AI in the investment process

Asset managers today face a paradox: never has so much information been available, yet the ability to process and interpret it remains constrained. The result is a growing imbalance between data and attention—one that is reshaping how investment research is conducted and where edge can be found.

Artificial intelligence (AI) is beginning to change that equation, compressing the time required to gather, organize, and analyze information. It has the potential to transform the scale and speed of research—but not without trade‑offs. While it may raise the baseline across the industry, it also risks compressing traditional sources of differentiation.

The question for asset managers is not simply whether to adopt AI, but how to do so in a way that strengthens judgment rather than dilutes it. Against this backdrop, it is useful to begin by looking at how T. Rowe Price has approached this transition in practice.

Our approach

T. Rowe Price’s AI journey didn’t begin with generative AI hype—it has been years in the making. In fact, we have been leveraging various methods, including machine learning and natural language processing (NLP)—to support portfolio management and research since 2006.

In 2017, we established our New York‑based Technology Development Center to explore machine learning and advanced analytics across the organization. That work moved closer to the investment process in 2019, with the creation of a dedicated group, the Investment Data Insights team. They worked with our fundamental investment teams to apply predictive analytics and natural language processing to tasks such as analyzing earnings calls and generating insights from large volumes of unstructured research data at scale.

The objective has not been to create isolated use cases, but rather to embed AI across investment teams so that it becomes part of how research is conducted.

The launch of large language models (LLMs) in late 2022 marked a clear acceleration point. Because the underlying capabilities were already in place, we were able to move quickly, rolling out internal tools such as Investor Copilot—now Chat TRP—within months. The next phase was not building everything internally but integrating best‑in‑class tools into investor workflows. That shift led to the creation of the Investment AI Solutions group last year, designed to embed AI ownership and accountability directly within the investment organization.

In practice, this has changed how investors allocate their time—reducing manual information gathering and creating more space for iteration, learning, and idea development. Several teams now prioritize faster synthesis and more frequent questioning over exhaustive first‑pass analysis.

Looking ahead, we are experimenting with agent‑based AI to support more complex, multi‑step workflows, particularly where automation can free up time for higher‑value analysis. In parallel, the firm is investing heavily in education and adoption, recognizing that access to tools alone does not drive impact. New “AI Adoption Specialist” roles are being introduced across major regions, including Baltimore, London, and Hong Kong, to provide hands‑on, role‑specific guidance. These specialists help investors identify the right tools, experiment safely, and share best practices across the platform.

In parallel, we are expanding the range of AI tools and capabilities available to our investment teams, testing multiple platforms and integrating them directly into investor workflows—from research and content analysis to financial modeling and coding. This includes a growing focus on agentic applications, where AI supports multi‑step processes such as portfolio analysis and trading insights. Importantly, development is increasingly led by teams closest to the investment process, combining domain expertise with scalable technology to build solutions that are both practical and high impact.

A central priority is integrating decades of proprietary research, investment frameworks, and institutional knowledge into AI‑enabled tools. While many AI technologies are becoming widely available, differentiation comes from how they are applied—combining advanced technology with deep domain expertise, disciplined processes, and a long‑term investment mindset. Our ambition is to create the most AI‑forward investment team in the industry, using evolving technology to reinforce, rather than replace, core strengths in research and judgment.

Our AI journey began long before the hype

Objective: Embed AI across investment teams so it becomes part of how research is conducted.

With underlying capabilities in place, we moved quickly when LLMs launched in late 2022. We then shifted from building everything internally to integrating best-in-class tools into investor workflows.

Outcome: Reduced manual information gathering: more space for iteration, learning, and idea development.

How AI maps to the investment process

But beyond any single firm’s strategy, the practical question is how AI is functioning day to day across asset management. One way to think about it is to break down AI evolution into three distinct waves: Predictive, Generative, and Agentic.

Predictive AI is the most mature. The industry has been using models to forecast outcomes, identify patterns, and support investment decisions for years. Predictive AI is now foundational across many investment processes.

Generative AI is driving the most visible change. Its impact lies in synthesizing information at scale: summarizing earnings calls, digesting unstructured research, comparing disclosures across years or regions. These tools go well beyond novelty chatbots. They compress the time required to absorb large volumes of information. This is where the challenge to active managers begins.

For most research‑driven investors, step one is the assimilation and organization of huge amounts of data, understanding the competitive dynamics of an industry and the competitive advantages of a business. Step two is using this information, in combination with sector expertise and insight, to model forecasted financial outcomes. True insight occurs when differentiated views can be obtained that vary from consensus with a higher‑than‑average degree of certainty. The compression of research cycles and information half‑life is a challenge that we, as an industry, are striving to meet.

For step one, given that the same tools are available to all market participants, it is difficult to argue that generative AI is anything other than a leveler that raises the baseline across the industry.

As for step two, newer tools are increasingly capable of building and maintaining detailed financial models, including working directly in Excel. However, left unguided, they can still default to consensus views rather than challenge them. Speed amplifies the cost of error; without clearly defined data sources, audit trails, and verification, AI can introduce inaccuracies or reinforce existing assumptions. Human oversight therefore remains essential.

Given the pace of development, however, it is reasonable to expect that tools will soon emerge that can build and maintain models that are both detailed and accurate. Even then, the real advantage lies in applying deep industry expertise to identify where a truly differentiated insight exists. In practice, that often comes from proprietary research—analyst insights, company meetings, fieldwork, and industry conversations that are not captured in public data. It is the interpretation of this differentiated information, rather than access to information itself, that ultimately drives edge. At this point, we believe this remains a human capability.

The third wave, Agentic AI, is at an earlier stage. Rather than assisting with tasks, “agents” execute multi‑step workflows autonomously. In fund management this extends to both analytical and decision‑making functions. Data ingestion, maintenance of newsflow, financial modeling, scenario setting, and assumption testing are tasks well suited to such systems.

In portfolio management the actions of buy, sell, position sizing, and portfolio construction as a means of risk control are also increasingly capable of being driven by agents—if not fully automated, then used as co‑portfolio managers and sounding boards. If, therefore, generative AI democratizes rapid information assimilation and Agentic AI automates the portfolio management process, what remains as durable edge for an active investment shop?

Where EQ meets IQ

Efficient market hypothesis suggests that everything knowable about a security is reflected in its price. AI increases both the scope and speed of what is knowable. In practice, investors must deal with intangible factors that may or may not be knowable, but are certainly not measurable, while market behavior reflects human emotion and shifting regimes that are not easily modeled.

For AI, these create blind spots. These can be grouped into three broad areas: regime change—particularly at moments of market stress, when historic correlations break down; areas that require qualitative judgment, such as assessing corporate culture, political risk, or technological change; and unconventional time horizons that extend beyond most datasets or precede when those intangibles are reflected in the numbers. LLMs still struggle with ambiguity and nuance—this is where knowledge, experience, and human judgment remain critical.

A case in point was the recent credit round by the issuer First Brands. Running the numbers from prospectus and publicly available information through an LLM showed nothing out of the ordinary and that the balance sheet was sound. However, our sector specialization meant that the relevant portfolio manager had history with the company. Triangulating that knowledge with industry contacts led to the suspicion there was off‑balance‑sheet financing. We could not know the extent of the financial leverage, but the point is that it did not pass the “smell test” and we passed up the deal. That experience, built across cycles, is difficult to train or infer. Judgment becomes more central, not less.

That is not to say that human judgment is infallible. When the model cannot rely on a mathematical answer, it has to deduce a solution from its training data. And the training data are mostly derived from human trading records and thus codify the same biases as humans. Investors are subject to cognitive biases—anchoring, confirmation bias, recency bias, overconfidence, and loss aversion—that can distort outcomes. Garbage in, garbage out.

The best investors are distinguished not by the absence of bias, but by their discipline in mitigating it. AI can help counter behavioral biases by challenging assumptions and surfacing contradictory evidence. Properly designed, systems can challenge assumptions, surface contradictory evidence, and enforce analytical consistency. In this sense, AI does not merely support judgment; it can help refine it. If models can be trained to understand an individual investor’s behavioral biases and call them out pre‑trade, they could reduce the influence of emotion—combining the best of human judgment and machine learning.

AI’s fundamental blind spots

Markets often involve intangibles that may not be knowable or measurable, challenging efficiency assumptions. Notable blind spots for AI include:

• Regime change: particularly at market stress moments where correlations break down and leadership shifts unpredictably
• Qualitative judgment: areas requiring assessment of corporate culture, political risk, or technological change where models struggle with ambiguity
• Extended time horizons: extending backward beyond datasets or forward before intangibles are reflected in numbers

Organization and culture matter

Part of how humans form judgment is through the thinking time between data points, and their ability to apply intangible knowledge to them. That is not replaced by AI doing some of the legwork, any more than open‑source software has ever replaced enterprise software over the past 30 years. The important point is to ensure people continue to develop judgment, because arguably the legwork is how we create the thinking time to form it. At T. Rowe Price, we pride ourselves on our programs of mentorship and training, and on providing the tenure, time, and space needed to learn from mistakes.

We noticed that at the onset of the recent Middle Eastern war, all the LLMs were recommending the same trading strategy: that the market would be marked down and that investors should fade the sell‑off (i.e., buy the dip). This did, initially, reflect market conditions as many investment decisions are increasingly informed or executed by AI systems. Here, the models were exhibiting recency bias, having been trained on recent geopolitical events that reinforced this strategy but without the benefit of longer historical datasets. As a result, they exhibited groupthink.

At T. Rowe Price, we have a policy of “no house view.” In a world of AI, where assimilation of data is being accelerated and commoditized, cognitive diversity matters. It is not the information itself that matters, but the way it is interpreted, challenged, developed, and ultimately implemented. The danger is not that AI fails, but that it succeeds too uniformly across competitors, creating crowded trades and systemic fragility.

While consumer generative AI such as ChatGPT remains free, establishing workable systems to enhance the investment process requires real investment. Scale matters because most of the economic, technical, and organizational benefits of AI in asset management are convex: fixed costs are high, marginal costs are low, and many advantages only show up once AI is deployed widely across the platform. We have described the age of AI as analogous to electricity—something that will become ubiquitous and pervasive. AI will function best when embedded across processes as an operating system, rather than as a separate dashboard.

In the age of AI, client trust will matter even more. As fiduciaries and custodians of client capital, we remain responsible for investment risk. AI systems must be interrogable and explainable, the independence of risk oversight and functions must be preserved, and clear boundaries between full automation and human sign‑off must be established.

In the end, the age of AI will not abolish active management so much as clarify what it really is. The tools will converge, the information edge will compress, and anything that can be automated will be—but judgment, context, and culture will not. The durable advantage will sit with those investors who can marry systematized, scalable intelligence with genuinely independent thinking, and who can use machines both to see more and to see their own biases more clearly.

For firms, that requires investment not only in models and data, but in mentorship, cognitive diversity, and organizational design that resists herd behavior even when the herd is algorithmic. For clients, it raises the premium on fiduciaries who can explain how decisions are made, take responsibility when they are wrong, and adapt as regimes change in ways no back test can fully anticipate. The task is to embed AI, to build an investment organization that is better because of it—more curious, more disciplined, and more resilient when future paradigms arrive and the patterns in the data suddenly break.

There is no room for complacency. The age of AI is moving so fast that today’s blind spots could narrow, shifting the frontier of truly human alpha yet again.