MIT Technology Review: In Enterprise AI, the Edge Comes From the Operational Layer, Not the Model

In corporate AI, long-term advantage comes not from the model itself but from control over the layer through which intelligence enters real business processes. That is where it is decided which data the system sees, who confirms disputed decisions, how human corrections are taken into account, and whether isolated good answers turn into a stable operational practice. An article for MIT Technology Review describes this layer as a combination of workflow software, data collection, feedback loops and governance rules, sitting between the model and real work.

Public debate 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 called via API as a one-off service, it can be very strong, yet it remains loosely connected to the day-to-day operating environment and barely accumulates state from case to case. What matters far more is whether knowledge is reset with every new request or the system genuinely learns as work goes on.

This leads to the key difference between two approaches. The first treats AI as an on-demand service: there is a task, a model call and an answer. The second embeds AI in the company's operational layer, where every exception, correction, approval and disputed decision becomes a learning signal and a reason to refine the rules. In such an architecture, value comes not only from model quality but from how deeply a company can instrument its own processes, collect data on how work unfolds, and turn employees' decisions into reusable policy.

Against this backdrop, the authors challenge the popular thesis that AI-native startups will inevitably overtake incumbents. If AI is seen purely as a model race, that scenario looks plausible. But in the enterprise, the problem is more often systemic: integrations, access rights, quality evaluation, change management, SLAs, cost control and compliance. The advantage goes not to those who simply plugged in a new model faster, but to those already embedded in high-volume, high-risk processes and able to turn that position into a continuous learning loop.

This also inverts the usual logic of work. Traditionally, expert service companies operate like this: a person uses software to do complex work, and technology is merely the environment. An AI-native platform works the other way around: it takes the case, applies accumulated domain knowledge, autonomously does what it is confident about, and hands the human only narrow subtasks where judgment, context or accountability are still required. In essence, AI executes and people arbitrate.

Particularly important is the thesis about three assets that large incumbents already have. These are proprietary operational data, a large group of domain experts who produce training signals every day, and accumulated tacit knowledge of how work is actually done in complex conditions. But on their own, these assets do not yet create a moat. They start to work only when the company can translate scattered decisions, exceptions and heuristics into machine-readable signals, and then feed the result back into the operating system.

Healthcare revenue cycle management is given as an example. Ensemble's approach is to first fill the system with explicit domain knowledge, then, through daily interaction with operators, identify gaps, ask targeted questions and cross-check the answers of several experts, in order to capture not only the general consensus but also nuances in edge cases. This builds a living knowledge base that reflects not just the final decision but the logic behind the expert's action. Once the system becomes sufficiently bounded and manageable, every decision by an experienced employee turns into a potentially labeled example for further improvement.

The practical takeaway for CIOs, CPOs and platform leaders is fairly stark: the argument over who has the better base model increasingly determines the outcome of the enterprise race less and less. The main question now is who owns the AI operational layer inside the company, meaning who controls data, permissions, cost, routing, audit and learning loops. Lasting advantage will go to organizations that can turn their knowledge, decisions and everyday expertise into infrastructure that improves with use.