AI is coming for FinanceSimone DaneseFollow6 min readJust now--Salutations,Artificial Intelligence has been all the rage for the past couple of years, but there seem to be significant disagreements in terms of its prospects and its potential impact on economic output.Critics argue that productivity has so far been minimally affected, and legal or economic bottlenecks will significantly slow down adoption even if the economic value were clear. Others complain about soaring capital expenditures even in the absence of a so-called killer application that could produce sufficient revenue streams to justify the continuous investment.The purpose of this article is not to directly debate these claims or support a hypothetical first killer application (which many now believe will be software engineering). Instead, Ill try to make the case that one of the most powerful upcoming points of application for AI will be the financial sector.Note that for the sake of being concise, Ill assume that you are already familiar with Large Language Models, how they work, as well as their current capabilities and limitations. There are plenty of freely available resources that will help you understand these topics with higher accuracy and precision than I ever could. (If you struggle to find something specific, write in the comments, and Ill link some relevant material.)With that being said, let me focus on the reasons why I believe the financial sector is strongly placed to absorb and benefit from AI capabilities.#1 Finance is almost entirely digitalFinance is fundamentally about gathering information, analysing it and making decisions about capital allocation.All of this is done almost entirely on digital infrastructure: the vast majority of information is on computers, analysis happens on computers and even the infrastructure to exchange resources (e.g. payment systems) has been almost entirely digitalized. In other words, the work of a modern finance worker is (almost) fully done in front of a monitor.Accordingly, once AI agents achieve a good level of performance on general computer use, there is no physical barrier preventing their deployment in the financial sector. There is no need for robotic arms, sensors, or any other form of physical embodiment.This might seem like a minor and obvious consideration, but it is actually very relevant: there is no reasonable argument for real economy bottlenecks in the use of AI in finance.#2 Finance can handle a rapid scale-upLets say you believe, as I do, that we can shortly achieve a satisfactory level of agentic performance. Then, a team of AI agents might realistically perform most of the tasks of a junior financial analyst.Whats the difference?AI agents can work 24 hours a day, 7 days a weekAI agents are much faster at processing dataAI agents can process much more data at onceYou can relatively seamlessly increase the number of AI agents in your team (aka hiring AI agents is much easier than hiring humans)This can lead to an explosion in the scale of analytical work performed.The financial sector is particularly well-suited to a fast scale-up:Its a rich industry with lots of liquidity to invest.Returns to capital invested are much faster than in other industries (precisely because many assets are uniquely liquid). In other words, the cycle from initial investment to profitability to reinvestment is much shorter than in other sectors, promoting rapid scaling. This would be further enhanced by a boom in market liquidity, reasonably driven by an increase in analytical capabilities.Financial decision-making can be considered a general-purpose technology, applicable across all sectors of the economy. Accordingly, an upgrade to this technology can be easily deployed on a very large scale. (As opposed to AI golf clubs, for which there is a bottleneck on market demand).#3 Finance is VerifiableWithin the realm of what is currently considered financial performance aka monetary risk-adjusted real returns performance in finance is easily verifiable.In other words, there is a clear and objective right answer to a financial problem. This is incredibly valuable for AI because it allows for iterative self-improvement via Reinforcement Learning without the need for human feedback. This, in turn, can dramatically speed up performance improvements.Now, there is the important caveat that financial performance is dependent on a number of factors that themselves are not easily verifiable (e.g., consumer taste). Nevertheless, the fact that there is a top-line performance metric that is verifiable means that the model is still able to learn on its own and discriminate which underlying factors to focus on.#4 Humans are not good at FinanceBelieve it or not, financial decision-making by which I mean mathematically-optimised rational unbiased decision-making on intertemporal capital allocation is something humans are not that good at. Were just not built for it.Were built for hunting, gathering, and socialising. Were built for arts and politics, not for analyzing endless spreadsheets. Our ability for future thinking and precise measurement of value is strongly constrained by our biological hardware and riddled with heuristics and biases.Finance is one of those fields where emotions only make things worse, and our brains are awfully adapted to make rational decisions by looking at a screen full of numbers when the stakes are so high.But guess what? Thats not the case for AI.AI agents dont get tired. They dont get emotional. They dont get distracted. Humans are constrained by physical, social and emotional needs that AIs are not.They arguably dont suffer from confirmation bias, scope insensitivity, or overconfidence. They can be rational in all circumstances. Similarly, they could be able to cooperate much more efficiently than a team of humans would, devoid of personal agendas, jealousies, conflicts, or miscommunication.Im saying arguably and could because currently there is little understanding of the inner workings of these models. Interpretability research is awfully underfunded and lagging compared to performance research (to the despair of AI safety experts). It is indeed unreasonable to assume that these models dont have their own biases (they probably do).But at the same time, we know that they dont think like humans, therefore, it seems safe to assume that they wouldnt be subject to the exact same cognitive biases and limitations as humans. More importantly, we can structurally modify and improve an artificial mind to correct biases when we find them. This, as far as I know, its impossible for humans.LimitationsFor all my evident optimism and confidence in the potential for AI to revolutionise finance, it must be noted that current applications dont quite meet the bar required for high-stakes financial decision-making.Models hallucinate (make stuff up), AI agents are still unreliable, and fine-tuning a model for a specific domain is still very data-inefficient and prone to errors. There are still considerable privacy, security, and legal liability concerns that have not been fully addressed by either the private or public sector.But finance is all about the future, and it would be unwise not to consider the trajectory that the AI industry seems to be on.The progress of this technology has been nothing but astounding in the past few years, and industry leaders have repeatedly expressed their confidence that this progress will continue (and possibly accelerate).I personally believe the industry is at a point of no return in terms of the scale of investment and public attention it has received. With hundreds of billions of dollars in investment and thousands of the smartest minds working on these problems, the sector has unprecedented resources and incentives to keep making progress at a tremendous speed.I wouldnt bet against them.The Dream of FinanceFinally, a word on what the consequences could be for the wider economy.To begin with, I think the value created will be potentially enormous.We are used to considering technology as the primary driver of economic growth (see Solow Growth Model), but we often forget that technology in economics does not mean gadgets but production processes. Namely, your ability to transform inputs into more valuable outputs.Technology then defines the upper bound of what is possible (production-possibility frontier), but in the real economy, most production processes are not on the efficient frontier (mostly because of asymmetric information and capital misallocation).An improved financial system will be able to better identify the best companies and allocate capital to them. This, in turn, leads to a higher economic surplus, which can be re-invested and start a virtuous cycle of faster progress and economic development.This is the dream of finance, this is why I love finance and why I dont like how it has come to be viewed as a leech on the real economy (and domain of finance bros).In fact, as financial savviness and rationality become pervasive via AI agents on both the demand and the supply side, we can (hopefully) expect the financial sector to cut the bulls**t: less questionable accounting decisions, less needlessly complex financial products, less creative revenue projections and for the love of dogs less trading gurus.Or at least I hope so.Cheers, Simone out