environment and barely accumulates state from case to case. What matters far more is this: does knowledge reset with each new request, or does the system actually learn as it works.

Hence the key distinction between two approaches. The first treats AI as a request-based service: there is a task, there is a model call, there is an answer. The second embeds AI into the company's operational layer, where every exception, correction, approval, and disputed decision becomes a signal for learning and a reason to refine rules.

In such an architecture, value is created not only by the quality of the model, but by how deeply the company can instrument its own processes, collect data about work progress, and turn employee decisions into reusable policy.

Against this backdrop, the authors dispute the popular thesis that AI-native startups will inevitably outpace incumbents. If you treat AI as purely a model race, such a scenario seems plausible. But in the enterprise environment, the challenge is usually systemic: integrations, access rights, quality assessment, change management, SLA, cost control, and compliance requirements.

Here, advantage goes not to those who simply connect a new model faster, but to those already embedded in high-volume, high-risk processes who can turn that position into a continuous learning loop.

From this follows an inversion of familiar work logic. Traditionally, expert service companies are structured so that people use software to perform complex work, while technology is merely the environment. An AI-native platform works the opposite way: it accepts a case, applies accumulated domain knowledge, autonomously executes what it is confident about, and hands to humans only narrow subtasks where judgment, context, or responsibility is still required.

Essentially, AI executes, and people arbitrate.

Particularly important is the thesis about three assets that large incumbent companies already possess. These are proprietary operational data, a large group of domain experts daily generating training signals, and accumulated implicit knowledge about how work actually gets done in complex conditions.

But these assets alone do not yet create a moat. They begin to work only when the company knows how to translate scattered decisions, exceptions, and heuristics into machine-readable signals, and then return the result back to the operational system.

Revenue cycle management in healthcare is given as an example. The Ensemble approach is to first populate the system with explicit domain knowledge, then through daily interaction with operators identify gaps, ask targeted questions, and cross-check answers from multiple experts to capture not only broad consensus but also the nuances of edge cases.

This forms a living knowledge base that reflects not just the final decision, but the logic behind expert action. When the system becomes sufficiently constrained and manageable, each decision by an experienced employee becomes a potentially labeled example for further improvement.

The practical takeaway for CIOs, CPOs, and platform leaders is quite harsh: the debate over whose base model is better increasingly determines less of the outcome of the enterprise race. The main question now is who owns the AI operational layer inside the company — who controls data, access rights, cost, routing, audit, and learning loops.

Sustainable advantage will go to organizations that can turn their knowledge, decisions, and everyday expertise into infrastructure that improves with use.