What are we actually studying when we study AI in finance?

Artificial intelligence is changing how financial work is performed, but its importance extends beyond the automation of familiar tasks. AI is also becoming a source of economic disruption, influencing productivity, competition, capital investment and financial markets, while offering researchers new ways to measure economic information that was previously difficult to quantify. Understanding these developments requires distinguishing the different roles AI can play in financial research.

Consider something as ordinary as a company's earnings announcement. An investment analyst might use a language model to prepare a briefing, an economist might investigate how the company's investment in AI affects productivity or competition, while another researcher might use similar technology to measure uncertainty in management's outlook or identify economic relationships with other firms.

Although the same document, and potentially even the same model, might be involved in each exercise, the underlying research questions are fundamentally different. In the first, AI performs financial work; in the second, it is a technology whose economic consequences are being studied; in the third, it becomes an instrument for measuring something about the economy.

This distinction is useful because discussions of AI in finance often bring together research that shares little beyond its use of machine learning. Much of the established literature has focused on forecasting asset returns, estimating risk or improving portfolio decisions, but the research agenda now extends well beyond whether a model can predict financial outcomes more accurately.

Across recent academic work, industry research and conferences, three broad roles emerge: AI as a tool for financial practice, AI as a source of economic change, and AI as a means of measurement. These roles overlap, but each raises different questions about what constitutes a useful result and how that result should be evaluated.

AI as a tool for financial practice

The first role is perhaps the most familiar, involving the use of artificial intelligence to perform, support or reorganise financial work. Banks process applications and monitor transactions, investment managers analyse information and construct portfolios, while compliance teams examine documents and assess whether activities meet regulatory requirements. Many of these processes contain tasks that AI can accelerate or partially automate.

Financial automation itself is not new. Conventional robotic process automation has long been used to execute predefined rules, while statistical models have supported credit scoring, fraud detection and trading decisions for decades. What distinguishes newer language-model applications is their ability to interpret comparatively unstructured information and respond to situations that have not been individually programmed, allowing automation to extend into activities previously requiring substantial human judgement.

The emerging interest in autonomous AI agents takes this further. Instead of producing a summary or answering a question, an agent might retrieve a financial filing, extract contractual obligations, consult internal records and prepare an exception report for review. The significance lies not simply in the model's ability to interpret language, but in its capacity to coordinate several steps within a financial workflow, which introduces questions about access permissions, supervision, auditability and responsibility for errors.

The CFA Institute Research Foundation's AI in Asset Management surveys applications across investment management, including machine learning, natural language processing and reinforcement learning. Its research on explainable AI highlights an important practical constraint: a model that performs well statistically may still be difficult to incorporate into a process in which investment professionals, risk managers and regulators need to understand or challenge its decisions.

Research in this area is therefore concerned with more than predictive accuracy. It asks whether AI can perform financial tasks reliably, how human and automated decisions should be combined, and whether the resulting systems improve efficiency without compromising governance or accountability.

AI as a source of economic change

The second role shifts the emphasis from using AI within financial institutions to studying its consequences for the wider economy. If artificial intelligence changes the cost of producing information, software, financial advice or other services, its adoption may affect productivity, employment, competition and ultimately the distribution of economic value between firms and households.

Evidence of productivity improvement is already emerging, although translating these gains into economy-wide effects remains considerably more difficult. In a study of 5,172 customer-support workers, Brynjolfsson, Li and Raymond found that access to generative AI assistance increased the number of issues resolved per hour by approximately 15% on average, with particularly substantial improvements among less experienced workers.

These results demonstrate that AI can improve productivity in specific activities, but they do not imply an equivalent increase in aggregate economic output. Daron Acemoglu's analysis of the macroeconomics of AI arrives at more modest estimates of overall productivity gains, reflecting the proportion of economic tasks that can be affected, the savings achieved and the costs of implementation. The difference illustrates why improvements observed within individual workplaces cannot simply be extrapolated across the economy.

The distribution of those gains is another important consideration. Babina, Fedyk, He and Hodson, studying firms' investment in earlier generations of AI, find evidence of growth through product innovation, with benefits concentrated among larger firms and associated with increasing industry concentration. A technology may improve productivity while simultaneously altering competitive advantages, market shares and the balance between established firms and new entrants.

For finance, however, AI also matters because its anticipated economic benefits are already influencing investment decisions on an extraordinary scale. Long before the full productivity effects can be measured, companies must finance the infrastructure required to develop and deploy increasingly capable systems.

Although AI is commonly encountered through software, its expansion requires substantial physical investment in semiconductors, memory, networking equipment, data centres, electricity and cooling. Constraints on access to computing capacity can consequently affect both the economics of AI services and the competitive positions of the companies providing them.

For financial researchers, this second role consequently encompasses far more than the use of AI in banking or trading. It includes the effects of AI on productivity and industrial organisation, the changing demand for physical and financial capital, and the possibility that risk concentration create vulnerabilities within financial markets.

AI as a means of measurement

The third role is less immediately visible in everyday financial practice, but represents a particularly interesting development for researchers. Rather than using AI to make a decision or examining its economic consequences, the objective is to extract economically meaningful information from data that conventional statistical methods struggle to represent.

Economics and finance contain many important quantities that cannot be observed directly. Expectations, uncertainty, competitive relationships, technological exposure and the economic connections between firms must often be inferred from imperfect evidence. Financial researchers have long used market prices, accounting data and surveys to construct such measures, although substantial information remains embedded in documents, news reports, corporate communications and other unstructured sources.

The use of language as economic data has an established history. In their 2011 paper, Loughran and McDonald demonstrated that general-purpose sentiment dictionaries frequently misinterpret financial terminology, with words that appear negative in ordinary English carrying quite different meanings in corporate disclosures. Their work highlighted a methodological problem that remains relevant today: processing language consistently does not guarantee that the resulting score measures the intended economic concept.

Later research extended textual analysis beyond sentiment. Hoberg and Phillips used companies' descriptions of their products to construct changing networks of competitors, allowing economic similarity to emerge from firms' activities rather than relying exclusively on fixed industry classifications. Their approach was supported by comparisons with other evidence of product-market competition, illustrating how a measure derived from language can be given an economically meaningful interpretation.

Similarly, Hassan and co-authors developed firm-level measures of political risk using earnings-call discussions, examining the content of those conversations and their relationships with subsequent corporate behaviour. The objective was not simply to predict returns or classify documents, but to construct a variable representing a particular dimension of economic uncertainty.

Modern language models expand the opportunities for this type of research by representing passages through learned relationships between words and concepts, rather than depending entirely on predefined vocabularies. This makes it possible to compare descriptions expressed in different language, distinguish topics within documents and represent companies through collections of information rather than a single numerical score.

For example, news about a diversified company may concern several unrelated activities, each with its own competitors, customers and sources of risk. Treating the articles as a collection of semantic representations offers a way to investigate whether that company shares meaningful economic exposures with different groups of firms, including relationships that conventional industry classifications might overlook.

Such approaches could ultimately complement financial models of similarity, dependence and diversification, although the interpretation of a language-derived relationship requires care. Companies appearing close together in a semantic representation need not have economically similar risks, just as firms frequently mentioned together in news reports do not necessarily have a direct commercial relationship.

This distinction is fundamental to AI as measurement. A model that generates a score for corporate uncertainty, technological exposure or competitive similarity has produced a candidate measure, but whether that score represents the proposed economic construct remains an empirical question.

Validation therefore needs to extend beyond statistical prediction. Researchers must establish what is being measured, examine the data and assumptions used to construct the variable, compare its behaviour with independent evidence and test whether the interpretation remains reliable across different periods, industries and models. Language-model applications introduce additional complications, including changes between model versions and the possibility that information from later periods has entered a model's training data.

The importance of this research direction lies in making economically relevant information available for systematic analysis, while retaining the theoretical and empirical discipline needed to interpret it. AI can make previously inaccessible information measurable, but the resulting numbers are only as useful as the economic meaning that can be established for them.

Different roles, different evidence

These three roles are not intended as mutually exclusive categories. A researcher might use a language model to measure firms' exposure to AI, examine whether that exposure explains differences in productivity or competition, and subsequently incorporate the findings into an investment process. The same research programme would involve measurement, economic consequences and financial practice, but each stage would require different forms of validation.

What makes the present period particularly interesting is that these roles are developing simultaneously. Financial institutions are finding new ways to incorporate AI into their operations, businesses and investors are committing capital in anticipation of potentially substantial economic changes, while researchers are beginning to explore whether the same technologies can provide new measures of the economy undergoing that transformation.

Distinguishing these roles helps establish what evidence we should expect from each, while also revealing the connections between them. As AI becomes both an economic force and an instrument for studying the economy, those connections may prove more important than any individual improvement in forecasting performance.