Developers have grown accustomed to the seamless flow of ChatGPT and Claude assisting with boilerplate code and documentation. This shift has moved beyond simple productivity gains into a new era where AI agents do not just suggest text but execute tasks and make decisions. As these agents become the primary users of software, the foundational economic logic of the software industry is fracturing. The traditional per-seat licensing model, which has sustained the SaaS era for decades, is becoming obsolete because the value is no longer in the interface, but in the ability of the AI to judge and execute.
The Architecture of the Data Moat
Recent market movements highlight the extreme value of proprietary interaction data. xAI currently holds an option to acquire Cursor, the AI-powered code editor, for $60 billion. This valuation is staggering when compared to Cursor's annual recurring revenue of approximately $4 billion, but the price reflects something far more valuable than a subscription base. The core asset is the diff data—the granular record of every change a developer makes, every suggestion they accept, and every line they rewrite. While a competitor could easily clone Cursor's user interface, they cannot replicate the years of human-led corrections that serve as a goldmine for model refinement. Cursor has leveraged this diff data, training it on an open-source base to create a specialized model that understands the nuance of professional coding better than a general-purpose LLM.
This trend extends into the financial and legal sectors where specialized AI firms are avoiding the costly race to build foundation models from scratch. Harvey, valued at $11 billion, and Legora, valued at $5 billion, along with the financial AI firm Rogo, have adopted a different strategy. Instead of training a base model, they build a harness—a sophisticated outer layer that controls inputs and outputs while meticulously capturing how experts correct the AI's work. In Rogo's case, the act of a financial analyst editing a model-generated memo is converted into a proprietary data asset. These companies are not betting on the raw power of the underlying LLM but on the compounding value of expert judgment data.
As agents begin to handle real-world transactions, such as booking flights or ordering industrial parts, the need for a financial settlement layer has become urgent. Stripe has already launched protocols to facilitate these agent-to-vendor transactions, while Visa and Mastercard are fighting to establish the global standard for agent payments. OpenAI is already implementing a system where it takes a percentage fee from the transactions its agents execute, signaling a move away from simple API credits toward a transaction-based economy.
From Context Windows to Judgment Layers
There is a fundamental difference between providing a model with context and building a judgment layer. Most AI applications currently rely on RAG or large context windows to give the model information, but this does not solve the problem of accuracy. The real value lies in the corrections. When a user fixes an AI's output, they are not just correcting a typo; they are providing a scorecard that defines what correctness looks like in a specific professional domain. This scorecard serves two purposes: it acts as a training signal to optimize a borrowed model for a specific industry and functions as the only reliable test set to measure if the agent is actually improving.
This creates a layer of judgment that is essentially untrainable for outsiders. Because this data is generated through the iterative process of human-AI collaboration, it compounds over time. This is the only area of the AI stack where a company can build a moat that is not easily disrupted by the next leap in foundation model performance. When the judgment layer is owned by the application, the underlying model becomes a commodity utility, while the application becomes the indispensable intelligence layer.
This shift necessitates a total overhaul of how software is monetized. The per-seat model fails when a single human user deploys one hundred thousand AI agents to perform tasks. Charging for one hundred thousand seats is an impossibility, and as model performance increases, the value of the software's basic features drops toward zero. To survive, AI applications must evolve into a hybrid of data companies and fintech companies. They must secure the write-access to critical data and the payment rails that move money.
Shopify provides the blueprint for this transition. It began as store management software but expanded into a payment system and eventually launched Shopify Capital, which provides loans based on the store's internal revenue data. Because Shopify sees the actual flow of money, it can offer credit terms that a traditional bank cannot. Today, roughly three-quarters of Shopify's total revenue comes from these financial services rather than software subscriptions. Similarly, companies like Toast and Ramp have built their empires by combining network lock-in with fintech models. By tying data to the flow of capital, these companies maintain high margins even as the cost of the underlying technology falls.
The era of software as a tool is ending, and the era of software as an agent is beginning. The survival of any AI application now depends on whether it can move beyond simple feature sets to capture the judgment of its users and the movement of their money.
