HowtoruninvestmentresearchwithAI

AI does not do investment research for you. It removes a large part of the finding, reading and assembling, then leaves the judgement with the analyst. A useful workflow is simple: frame the question, give the agent reliable sources, draft from evidence, check every important claim, and make the final call yourself.

KEY TAKEAWAYS

•  AI is strongest at finding, reading and assembling evidence; the investment call still belongs to the analyst.

•  An answer can only be as reliable as the sources the agent was allowed to use.

•  Important claims should trace back to a passage, filing, transcript or dataset you can actually open.

•  The analyst’s role shifts away from collecting information and toward testing what the evidence means.

What AI changes in the research process, and what it does not

AI is most useful in investment research when it is given a clear job to do. Instead of asking it to “do the research,” it makes more sense to use it for the parts that usually take up time: finding the right documents, reading across different sources, pulling out relevant evidence, comparing information and putting together a first draft. With the right sources behind it, an agent can take a lot of this work off an analyst’s plate.

The harder part is knowing what that information actually means. An agent can lay out the bull and bear case, but it does not have the same understanding of your portfolio, strategy or priorities. Those decisions still need an analyst’s judgement.

That is also why a polished AI answer should not automatically be treated as a finished piece of research. It may look complete while still missing context or important details. A better approach is to let AI handle the research groundwork and first draft, then have an analyst review the evidence, challenge the conclusions and decide what actually matters.

Step 1: frame the question before you search

Good research starts before the first document is opened. If the question is vague, the agent has to guess what evidence matters. Instead of “research this company,” define the decision you are trying to support and the period that matters.

A tighter question also gives you a stopping rule. You know which period, companies and evidence types count, so the agent has less room to wander into interesting but irrelevant material.

Step 2: give the agent sources that match the question

The sources you use should depend on what you are trying to find. If you need a reported number, the filing or company release is usually the best place to start. If you want to understand management’s thinking, the earnings call and investor materials give you more context. For recent developments around the company, you can bring in licensed news or other sources your firm is allowed to use. Internal research is useful too, especially for understanding previous views and assumptions, but it should add context rather than replace the original evidence.

• Filings and financial statements for reported figures and accounting details.

• Earnings transcripts for management commentary, guidance and Q&A.

• Company presentations and releases for current messaging and operating updates.

• Licensed news and research for relevant developments outside company materials.

• Internal research for previous views, assumptions and institutional context.

This is where the data behind the AI starts to matter. In a Bigdata.com test published in May 2026, the same prompts were run through Claude using web search and connected financial data. Keeping the model and prompts the same made it possible to see what changed when the underlying source of information changed.

The takeaway is simple: the model is only part of the research setup. Give the same model different evidence, and you can end up with a very different research output.

Step 3: draft from evidence, not from memory

Once the question and sources are fixed, the agent can do the assembly work. A practical sequence is to retrieve the relevant passages first, organise them by the research question, and only then draft the note. That keeps the prose downstream of the evidence instead of letting a fluent answer lead the research.

1.  Pull the evidence for each sub-question separately.

2.  Keep the reporting period, units and company identity attached to every figure.

3.  Ask for contradictions and missing information, not only supporting evidence.

4.  Draft the note with citations next to the claims they support.

5.  Separate reported facts, management statements and your own interpretation.

The same pattern works at different scales. A sector deep dive may gather comparable evidence across a peer set. A catalyst calendar turns dated company and regulatory events into a forward view. A quarterly review updates a fixed structure with new evidence. The output changes, but the workflow is the same: retrieve, organise, draft, verify.

Which parts of the research process does AI actually speed up?

AI tends to save the most time on the repetitive parts of research. An agent can go through several quarters of earnings calls looking for the same topic, compare changes in risk factors, track guidance updates or put together an initial timeline of events. Instead of doing each search manually, the analyst gets a useful starting point much faster. It can also make recurring research more consistent by following the same questions, fields and source rules each time.

Where AI helps less is with judgement. It can show you that management’s explanation has changed, for example, but deciding whether that change is important is a different question. The same goes for deciding whether management is being credible, whether a small update could become a bigger issue or whether new information is enough to rethink a position.

Those decisions depend on a wider understanding of the company, the portfolio and the investment strategy. AI can make the information easier to find and compare, but the analyst still has to decide what matters and what to do with it.

How do you check an AI-written research note before it goes out?

Treat the draft like work from a fast junior analyst: useful, but not self-validating. The review should be systematic enough that a polished sentence cannot slip through simply because it sounds right.

The core rule is simple: if a claim cannot be traced to evidence you can inspect, it should not carry the same weight as one that can.

Where does the analyst still have to make the call?

The analyst becomes most important when the evidence has to turn into a decision. An agent can show that guidance has changed, a competitor is saying something different, margins are under pressure or an important catalyst is coming up. But it cannot decide whether that information is enough to change the investment thesis, adjust the valuation, rethink the risk or change the timing of a position.

That is where human judgement still matters. The real value of AI is not in replacing the analyst, but in taking away some of the repetitive work that sits around the research process. If less time is spent searching through documents, collecting figures and comparing updates, more time can go into understanding what those findings actually mean.

The important questions are still the difficult ones: What really matters here? What information are we missing? What has the market already priced in? And what would need to happen for us to change our view? AI can help bring the evidence together, but those decisions still sit with the analyst.