Agentic AI in Investment Decisions: A Partner, Not a Shortcut
By Kevin Bauer
19 November 2026
Investors are increasingly using AI systems to gather information and inform their decisions. Investment firms already offer connections to AI agents that can act independently within defined limits. This is making sophisticated analysis more widely accessible. But broader access to analysis does not automatically democratise sound investment judgement, argues Kevin Bauer.
Imagine two retail investors asking an AI system, such as ChatGPT or Claude, the same question: “I have €10,000 available. Should I invest in Company X?” One receives a cautious assessment focused on valuation risk. The other gets a more positive assessment highlighting growth prospects and recent earnings. Both perspectives may be relevant. What calls for explanation is when the same information and investment objectives lead to opposing recommendations without a clear reason.
Generative AI systems, which can produce different answers to the same question, are becoming an established part of how people process financial information. A recent study of more than 2,000 retail investors found that 47% of respondents reported using generative AI to process financial information or support investment decisions. An analysis of more than 400,000 queries submitted to a brokerage’s AI chatbot also shows that investors already use these systems to interpret market developments, screen stocks and streamline their research.1
Increasingly, however, AI does more than answer individual questions. More advanced systems can go considerably further. An AI system can decide what information it needs, retrieve financial reports, compare competitors, check recent news, calculate valuation metrics, reassess its initial view and ultimately make a recommendation. Such systems are described as agentic AI: they can pursue a specified goal over multiple steps, taking actions, evaluating feedback and adjusting their approach accordingly.2 This discretion creates additional scope for analyses to follow different paths. At the same time, predefined data sources, calculations and verification procedures can help make results more consistent.
From Answering Questions to Pursuing Goals
A basic AI chatbot starts by producing an answer to a request. Ask a large language model, or LLM, to summarise a company report, explain the price-to-earnings ratio or list the risks associated with a stock, and it generates a response. An LLM-based AI agent also uses generative AI, but it can be assigned a broader goal, decide within specified limits how to pursue it and take the necessary actions. Research on LLM-based agents examines systems that combine language models with planning capabilities, memory, external tools and the ability to interact with digital environments.3 An investment agent might, for example, gather the latest company reports, compare profit margins with those of competitors, search for relevant news, perform calculations and revise its analysis as new evidence emerges. Approaches such as ReAct, which connects reasoning with action, illustrate this process. Research on tool use also shows that an LLM can learn when and how to call external services.4
Through software interfaces connected to a brokerage account, agents can potentially also place orders, monitor positions or adjust a portfolio. These capabilities are already being offered in practice. In 2026, brokerage platforms introduced various forms of agentic trading and AI connections to investment accounts.5
The degree of autonomy can vary considerably. Agents need appropriate permissions; in some configurations, individual trades still require human authorisation. Others allow agents to act within predefined permission limits or through separate accounts or account partitions.
This article focuses on a particular arrangement: the agent has substantial discretion over how it analyses an investment, but a human retains the final investment decision.
Why Your AI’s Analysis May Differ from Mine
It is tempting to treat an AI-generated investment analysis like the output of a calculator, where the same input always produces the same result. Large language models and the agentic systems built around them work differently.
First, response generation can be stochastic. In other words, repeated requests can produce different results even when the prompt remains unchanged. Methodological research shows that such variation can arise, among other things, from probabilistic sampling during text generation. Model updates and features of the underlying computing infrastructure can also change the results.6
Second, differences in how users describe their goals can matter. A 2026 study of AI financial advice found that recommendations varied systematically with users’ prompts, financial literacy, previous experience with AI and personal characteristics.7
Third, AI systems can be personalised. Research on personalised language models examines how systems can infer explicit and implicit user preferences from earlier interactions, take them into account within a conversation and refer back to them when needed. It also shows that reliably incorporating such preferences remains a challenge.8 In investing, one user’s AI agent might therefore prioritise capital preservation, while another’s places greater emphasis on growing wealth. These differences can be appropriate when they reflect real differences in investment objectives and risk preferences. They become problematic when they rest on inaccurate or outdated assumptions about the user.
Finally, the model itself is only one component of the system. The surrounding harness—the technical environment that connects the AI model to tools, context and rules, and manages its execution—defines which tools the agent can access, how information is stored and retrieved, what instructions constrain it and which verification procedures are applied. It can also restrict the sources available to the agent.9 Different models, versions, information sources, stored memories, prompts and tool configurations can therefore produce different analytical paths and answers. Agentic systems can differ both in what they investigate and in the conclusions they draw.
Democratising Analysis Is Not the Same as Democratising Judgement
Generative AI systems and agents can substantially reduce the effort required to obtain and process financial information. Tasks that once involved working through lengthy annual reports, comparing multiple sources or deciphering technical terminology can increasingly be supported by AI.
Evidence from retail investors supports this interpretation. Research shows that investors particularly value generative AI for making information processing faster and easier.10 The UK’s Financial Conduct Authority likewise recognises that AI can summarise complex information, speed up research and make technical material easier to understand. 11
This raises the question of whether easier access to sophisticated analysis also makes sound investment decisions easier. Assessing an AI-generated investment thesis still requires someone to judge whether the relevant data are current, whether important information has been omitted, whether a comparison is appropriate, whether assumptions are reasonable and whether the conclusions follow from the available evidence. AI can help with these checks too. But a convincingly written analysis is not, in itself, evidence that it is reliable.
Expertise therefore remains important, particularly when evaluating the results. Yet even basic prior knowledge can matter from the moment a user frames a request. In the study by Choukhmane et al. (2026)12, prompts and the resulting recommendations differed with respondents’ financial literacy. Simulations based on those recommendations produced less favourable wealth outcomes for the lower-literacy group. At the same time, the study shows that AI advice can improve outcomes relative to respondents’ reported baseline behaviour. Its focus, however, is lifetime financial planning, rather than agents with persistent memory researching individual companies.
More broadly, research on human–AI collaboration shows that the benefits also depend on how well people understand where AI is useful and where its limits lie. In a large experiment involving management consultants, participants using AI achieved substantial productivity and quality gains on tasks within the tested system’s capabilities. On a task beyond those capabilities, participants using AI performed worse. Within the system’s capability frontier, participants with lower baseline performance benefited particularly strongly.13
The experiment used GPT-4 in 2023, so its findings cannot be applied directly to today’s investment agents. It nevertheless illustrates a fundamental challenge: people may find it difficult to recognise whether a task falls within or beyond the capabilities of the AI they are using. A fluent and persuasive answer alone does not reveal how reliable it is. Factually incorrect or unsupported statements, often referred to as “hallucinations”, can also sound convincing. Human judgement and expertise therefore remain important for critical evaluation. Retaining the final decision, however, does not by itself ensure effective oversight: the basis for a recommendation must also be understandable and open to scrutiny.
AI as a Partner, Not a Shortcut
A productive question for investors might therefore be: “Which parts of the AI’s investment analysis can I rely on, and which should I challenge or refine?” An agent can research widely, process information quickly, compare alternatives, identify contradictions and update an analysis. A human can define objectives, challenge assumptions, contribute contextual knowledge and judge whether the evidence is relevant to the intended goal.
For this collaboration to work, the sources used and the dates of the information should be clear. Facts, assumptions and conclusions should be distinguishable; counterarguments and significant gaps in the data should also be visible. This makes meaningful scrutiny easier, rather than reducing the human decision to approval of a ready-made recommendation.
Seen this way, AI becomes a new kind of analytical partner. It can broaden an investor’s research and support their understanding of financial relationships, while uncertainty, market risk and the need for financial literacy remain. The Financial Conduct Authority recommends treating general-purpose AI as a starting point for further research and checking financial information against trusted sources.14 Securities regulators also warn investors about products claiming that AI can deliver guaranteed or risk-free returns.15
The next frontier is already here. An AI agent that analyses an investment, selects relevant information and formulates a recommendation can also, in principle, be authorised to execute trades, monitor outcomes and adjust a portfolio. Current brokerage offerings and trials show that parts of this transition are already under way.16 At the same time, regulatory bodies explicitly identify the risks that arise when agents act without human review and approval.17
A central challenge is how people can collaborate effectively with agents that analyse investments on their behalf. As agents gain greater discretion over autonomous investment decisions, understanding and directing their choices will become more demanding. It will be particularly important to understand how the personal information agents collect and retain about their users influences investment decisions. Persistent user profiles and stored memories may play a role in unexpected ways that users find difficult to detect.
Users therefore need ways to inspect and correct stored information and inferred preferences—both when these misrepresent their needs or circumstances and when those needs or circumstances change over time. They should also be able to understand which of these details influenced a recommendation. Without such mechanisms, agents could base investment decisions on inaccurate or outdated information and act against their users’ interests.
Suggested citation
Bauer,Kevin (2026): Agentic AI in Investment Decisions: A Partner, Not a Shortcut. In: eFin-Blog. https://zevedi.de/en/efinblog-agentic-ai-in-investment-decisions/ [19.11.2026].https://doi.org/10.83253/cvfj-e987
Sign up for the eFinBlog RSS-Feed Back to the eFinBlog Starting Page
More about the Project “eFin & Democracy – Democracy Issues of the Digitalised Financial Sector”
1 Blankespoor, E., Croom, J. and Grant, S.M. (2026) ‘Generative AI and investor processing of financial information’, Journal of Accounting and Economics, 82(2).2 Financial Industry Regulatory Authority (FINRA) (2026) 2026 FINRA Annual Regulatory Oversight Report: GenAI: Continuing and Emerging Trends.3 Zhou, C. et al. (2026) ‘Externalization in LLM agents: A unified review of memory, skills, protocols and harness engineering’, arXiv preprint, arXiv:2604.08224.4 Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K.R. and Cao, Y. (2023) ‘ReAct: Synergizing reasoning and acting in language models’, International Conference on Learning Representations (ICLR).5 Robinhood Markets, Inc. (2026) ‘Robinhood is now open to agents’, 27 May 2026; Interactive Brokers LLC (2026) ‘Interactive Brokers integrates AI into client portfolios – informed by agentic technology, controlled by the client’, press release, 1 June 2026; Scalable Capital (2026) ‘Scalable Capital becomes first bank in Europe to open platform to AI assistants via Agentic Investing’, 25 August 2026.6 Coqueret, G., Llull, J., Oswald, F., Pérignon, C., Scheuch, C. and Vilhuber, L. (2026) ‘Randomness in large language models: What researchers need to know (and report)’, arXiv preprint, arXiv:2607.24372.7 Choukhmane, T., de Silva, T., Lin, W. and Akuzawa, M. (2026) AI Financial Advice: Supply, Demand, and Life Cycle Implications. NBER Working Paper No. 35574. Cambridge, MA: National Bureau of Economic Research.8 Zhao, S., Hong, M., Liu, Y., Hazarika, D. and Lin, K. (2025) ‘Do LLMs recognize your preferences? Evaluating personalized preference following in LLMs’, International Conference on Learning Representations (ICLR).9 Zhou, C. et al. (2026) ‘Externalization in LLM agents: A unified review of memory, skills, protocols and harness engineering’, arXiv preprint, arXiv:2604.08224.10 Blankespoor, E., Croom, J. and Grant, S.M. (2026) ‘Generative AI and investor processing of financial information’, Journal of Accounting and Economics, 82(2).11 Financial Conduct Authority (2026) Using AI for Investment Research. InvestSmart. Last updated 30 July 2026.12 Choukhmane, T., de Silva, T., Lin, W. and Akuzawa, M. (2026) AI Financial Advice: Supply, Demand, and Life Cycle Implications. NBER Working Paper No. 35574. Cambridge, MA: National Bureau of Economic Research.13 Dell’Acqua, F., McFowland, E. III, Mollick, E., Lifshitz, H., Kellogg, K.C., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K.R. (2026) ‘Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality’, Organization Science, 37(2), pp. 403–423.14 Financial Conduct Authority (2026) Using AI for Investment Research. InvestSmart. Last updated 30 July 2026.15 Securities and Exchange Commission (SEC), North American Securities Administrators Association (NASAA) and Financial Industry Regulatory Authority (FINRA) (2024) ‘Artificial Intelligence (AI) and Investment Fraud: Investor Alert’, 25 January 2024.16 Robinhood Markets, Inc. (2026) ‘Robinhood is now open to agents’, 27 May 2026; Interactive Brokers LLC (2026) ‘Interactive Brokers integrates AI into client portfolios – informed by agentic technology, controlled by the client’, press release, 1 June 2026; Scalable Capital (2026) ‘Scalable Capital becomes first bank in Europe to open platform to AI assistants via Agentic Investing’, 25 August 2026.17 Financial Industry Regulatory Authority (FINRA) (2026) 2026 FINRA Annual Regulatory Oversight Report: GenAI: Continuing and Emerging Trends.
