A Practical Guide for Fundamental Managers Leveraging AI in Their Investment Process
TLDR: AI has moved from pilots to workflows in investment research, but the reliable applications are narrower than the prevailing sentiment suggests. Extraction, synthesis, and pattern detection over a fund's own research do deliver value today. But producing output that teams can trust when taking investment risk remains harder. EDS wrote this guide leaning on our decades long experience working with the world’s elite money managers and helping them on their journey to fully leverage AI in their investment management process. The guide maps the AI capability spectrum, explains why the line between language and calculation matters more than a product feature list, and suggests key questions managers can ask when evaluating financial research software in 2026.
The Disconnect Between AI-Powered Research Software Demos and Real Investment Workflows
Every fundamental manager has sat through several AI tool demos by now and most of them are impressive at first glance. There are several off-the-shelf tools, some built on frontier models like Anthropic's Claude that show promise with automations, research, and agentic workflows. During a demo you may have seen an AI assistant summarize an earnings call in seconds, search through filings and answer questions about a 10-K. Agents can also draft a decent thesis memo that a competent second-year analyst could have written. The demo is overall impressive but when real work needs to be done the disconnect becomes evident, and a realization sets in: the output is not really usable for investment workflows.
One example that you may have already seen working with foundational models is that a model can quote a revenue figure that appears nowhere in the filing. Sometimes they make numbers up just to come back with something plausible for the user. Compliance wants to know where the data came from and nobody can explain how the tool arrived at its answer. This happens when the model does not know or is not explicitly told where to get that data, so it gets creative. That becomes a bigger problem if the output is ever sent externally.
Another issue for investors is that no matter how good the model is, it needs data or at least access via connectors to datasets that you use to run your investment process. Some of this data, like portfolio risk and holdings data is not easily uploadable to the cloud and most tools do not come with connectors to your most critical data off the shelf.
Now, this does not mean AI has no place in investment research. It just means the market is crowded with tools that look and feel similar and most due diligence processes have no framework for telling them apart. This guide helps when choosing a tool by breaking down what AI does reliably today and how investment teams are leveraging it. We also review where it commonly breaks, and what to ask a vendor before putting any AI-powered financial research software in front of your investment team.
Five Categories of Capability for AI-Driven Investment Tools
AI investment research tools fall into different categories, from modest enhancements, to the chatbots most people already use, to ambitious autonomous AI agents. Investors have to be very careful with autonomous agents depending on how much data they have access to. The more powerful the tools, the higher the risk and the more stringent the guardrails should be.
Level 1: Search enhancement. Low risk. These tools offer advanced semantic search across notes, filings, transcripts, and emails. Instead of matching keywords, the system understands that “pricing pressure in the datacenter segment” and “hyperscaler capex pushback” describe the same concept. It then pulls related documents. This capability is now very mature, low-risk, and by now table stakes.
Level 2: Extraction and structuring. Pulling structured data out of unstructured documents: guidance changes from a transcript, KPIs from a filing, key claims from an expert call. This is where AI starts saving meaningful hours, and because every extracted field can be checked against the source document, errors are catchable. Every tech-savvy team should already be doing this.
Level 3: Synthesis grounded in data. Summaries and briefings assembled from a defined corpus, ideally your own research. A morning briefing across your coverage. A “what changed this week” memo for a position. A recap of every note your team has written on a name ahead of earnings. This works when the model is restricted to your documents and cites its sources, and fails when it free-associates from training data. Some models like Google’s notebook LM are very good at this, others require specific guardrails. As far as we see, the field is split. Not all managers are doing this in a scalable way across their most critical data.
Level 4: Analytical reasoning. Some risk. Natural language questions over real analytics that takes thought above a simple look up or summary: “What's driving my underperformance this week and how likely is this to continue?" or “Which positions carry the most factor risk and why?” Done right, the model interprets the question, understands it, then calls the right analytics engines, and narrates the verified result. This is the level where architecture matters most, as the model should not be inventing an answer or reasoning outside of the granted data stores. You need to be very careful about hallucinations and be explicit with what data the model leans on. EDS has recently released a solution for this, more on that later.
Level 5: Autonomous workflows. High risk. Agents that reason, form investment views, size positions, or trade. This is the most ambitious category, and the one where governance and track record are still being established and the few exceptions prove the rule: we are not there yet. The next section covers it more.
For a fundamental shop, practical value today concentrates in levels 2 through 4. Level 1 you should already have. Level 5 deserves the same care and consideration you'd apply to any black box system asking for discretion over the book. Most managers may never fully get there, no matter how capable the models get.
What Works Today for Investment Managers Using AI
Structured data extraction. The unglamorous workhorse of AI investment research is pulling and prepping data. In a survey of banks, investment firms, insurers, and hedge funds reported by The Fintech Times, two-thirds of respondents said their quants and analysts spend between a quarter and half of their time collecting, preparing, and quality-controlling data rather than analyzing it. Extraction addresses that sunk cost: transcripts arrive summarized against your thesis, guidance revisions are flagged the morning after the call, and expert calls are searchable, tagged tearsheets instead of the usual PDF. The productivity gains are quantifiable and compounding, which is exactly what you want from infrastructure.
Automated synthesis over your own research. A general-purpose model summarizing “what the market thinks” produces generic, consensus-level summaries, because it was trained on consensus and fit for a retail investor, not an institution. A model grounded in your research notes, internal models, and position history can answer a more valuable question: what do we believe, and has anything changed in light of our exposure?
Of course, there are many other use cases for AI in investment research and each fund is different in how they use it, their risk tolerance to uploading data, and their pain points that need to be automated most. However, we are still not at a point where AI agents can run portfolios themselves and nowhere near the point where LPs trust their capital to AI fully.
What Still Doesn't Work: Autonomous Stock Picking
Recently, Magnetar made news with a planned agentic AI fund, still pre-launch as of this writing in July 2026. Notably, even in that design, hundreds of AI agents handle the research while humans retain final authority over every trade. The answer to “can AI pick stocks as well as a seasoned PM?” is still no, and just as importantly, LPs have not yet shown they will trust their capital to an LLM model, no matter how advanced.
Large language models are probabilistic by design and generate the most plausible next token, which makes them superb at language but untrustworthy at consistently giving you the right numbers. A model that gets numbers right 95% of the time sounds great until you consider what a 5% error rate does to a report that an investor or regulator will read. It is important to note that because of their non-deterministic nature, the answers these models provide may never get to be consistently 100% accurate, but there are ways to guardrail and guide the model to be much more accurate than what you get by default.
Markets add another wrinkle: they are competitive, and information available to everyone is priced in quickly. Anyone can rent the same frontier model, so whatever signal a general-purpose model extracts from public information loses its value roughly as fast as it can be generated. Long before LLMs, quants and HFTs have figured out how to arbitrage away inefficiencies in milliseconds. Durable edge for a fundamental manager still comes from proprietary research, differentiated views versus consensus, and process discipline. AI can help amplify all three but it cannot substitute for any of them.
- "Durable edge for a fundamental manager still comes from proprietary research, differentiated views versus consensus, and process discipline. AI can help amplify all three but it cannot substitute for any of them."
Then there is accountability. A portfolio manager owes an explanation to the investment committee, to investors, and to regulators, and “the model said so” doesn’t cut it. This is why a sensible division of labor has emerged across the industry: the model handles synthesis, deterministic engines handle numbers, and humans retain judgment and accountability.
How to Evaluate AI Tools for a Research Workflow
The following are six questions that help separate financial research software fit for institutional workflows from a general-purpose AI tool.
1. Where do the numbers come from? The single most important question if you are considering relying on the tool for investment research. If the answer is “the model does its own research online or through a market data connector,” that should give you pause. If the model calls verified analytics engines and every figure carries a traceable computation path, you're looking at something you can more likely trust. This is the design principle behind EDS's Fusion AI, summed up internally as “the AI synthesizes, EDS calculates,” and it's the right test to apply to every vendor in the category.
2. What is the data the AI works with? It should be YOUR data. Your notes, models, estimates, and portfolio positions. The corpus of the open internet is a good fit for a retail investor researching a name for the first time, but it lacks the context a hedge fund needs. Test the system: when you ask about “YTD attribution,” does the tool map that to the correct period, portfolio, benchmark, and factor model automatically, or does it guess? Context resolution is where generic tools often fail, and no amount of prompt engineering by your analysts will fix it.
3. How does it handle entitlements? An analyst should not be able to prompt their way into positions they aren't permissioned to see. Access controls need to be inherited from the platform's architecture, not enforced by a system prompt asking the model nicely.
4. Can you audit the answer? Every output should cite its sources and log its tool calls, so any answer can be reconstructed after the fact. If you can't reproduce how a number was made, you can't use it for anything that matters.
5. Does it meet the institutional bar? SOC 2 Type II, zero-retention inference so your data never trains someone else's model, SSO/SAML, and human approval gates on any portfolio-affecting action. These are prerequisites, not differentiators.
6. Does it work where your team works? Analysts live in Excel, Outlook, and increasingly in AI assistants. Look for plug-ins into existing workflows and, more important each quarter, an MCP (Model Context Protocol) server, which lets AI agents like Claude or ChatGPT query your research and analytics under the same permissions and audit trail as the platform itself. MCP is fast becoming the standard interface between institutional data and AI agents, and a platform that cannot serve as a governed, permissioned context for AI agents is already behind, no matter how strong its other features are.
Moving From System of Record to System of Reasoning Unlocks More Value For Investors
Taking a step back we can see how far this industry has come in recent years. For two decades, the research platform was a system of record: the place where notes, models, and price targets were stored and occasionally retrieved. AI changes the value of that historical data by giving it a voice, and the ability to answer your questions on demand. Every thesis, revision, hit, miss, and post-mortem becomes the data foundation a model can reason over and give you an intelligent answer. Institutional memory that is codified compounds value.
Structured note-taking and centralized data used to be a tax paid for compliance and continuity and now these are the prerequisite for every AI capability described here. A fund whose research is spread across personal OneNote files and local desktop spreadsheets has much less data for AI to reason over. While a fund whose process is codified in one centralized, governed platform gets better answers from every new model release without changing a thing. That difference in positioning on AI means one shop will benefit from every frontier model release while the other barely notices anything different.
How EDS Approaches Using AI in Investment Management
Equity Data Science builds structure around the division of labor this guide argues for. Fusion AI, the intelligence layer on the EDS platform, answers natural language questions across research, risk, and performance, but every number comes from deterministic analytics engines with a traceable computation path rather than from the model itself "figuring it out”. This works in an institutional setting because outputs inherit each user's entitlements, carry citations and calculation lineage, and any portfolio-affecting action requires explicit human approval. The premise is that governed AI, grounded in a fund's own research and verified analytics, is the version institutional investors can put in front of an investment committee and a regulator.
EDS now also exposes its analytics through an MCP server, so funds can point Claude, ChatGPT, or their own internal agents at their investment system of record: morning briefings, risk alerts, and thesis-to-portfolio reviews built on proprietary data instead of the open internet. It's one implementation of the architecture described here, and worth evaluating against the six questions above.
See Fusion AI on your data →(link: https://equitydatascience.com)
FAQ
Can AI pick stocks on its own?
Not to an institutional standard yet. Large language models are probabilistic systems that generate plausible language, not verified calculations, and markets arbitrage away any signal available to everyone renting the same model. AI is effective at extraction, synthesis, and pattern detection within a research process. Position decisions belong with humans, supported by deterministic analytics. Some signs point to this as a possible development in the future as AI agents are connected to more data and better governance and guardrails are set up.
What is the best use of AI in investment research today?
Grounded synthesis over a fund's own research: briefings, thesis reviews, and “what changed” summaries with citations back to source documents. Conversations with your own portfolio need specialized systems like EDS’s Fusion AI. Close behind are structured extraction from transcripts, filings, and expert calls, and drift detection between internal views and consensus that you can get from tools off-the-shelf.
Are LLMs safe for hedge fund data?
They can be, with the right architecture: zero-retention inference so your data never trains the vendor's models, entitlement-aware access, SOC 2 Type II controls, and full audit logging. The unsafe version is analysts pasting fund data into consumer chatbots, which happens at any fund that hasn't provided a governed alternative.
What should I look for in financial research software with AI?
Deterministic analytics behind every number, grounding in your own research rather than the open web, inherited permissions, cited and auditable outputs, institutional security controls, and integration with the tools your team already uses, including an MCP server for AI agents.
What is an MCP server and why does it matter for investment research?
The Model Context Protocol is an open standard that lets AI agents query external systems. For a fund, an MCP server makes the research and analytics platform accessible to Claude, ChatGPT, or internal agents under the fund's own permissions and audit trail, so AI workflows run on proprietary data instead of public information.
