AI is changing the economics of software by making it possible to automate work before we fully understand how to specify the solution. Businesses can begin with general-purpose inference, learn from operating the application, and progressively convert recurring decisions into conventional software. What initially requires an expensive language model may eventually run as inexpensive code. The consequence is a different approach to software development, in which intelligence is used not only to perform work but to discover how that work can be automated more efficiently.
Traditional software development requires a substantial fixed investment in understanding requirements, designing solutions, writing code, and testing the result. Once built, however, software can perform millions of transactions at relatively low marginal cost. Large language models (LLMs) changed this calculation by allowing businesses to purchase reasoning as a recurring service rather than specify every decision in advance. Vibe coding changes it again by reducing the cost of creating specialized software. These are not necessarily competing approaches. They can represent successive stages in the economic life of an application.
Consider an insurer automating claims processing. Initially, an LLM might interpret documents, evaluate circumstances against policy provisions, and recommend decisions. The company can begin automating without knowing which cases admit reliable rules and which require judgment. Operating experience reveals recurring patterns, exceptions, and errors. Once the insurer establishes that a category of decisions can be specified and validated, it can implement those decisions in conventional software while retaining inference for ambiguous cases. The application becomes cheaper to operate as its underlying logic becomes better understood.
This progression follows the task-based economics developed by Daron Acemoglu and Pascual Restrepo. Firms allocate work among people and technologies according to their relative productivity and costs. AI makes that allocation more dynamic because it can both perform a task and help create a cheaper way to perform it. A task might begin with human judgment, move to general-purpose inference, and eventually become conventional software. Each transition becomes attractive when the alternative provides an acceptable outcome at lower total cost.
The boundary is determined by the economics of individual decisions, not the apparent complexity of an application. A network router handles enormous variability through well-defined rules, while a routine customer complaint may require substantial interpretation of intent and circumstances. High transaction volumes favor specialization because development costs can be spread across more executions. Strict latency and correctness requirements can favor it even at lower volumes. General-purpose inference retains its advantage when requirements remain uncertain, exceptions are frequent, or the cost of maintaining specialized software exceeds the savings from eliminating model calls.
Two additional economic principles explain why beginning with inference can be rational even when conventional software will eventually be cheaper. Kenneth Arrow’s learning-by-doing framework describes how productive knowledge accumulates through experience. Real options theory explains the value of postponing an expensive commitment while uncertainty remains high. An LLM application can combine these advantages: it performs useful work while generating evidence about what a more specialized system should do. The organization can defer part of its development investment until it understands the problem well enough to make that investment more confidently.
Inference can therefore serve as both a production method and a discovery mechanism. This is a meaningful change from software development processes that require extensive specification before useful automation begins. Businesses can test demand, observe decisions, and refine requirements through operation before committing to a permanent implementation. AI-assisted development then reduces the cost of converting what they have learned into executable software. The process resembles progressive capital investment, with each stage justified by the knowledge and demand established in the preceding one.
Recent research by Jens Lehmann illustrates the technical capability. Frontier language models constructed functioning neural networks directly from the rules of unfamiliar games, achieving up to 90% success on unseen episodes. Yet conventional Python programs written by the same models achieved 97% to 100%. The finding reinforces what vibe coding already demonstrates: the valuable capability is translating an understanding of a problem into executable computation, not necessarily using a language model or neural network to execute it. The research does not establish that models can independently discover correct rules from operating experience, but it demonstrates their ability to implement rules once those rules are understood.
The remaining obstacle is the cost of establishing correctness. Generating code is only part of building reliable software. Requirements must be validated, exceptions understood, systems integrated, and implementations tested against circumstances they may not have encountered. Faster code generation does not eliminate these investments, which explains why improvements in programming productivity need not produce equivalent reductions in total development costs. The next economically important advance may therefore be AI that can reliably test, verify, and maintain software, rather than simply generate it faster. Such capabilities would lower the fixed cost of specialization and make it attractive for progressively smaller categories of work.
This has consequences for employment because assistance and automation operate differently. An LLM that helps an employee process more claims increases that employee’s productivity. Software that independently processes those claims reduces the labor required for the task. Lower costs may stimulate additional demand and create other work, but those opportunities need not arise in the same occupations or organizations. As AI reduces the cost of turning human expertise into reusable software, businesses can increase output without proportionally increasing employment. The economic benefit depends increasingly on who develops, owns, and operates the resulting systems.
The same reduction in development costs also expands the market for software. Bresnahan and Trajtenberg’s theory of general-purpose technologies explains how foundational innovations create value through complementary products and production methods. Christensen’s new-market disruption describes how less expensive, initially inferior technologies can bring previously excluded consumers into a market. Vibe coding exhibits both mechanisms. Nonprogrammers can create applications that would never have justified conventional development, expanding software production beyond the organizations and activities that historically could afford it. These new producers gain access to capabilities that once required substantial technical expertise and investment.
The result is simultaneous substitution and market expansion. Established businesses can convert recurring inference into cheaper software, while new businesses and individuals can automate activities that were previously uneconomic. Both effects increase the productive reach of AI, but they have different implications for computing demand. Some applications will consume substantial inference during development and increasingly little during execution. Others will continue to depend on models because their decisions require ongoing interpretation, adaptation, or judgment.
Jevons’ paradox becomes relevant here, although it does not determine the outcome. Greater efficiency can increase aggregate resource consumption when lower costs stimulate enough additional demand to outweigh the savings. Cheaper inference and software development may enable so many new applications that total token consumption continues to grow, even as established applications become less dependent on inference. But expanding software consumption is not equivalent to expanding inference consumption. An application created with an LLM may execute millions of transactions without invoking that model again.
This distinction matters for companies investing in AI infrastructure. Forecasts that extrapolate inference demand from the growing number of AI applications implicitly assume that applications retain something like their initial inference intensity. Yet economically successful applications may progressively substitute conventional computation for model calls. The relevant measures are the number of new applications, their inference requirements during development, and how much recurring inference they retain as they mature. Infrastructure demand depends on whether new activity creates more inference than established applications eliminate through specialization.
The broader economic implication is that AI may change the process of creating productive software more permanently than it changes the technology used to execute that software. General-purpose inference makes previously inaccessible tasks economical to automate. Operating those applications generates knowledge about how the work can be performed. AI-assisted development converts that knowledge into specialized implementations, lowering execution costs and making additional tasks worth automating. What begins as a way to avoid expensive software development becomes a way to make software development itself less expensive.
This changes where competitive advantage may accumulate. If generating code becomes inexpensive, the scarce capability is no longer simply writing an application. It is identifying valuable problems, discovering which decisions can be reliably formalized, and validating software that performs them correctly. Organizations that can systematically convert operating experience into reusable computation will be able to reduce costs and improve their applications faster than those that continue paying for the same reasoning indefinitely.
The lasting economic contribution of LLMs may be that they make far more problems solvable through conventional software. Some tasks will continue to require general-purpose intelligence every time they are performed. Others will require intelligence primarily to discover, develop, and improve a solution. The distinction will determine not only how much inference the economy consumes, but how software companies compete, how businesses organize work, and where the economic returns from AI ultimately accumulate.