From Artificial Intelligence as a Tool to Intelligence as Productive Capital The economic trade-off is between autonomy and control. Excessive approval requirements can eliminate the productivity gains of automation by forcing employees to supervise every step. Insufficient controls can expose the business to losses that overwhelm those gains. The objective is risk-adjusted autonomy: routine, reversible, low-impact actions can be automated more freely, while sensitive, irreversible, or financially consequential actions receive stronger oversight. This framework is especially relevant to financial applications. An agent that retrieves account information or prepares a payment recommendation has a different risk profile from one that can initiate transfers or change investment positions. The required controls should scale with the potential harm, the reversibility of the action, and the reliability of the evidence supporting it. As adoption expands, trust may become a commercial differentiator. Businesses will favor systems that provide predictable behavior, transparent permissions, reliable audit trails, and credible recovery mechanisms. Security investments can therefore support revenue and retention as well as reduce expected losses. Evaluation, Model Improvement, and the Economics of Technological Diffusion A persistent agent creates a difficult evaluation problem. model may perform well on short, isolated tasks but fail when it must maintain consistency over weeks, respond to changing instructions, preserve critical information, and recover from unexpected events. Reliability over long horizons cannot be inferred directly from performance on brief demonstrations. I see evaluation as an essential link between technological progress and commercial deployment. When an agent makes an error, the failure should be classified, reproduced where possible, and converted into a measurable test. The underlying cause may lie in the model, the orchestration logic, the available tools, the memory system, the permissions structure, or the infrastructure. Correctly identifying the cause prevents organizations from treating every failure as a model problem. This process creates an improvement cycle. Operational failures reveal weaknesses. Engineers adjust the system, modify prompts, improve tools, or change model behavior. Evaluations test whether the correction works. New models can then be assessed against accumulated evidence of real-world performance. Over time, this feedback loop can improve the reliability of the complete product even when the underlying model remains unchanged. Subsequent model improvements may extend the range of tasks the product can handle, creating another opportunity to redesign the system and expand its usefulness. However, rapid model releases introduce a governance challenge. If a business deploys an agent that operates for weeks or months, it may be difficult to fully evaluate a new model before the previous version's long-running tasks have completed. The organization must balance the benefit of adopting improved capabilities against the risk of changing behavior during extended operations. One possible response is to evaluate long-horizon behavior through carefully designed simulations, representative task suites, staged deployments, and monitoring of real-world outcomes. Another is to separate model updates from active workflows or maintain stable execution environments for tasks that require predictable behavior. No single technique eliminates the problem, but the broader principle is clear: the faster the technology evolves, the more important disciplined release management becomes. I also see a potential competitive advantage in organizations that build strong evaluation systems early. company with reliable measurements of task success, error rates, recovery performance, security, and cost can improve its product systematically rather than relying on demonstrations or subjective impressions. Evaluation becomes part of the company's operational knowledge and may be difficult for competitors to reproduce quickly. The broader historical parallel is the maturation of earlier computing technologies. Early products often emphasized what the technology could do. Durable commercial systems emerged as developers learned how to make those capabilities reliable, maintainable, secure, and economically repeatable. AI agents appear to be moving through a similar transition from capability discovery toward operational discipline. Enterprise Knowledge as an Intangible Capital Asset One of the most promising opportunities lies in converting tacit organizational knowledge into structured, reusable information. Many businesses depend on procedures that exist primarily in employee experience, informal communication, and accumulated judgment. These processes can be difficult to scale because new employees must learn them through observation, repetition, and direct supervision. If a company documents how its work is performed, what distinguishes a successful outcome, which exceptions require escalation, and which constraints must never be violated, agents may be able to apply that knowledge repeatedly. The result is a form of operational codification that makes selected capabilities less dependent on the continued availability of a particular employee. I would treat these documented procedures as intangible capital. Their economic value comes from their ability to improve future output without requiring the organization to rediscover the same knowledge every time a similar task arises. They can reduce training costs, shorten execution time, standardize quality, and support expansion across teams or locations. Yet codification has limitations. Some knowledge depends on context, relationships, ambiguous evidence, or judgment that is difficult to express in a fixed procedure. An overly rigid workflow can fail when conditions change. Businesses therefore need a mechanism for updating documented processes, recording exceptions, and distinguishing stable operating principles from assumptions that require periodic review. The most effective organizational approach may combine standardized skills with controlled flexibility. Agents should be able to follow proven procedures for recurring work while recognizing when a task falls outside the conditions under which those procedures were validated. This is where evaluation, memory management, and escalation rules become economically important. For smaller businesses, this could reduce a traditional disadvantage. Large organizations often possess deeper documentation, more specialized staff, and greater resources for building internal systems. If AI tools make it cheaper to encode and reuse operational knowledge, smaller firms may gain access to capabilities that previously required substantial administrative infrastructure. At the same time, the advantage will not necessarily accrue to the companies that collect the most data. It will accrue to those that convert relevant information into accurate, accessible, and continuously improved operating procedures. The distinction between data accumulation and productive knowledge is likely to remain important. Investment Implications: Productivity, Valuation, and Competitive Moats When evaluate the investment implications of AI agents, separate the growth of the technology from the attractiveness of individual securities. large addressable market does not guarantee high returns, and a genuine productivity revolution does not ensure that every company exposed to it will become more profitable. The first question is where economic value is created. Model developers may benefit from demand for intelligence. Infrastructure providers may benefit from rising computation requirements. Software platforms may benefit from workflow integration. Enterprises may benefit from lower operating costs and improved output. Customers may capture some of the gains through lower prices. These outcomes can occur simultaneously, but the division of economic surplus will depend on competition and bargaining power. The second question is how much of that value is already reflected in valuation. An equity price represents expectations about future cash flows, growth, risk, and the time required for those cash flows to materialize. company can deliver excellent operating results and still produce disappointing investment returns if its valuation already assumes an even stronger outcome. This is particularly relevant to high-growth technology companies whose valuations depend heavily on distant earnings. The longer the expected cash flows extend into the future, the greater the sensitivity of their present value to discount rates and changes in growth expectations. Such companies can exhibit substantial duration risk even when they carry relatively little conventional balance-sheet debt. If monetary policy tightens, real interest rates rise, or risk premia expand, investors may demand a lower valuation for the same expected stream of future earnings. Conversely, a decline in discount rates can support valuation multiples even before a company's operating performance changes materially. The interaction between AI adoption and financial conditions therefore matters as much as the underlying technology for many growth-oriented investments. The third question is whether productivity gains translate into durable margins. If competitors can adopt the same models and tools at comparable cost, initial efficiency advantages may be competed away through lower prices. If a business combines AI with proprietary data, distribution, specialized workflows, trusted customer relationships, or superior implementation, it may retain a greater share of the gains. I would pay particular attention to return on invested capital. AI-related capital expenditure can be economically justified when the resulting incremental operating profits exceed the cost of the capital deployed. However, rapid spending on computing capacity, software development, and organizational restructuring can destroy value if demand fails to materialize or if competitive pressure prevents the company from capturing the benefits. For infrastructure businesses, the central questions include utilization, pricing power, capital intensity, depreciation, energy costs, customer concentration, and the pace of technological obsolescence. High demand can support strong near-term revenue while simultaneously encouraging competitors to add capacity. If supply expands faster than durable demand, margins can compress even when aggregate AI usage continues to grow. For application businesses, would examine customer retention, measurable customer return on investment, inference costs per completed task, the frequency of human intervention, and the defensibility of the workflow. Products that demonstrate sustained economic value may support recurring revenue more effectively than products whose appeal depends primarily on novelty or temporary model advantages. I would also distinguish between cyclical and structural drivers. Near-term spending can be influenced by financing conditions, corporate budgets, and expectations about future demand. Long-term value depends more fundamentally on whether the technology produces sustained improvements in output, lowers costs, creates new markets, or changes the competitive structure of industries. My investment framework would therefore focus on the interaction of adoption, unit economics, competitive advantage, capital intensity, valuation, and financial conditions. No single metric captures the opportunity. technologically advanced company can be a poor investment at an excessive price, while a less visible infrastructure or enterprise business may offer attractive returns if its role in the value chain is durable and its future cash flows are underappreciated. Capital Cycles, Overinvestment, and the Risk of Misallocated Resources Technological transitions frequently generate periods in which expectations rise faster than the capacity to generate sustainable profits. Businesses invest in anticipation of future demand, investors finance expansion, and suppliers increase production to capture an emerging market. If the resulting capacity exceeds economically justified demand, the industry can experience declining prices, underutilized assets, and disappointing returns on capital. I see this as a relevant risk for AI infrastructure. Rapidly increasing demand for computing can encourage substantial investment in data centers, networking, power capacity, and specialized hardware. Some of that investment may be essential to support future productivity. Some may be based on optimistic assumptions about utilization, pricing, or the speed at which AI applications become commercially viable. The distinction between useful investment and overinvestment cannot be established solely by observing the size of capital expenditure. Infrastructure may be built ahead of demand and still become productive over time. Conversely, a project can appear commercially attractive during a period of scarcity but generate poor returns once competing capacity comes online. I would examine marginal returns rather than aggregate spending alone. If each additional unit of capital generates progressively less incremental cash flow, the industry may be approaching a point of diminishing returns. If utilization remains high, customers continue to pay for scarce capacity, and productivity gains support expanding demand, continued investment may be justified. Technological obsolescence introduces another layer of risk. AI infrastructure is exposed not only to ordinary demand cycles but also to rapid changes in hardware efficiency, model architecture, and the amount of computation required to perform a given task. An asset that is economically valuable today may become less competitive if newer systems deliver equivalent output at substantially lower cost. At the same time, efficiency improvements can expand total demand enough to support additional capacity. Lower inference costs may make new applications viable, increasing the amount of work performed by AI systems. The resulting relationship between efficiency and infrastructure demand is therefore not linear. Investors must consider both the reduction in resource requirements per task and the expansion in the number of tasks performed. The broader capital-cycle lesson is that technological progress and investment returns are not synonymous. technology can transform the economy while many companies investing in it earn inadequate returns. The beneficiaries may change as scarcity moves through the value chain, from computing hardware to energy infrastructure, from model access to workflow orchestration, or from basic capabilities to specialized enterprise applications. For investors, disciplined capital allocation requires distinguishing temporary bottlenecks from durable competitive advantages. Scarcity can support exceptional returns, but those returns attract capital. The eventual outcome depends on whether barriers to entry, proprietary capabilities, or sustained demand prevent competition from eroding margins. Financial Conditions, Leverage, and the Funding of Technological Change The pace of AI adoption is also influenced by the availability and cost of capital. Large infrastructure projects require substantial upfront expenditure, while the associated productivity benefits may emerge over many years. The economic case for investment depends not only on expected operating gains but also on the financing structure, discount rate, and ability to withstand periods of weaker demand. When financial conditions are favorable, businesses may find it easier to finance expansion, investors may accept longer payback periods, and valuations may reflect more optimistic expectations for future growth. When rates rise or credit spreads widen, the cost of financing increases and investors may become less willing to fund projects whose returns are uncertain or distant. This creates an interaction between monetary policy and technological diffusion. technology may be economically transformative in the long run while its adoption remains sensitive to short-term financing conditions. Companies with strong balance sheets and internally generated cash flow may continue investing through tighter conditions, while highly leveraged businesses or speculative ventures may be forced to delay projects or reduce spending. Leverage can amplify both outcomes. Debt may increase equity returns when operating cash flows exceed financing costs, but it also introduces fixed obligations that remain even when demand disappoints. business investing heavily in rapidly evolving technology may face additional risk if its assets lose value faster than anticipated or if the revenue required to service its debt arrives later than expected. I would therefore evaluate AI-related investment using a combination of operating economics and balance-sheet resilience. Strong demand forecasts are not sufficient if the business cannot survive a period of lower utilization, higher financing costs, or accelerated technological change. Conversely, a well-capitalized company may be positioned to acquire assets, talent, or customers at favorable prices when weaker competitors are forced to retrench. The macroeconomic implications extend beyond individual companies. If AI investment becomes a major driver of aggregate capital expenditure, changes in interest rates and credit availability could influence the timing of infrastructure construction, technology adoption, and productivity realization. Financial conditions would not determine the ultimate technological potential, but they could materially affect the speed and distribution of its economic benefits. Why Simpler Interfaces Can Conceal Greater Economic Complexity One of the most revealing aspects of agent-based software is that a simpler user experience can require a more sophisticated underlying system. The user may interact through a short message, a voice command, or a minimal interface, while the system coordinates memory, permissions, tools, specialized processes, external applications, and persistent workflows behind the scenes. I see this as a recurring pattern in technology economics. Complexity does not necessarily disappear when a product becomes easier to use. It is often transferred from the user to the provider. The provider absorbs the engineering, infrastructure, and operational burden so that the customer can focus on the desired outcome rather than the sequence of actions required to produce it. This transfer can expand a market by lowering the skill and time required to access a capability. Software that previously required technical expertise may become accessible to a broader range of employees and businesses. As adoption grows, the provider gains more opportunities to improve the product, collect operational feedback, and spread fixed development costs across a larger customer base. However, the provider must recover the cost of that hidden complexity. minimal interface can create an expectation that the service is inexpensive even when the underlying execution requires substantial computation and support. Commercial success therefore depends on matching the simplicity of the experience with a sustainable delivery architecture. The same principle affects product strategy. If model capabilities improve rapidly, interfaces built around today's limitations may become unnecessary. workflow that currently requires multiple screens and manual decisions may eventually be handled through a single objective. Products that hard-code too much of the current interaction model may need expensive redesigns as the technology evolves. I would favor modular product architectures that preserve flexibility without sacrificing reliability. The goal is not to anticipate every future capability in advance. It is to minimize the cost of adapting when new capabilities change what users need from the product. For established software companies, this creates both a threat and an opportunity. Existing interfaces and workflows may become less differentiated, but established distribution, customer trust, data access, and domain expertise may remain valuable. The strategic question is whether a company can use its existing advantages to deliver better outcomes in an agent-driven environment or whether its current business model depends too heavily on interactions that automation will make less necessary. Historical Perspective: General-Purpose Technologies and the Productivity Lag I would place AI agents within the broader history of general-purpose technologies: innovations that can be applied across many industries and that require complementary changes before their full economic effects become visible. Electricity, computing, and the internet all created opportunities that extended beyond the immediate capabilities of the underlying technology. The central historical lesson is that invention and productivity realization are separate processes. Businesses must invest in complementary infrastructure, redesign organizational practices, develop new skills, and establish standards before a new technology can materially change production. Early adoption may generate high costs and uneven results even when the long-run potential is substantial. AI agents appear to fit this pattern. The model is only one component. Businesses also need data access, secure execution environments, workflow integration, reliable evaluation, documented procedures, and clear accountability. These complements determine whether AI remains an experimental tool or becomes an integral part of production. The diffusion process may also be uneven. Industries with standardized information, clear success criteria, and well-defined digital workflows may adopt the technology more quickly. Industries with fragmented data, complex regulation, physical constraints, or high consequences of error may require longer implementation periods. This unevenness matters for macroeconomic forecasting. sudden improvement in model performance does not imply an immediate, uniform increase in measured productivity. The effect may initially appear in selected firms or tasks, while broader gains emerge only as organizations reorganize their operations and competitors adopt comparable methods. Historical parallels also caution against assuming that every early technology leader will remain dominant. As an industry matures, competitive advantages can migrate from invention toward distribution, reliability, cost control, customer relationships, and complementary infrastructure. The firms that establish the first compelling demonstration do not necessarily capture the largest share of long-term economic value. The lesson draw is that technological forecasting should be paired with institutional and economic analysis. need to understand not only what AI can do, but also which organizations can deploy it, what complementary investments are required, how quickly customers can change their behavior, and where the resulting surplus is likely to accrue. Strategic Priorities for Businesses and Investors For business leaders, would begin with workflows rather than with the desire to deploy AI everywhere. The most promising opportunities are processes where the business can define a useful outcome, provide the agent with the necessary information and permissions, measure the result, and recover safely from mistakes. These conditions make it possible to estimate the actual return on investment rather than relying on generalized claims about productivity. I would then measure the complete economics of the workflow. The relevant comparison includes labor hours saved, throughput gained, error costs, inference expenditure, implementation expense, human review, maintenance, and the value of faster execution. process that reduces labor costs but increases error rates or introduces substantial compliance risk may not be an improvement. Businesses should also decide which capabilities to build internally and which to obtain from external providers. Internal development may make sense where workflows are strategically differentiated, proprietary data creates an advantage, or deep integration is essential. External services may be preferable for standardized capabilities where competition is strong and building a replacement would divert capital from more valuable activities. For technology investors, would distinguish between companies that sell AI capability and companies that use AI to improve their own economics. The former depend on their ability to capture value within a competitive technology market. The latter may benefit through lower costs, better service, improved product development, or increased output, even if they never sell an AI product directly. For infrastructure investors, would focus on capacity discipline, utilization, energy availability, financing structure, and the risk that technological progress changes the economics of existing assets. Strong demand today should be evaluated alongside the possibility of future oversupply and lower prices. For software investors, would examine whether the company owns a valuable customer relationship, a differentiated data asset, a mission-critical workflow, or a trusted system of record. If its primary value comes from mediating routine interactions that an external agent can perform more cheaply, its competitive position may be vulnerable. For corporate management teams, would treat institutional knowledge, evaluation data, and secure workflow design as long-term assets. They can improve the productivity of current models while creating a foundation for adopting more capable future systems. Their value is likely to compound when they are maintained systematically rather than assembled through disconnected experiments. Across all these decisions, would keep three time horizons separate. In the near term, would monitor implementation costs, capital expenditure, financing conditions, and reliability. Over the medium term, would evaluate adoption, workflow redesign, competitive responses, and unit economics. Over the long term, would assess productivity growth, changes in labor demand, industry structure, and the distribution of economic surplus. Conclusion: The Competitive Advantage Will Belong to Those Who Convert Intelligence Into Output I view persistent AI agents as an important evolution in the relationship between software, labor, and capital. Their significance comes not simply from generating information but from coordinating the tools, memory, infrastructure, and decision processes required to perform useful work over extended periods. The largest economic gains will depend on the reliability of the complete system. Better models matter, but so do durable workflows, efficient context management, specialized execution, secure permissions, robust evaluation, and the ability to recover from failure. These components determine whether intelligence can be deployed at scale without imposing excessive computational, operational, or financial costs. At the business level, the opportunity is to redesign production around a different combination of human judgment and machine execution. At the industry level, the opportunity is to reduce integration costs, expand access to specialized capabilities, and create new software business models. At the macroeconomic level, the possibility is a sustained increase in productivity, although the timing, distribution, and inflationary consequences will depend on complementary investment and the pace of technological diffusion. At the investment level, the central challenge is separating genuine economic value creation from the expectations already embedded in asset prices. AI adoption may expand markets and improve corporate profitability, but competition, capital intensity, financing costs, and valuation determine how much of that progress investors ultimately capture. I would therefore avoid treating AI as a single trade, a single industry, or a guaranteed source of economic growth. It is better understood as a broad technological transition that will redistribute costs, capabilities, bargaining power, and investment opportunities across the economy. Some companies will benefit from selling the infrastructure. Others will benefit from embedding the technology into their operations. Still others may lose pricing power as automation lowers barriers to entry and weakens established intermediaries. The deepest structural change is the potential separation of productive capacity from continuous human attention. If digital systems can reliably preserve knowledge, initiate work, coordinate specialized processes, and deliver results with limited supervision, the cost of scaling many information-intensive activities may decline. That could expand the range of businesses capable of undertaking sophisticated work while changing the economics of established enterprises. Ultimately, believe the decisive question is not which organization has access to the most impressive model. It is which organization can consistently convert intelligence into measurable output, preserve the resulting knowledge, manage the associated risks, and earn an attractive return on the capital required to sustain the process. That is where the long-term economic and investment consequences of autonomous AI will be determined. 01 Stanford CS146S: How always-on agents work