In most fields, the better an agent works, the more its principal gains. Finance differs in four ways. First, the baseline for comparison should be the actual behaviour of individual investors, not the standard workflow of professionals; many people do no research at all, so an agent changes the depth of their research rather than the time it takes. Second, the market responds: the more people use the same kind of analysis, the less that analysis earns, and any edge in prediction shrinks as it spreads. Third, trading errors are hard to undo; a single mistake can cause enormous losses within minutes. Fourth, many agents use similar methods to reach similar judgments, and may act in the same direction at the same moment. The lasting value of a financial agent therefore lies not in prediction but in lowering costs, maintaining discipline, reducing behavioural bias, and making possible what people could not otherwise do.
Three kinds of delegation are common in finance: researching companies, placing trade orders and managing portfolios. Their amplification factors depend first of all on the baseline. Measured against the standard workflow of professionals, research and portfolio management both have high theoretical factors; measured against what individual investors actually do, the conclusion is entirely different. Laarits and Wurgler (2025) estimate that the median time an individual investor spends on research per trade is only about 6 minutes. A research agent mainly changes the depth of research, not the time spent. Placing an order takes only half a minute to begin with, so an agent can save little more than ten seconds. The value of a portfolio management agent is that people who would not otherwise allocate systematically can diversify and rebalance on a schedule.
The baseline decides how much can be saved
Three kinds of financial delegation. Switch the baseline: the standard workflow of professionals, or the actual behaviour of individual investors.
In other fields, when a technology lets everyone work better, everyone benefits. Investment research is different. Grossman and Stiglitz (1980) showed that information earns a return only because acquiring it is costly: the more people use the same kind of analysis, the faster the same information enters prices and the smaller the excess return each person can earn, until it barely covers the cost of the analysis. Once AI drives down the cost of analysis, more people will use the same kind of analysis, and its excess return will fall accordingly. The fundamental law of active management gives the other side of the same conclusion (Grinold, 1989). The most direct effect of AI is to widen breadth cheaply, covering thousands of stocks at once and reassessing them every day. But breadth raises the information ratio only by its square root: ten times as many independent judgments raise the information ratio only about threefold. And the information coefficient itself falls as the same kind of analysis spreads, offsetting even that gain. In a market close to efficient, the benefit of an agent that predicts more accurately is limited, and it shrinks as such agents spread.
Return to the decomposition of realized return: time compensation, risk compensation, information advantage, trading costs and luck. An agent cannot create time compensation, nor can it change luck; whatever it contributes to information advantage is consumed by competition. What can last lies in three other places. In risk compensation, it helps users stay diversified, rebalance as planned and not give up in the face of volatility. In trading costs, it chooses suitable order types and timing, reducing spreads and market impact. And between the return users actually earn and the return of the product, it narrows the gap caused by buying high and selling low. None of this depends on predicting better than others, so it does not disappear when everyone uses the same agent.
For many individual investors, the greatest value of an agent is not doing things faster but making feasible what they would not otherwise do. D'Acunto, Prabhala and Rossi (2019) studied a robo-adviser launched by an Indian brokerage. Investors whose holdings had been concentrated and poorly diversified became markedly more diversified after adopting it, and their volatility fell. Investors who had already been fairly diversified ended up holding fewer stocks; their volatility fell slightly, but they traded more often. After adoption, behavioural biases such as the disposition effect and trend chasing also declined. The value of the agent was concentrated among those who had been doing worst; for those who had been doing better, it could also bring new problems.
When an information task goes wrong, the error can usually be found and corrected, and the loss is limited. Trading is different: orders are filled within milliseconds, cannot be withdrawn once filled, and errors can be rapidly amplified by leverage and repeated execution. In August 2012, a deployment error in the trading system of a US market maker caused it to send millions of erroneous orders in about 45 minutes. It lost about $460 million and was subsequently acquired (U.S. Securities and Exchange Commission, 2013). Authority for financial agents must therefore be tiered: providing information and advice can be autonomous; changing holdings and moving funds require the person's own confirmation, together with per-transaction and cumulative limits and a switch that stops everything at once when something abnormal happens.
When many investors use similar models, similar data and similar rules, their trades will also be similar. Khandani and Lo (2011) analysed the "quant crisis" of August 2007: quantitative funds following similar strategies unwound their positions at the same time within a few days; their selling pushed prices down and forced still more funds to unwind. Kirilenko et al. (2017), studying the "flash crash" of May 2010, likewise showed that high-frequency traders passed positions back and forth among themselves during the fall, amplifying price swings. The spread of AI agents may amplify this sameness: each agent's rules may be reasonable, yet when all agents execute the same rule at the same time, the result can be bad for everyone.
The value of a financial agent is not sold on prediction; it lies in cost, discipline, behaviour and access. Execution needs tiered authority and limits, and must take into account the consequences of many agents acting at once.
A research agent's value lies in depth; a portfolio agent's value lies in letting people who do no allocation do it, not in saving a few minutes.
Place value in staying diversified, lowering costs and reducing behavioural bias, which do not disappear as similar agents spread.
Information and advice can be autonomous; changing holdings and moving funds need the person's confirmation, with per-transaction and cumulative limits.
Stagger execution times, vary rule parameters, and when markets swing violently, flag the risk and slow execution down.