Toward a Mature AI Economy: Policy Priorities for the Road Ahead

To move from artificial intelligence (AI) as a shiny new object to AI as a mature driver of economic value, a commensurate shift will need to occur from simply embedding AI persistently in preexisting processes to leveraging AI as a catalyst for uncovering innovations that lead to follow-on products and solutions. The goal is not just to use AI; the goal is to use AI as a tool to accomplish specific outcomes. With this forward-looking perspective, one can seek to extrapolate a future AI economy by considering the kinds of AI business models that are likely to have increased profitability in the long-term and then align policy priorities to get there. Against the backdrop of geopolitical competition, this is an important moment to shift attention from simply protecting underlying technical foundations to supporting rational follow-on uses.

Defining the AI Economy’s Next Phase

AI optimists and AI skeptics can agree on one thing: The state of the AI economy has not yet matured. There will be sufficient opportunities for both sides to claim victory through the ups and downs. Along the path to maturity, the development of AI technology—both the design of chips and the training of models—will continue to improve, but it will be use cases and business models that ultimately determine what the AI economy will look like. When it comes to business models, margins matter. A critical reason the software as a service (SaaS) market originally led to the explosion of Silicon Valley venture capitalist (VC) firms was the enticement of high margins. SaaS startups figured out how to design one piece of software and then sell it to everyone, with upfront investments that were relatively low. Add to that the ubiquitous subscription pricing model and digital advertisement ecosystem, and revenue grew steadily while marginal cost went down.

The rise of AI distorts and muddies the high-margin business models of prior internet booms. Current AI costs charged to users are artificially low, while the feverish buildout of data centers makes it seem as if AI use is limitless. As the true costs of AI tokens, API calls, and cloud compute resources become visible and are passed on to organizations, calculations will shift. Current enterprise AI use cases are predominantly attempts at improving preexisting processes through persistent agentic AI usage; in the long run, such approaches are likely to result in decreased margins, regardless of revenue, given the expectation of increased costs of large language models. Even once tiered pricing against model sizes and model capabilities allows for better management of token allocations and AI costs, the focus on process improvements should be only the first step.

From Improving Processes to Catalyzing Innovations

This is not to say that AI-enabled process improvements cannot be valuable on their own. Speeding up preexisting processes can open up time for employees to focus on higher value-added activities, highlighting a connection between process improvement and societal innovations (though the evidence on productivity gains from AI remains unclear outside of coding work). Regardless, for the economy as a whole, true transformation will occur when AI use leads to new innovations, new companies, new products, and new industries, enabling forms of Schumpeterian creative destruction. The business models that can define such transformations need to be ones with thoughtful and targeted AI use that account for long-term margins, enabling innovative full redesign of workflows that can make many preexisting processes obsolete. That is, processes require fundamental transformations, not just becoming faster with AI. Organizations must fully evaluate the places where AI really can support work by changing work, and where AI should not, and alter their business models in response.

Here, it is essential to differentiate between AI—and specifically agentic AI—as used within IT software and AI as applied to downstream use cases. The core activity of software design has a history of gaining speed and ease through progressive abstraction layers (i.e., coding languages abstracted away the need to understand machine language, and code libraries provide prewritten code for common tasks) and, in this sense, AI for software is essentially the newest, and potentially the most impactful, layer of abstraction in the evolution. This makes productivity and efficiency gains clearer in the case of writing and testing code; though, even here, the single-minded drive to use more and more tokens is in itself a symptom of measuring the wrong things.

For the rest of the economy, the real value of downstream applications of AI is currently less clear. A project autonomously generated by a reasoning model, even with results better and faster than those of an entry-level employee, could turn out more expensive than a knowledgeable human with the context, both to avoid AI-generated “work slop” as well as to have then internalized the insights of the project and its inputs for use in follow-on negotiations. This is compounded by the fact that, unlike other utility-like inputs—such as electricity, water, or broadband internet—outputs from the use of AI are not predictable nor standardized. Though AI is now driving growth in the VC market, the startups with the most excitement in the United States are the ones building out frontier models, indicating that there remains uncertainty about the long-term profitability of AI use in downstream practical applications.

Thus, business models that figure out how to use AI to catalyze innovations, which do not then depend on the most resource-intensive versions of AI for ongoing activity, are likely to become more valuable and better suited to driving sustained economic growth. AI should be viewed as a catalyst and not as a crutch. Knowing when not to use AI will be as important as knowing when to use AI. What a relief that would be from an environmental perspective, too, if the most pervasive uses of AI drive carry-on implications rather than require persistent energy and token consumption.

This matters little in the short term, as investors and enterprises are more concerned with the fear of missing out than the fear of betting on the wrong AI use cases. And some initial overuse of AI is likely necessary to start to narrow down the truly profitable opportunities. Hence, China’s full-stack approach to AI adoption and experimentation can push the country further along the AI journey, even as much of its adoption remains inefficient. But as the AI journey progresses from frontier AI model development, to broader adoption of AI models focused on process and productivity, to ultimately margin-sensitive and innovation-enabling utilization, the organizations and the countries that have thought deeply about applying AI in scalable manners are going to weather investment resets and bubble bursts better. This will likely mean that some organizations fail and new industries develop. Similarly, the country that manages to go through these steps first toward effective high-margin AI adoptions is likely to gain the upper hand in the geopolitical AI race by ensuring a more dynamic and vibrant economy.

Policy Priorities for a Mature AI Economy

Policymakers in the United States can encourage progress along the AI journey by creating an environment for long-term, sustainable uses of AI:

1. This would mean shifting policy attention away from the “tech sector” to industry-specific practical AI applications. Since the potential uses of AI are diverse, as are the risks, regulating and enabling across industries becomes overly generic. How to enable or restrict AI in medicine production by the pharmaceutical industry would be quite different from how to create opportunities for self-driving cars on the road, for example. Shifting to industry-specific AI would also allow for more effective management of public skepticism about AI, as it encourages practical conversations anchored on resolving concerns from concrete impacts and moves away from ambiguous anxieties of AI in general terms. Ensuring there continues to be a vibrant space for startups within industries is particularly important to drive innovation. Thoughtful competition-focused legislation is key here; California’s proposed Blocking Anticompetitive Self-preferencing by Entrenched Dominant platforms (BASED) Act, with some smart and some flawed rules, is ahead in recognizing that AI can enable more lock-in to platforms, which can, in turn, limit innovation.
2. Recognizing that data privacy becomes more critical than ever, not less, as data is crucial to the improvement of AI models and the design of new AI-enabled business models. While AI provides the tooling, it is regulatory certainty on data privacy, as opposed to the current regulatory vacuum, that can enable the necessary experimentations. Regulations about which kinds of information can be used to target individuals and groups, and how people can control the extent to which their data is used, will give people the confidence to trust the outputs of AI. This also ties back to the industry-specific approach, as data will be used differently in different industries. For example, using AI to go through large volumes of patient and drug trial data in order to create individualized drug treatments could be an innovation in pharmaceuticals that makes it more logical to store health data in ways that would be unacceptable for, as another example, data on shopping habits.
3. Investing in education and training of AI fundamentals. Policies that can spread the wealth, so to speak, would deepen the economic benefits for society, as greater investments in AI training are critical so that AI becomes a tool that everyone can leverage to solve real-world challenges. This would require both changing how and when AI and new ways of coding are taught in school, as well as creating upskilling opportunities for existing employees. Large-scale workflow redesigns to enable innovations often fail due to inertia in organizations, which can only be managed once upskilled employees are involved in the necessary changes. These structural pieces will create the long-term scaffolding to manage the maturation of the AI economy.

The AI economy is still embryonic. In the progression, greater value from this new technology can be spurred from business models focused on innovation and not just process efficiency. Policymakers should set in place the foundations for encouraging long-term AI use that seek higher margins through innovative redesigns of entire workflows. In the long run, the focus of geopolitical competition should be about enabling scalable AI, not just more AI.

Yinuo Geng is an adjunct fellow (non-resident) with the Strategic Technologies Program at the Center for Strategic and International Studies and a managing vice president with Gartner in Washington, D.C. Views expressed are the author’s own.