.jpg)
The practice of producing investment research on listed stocks has evaded real change over the last 30-40 years. I am an ex-II-ranked analyst who grew up in the world of global investment research in the mid-1990s. In my time, an average sell-side analyst (at a global investment bank) could cover no more than 12-15 stocks at a time. This has been a major business model constraint for decades, further compounded by continuously declining payouts from the client base for written content.
I believe Gen AI, and its various configurations, could potentially disrupt existing practices to boost productivity, analytical insights, and rating conviction across analysts. I make this statement after having done in-depth work on AI and how best it can be leveraged, intelligently and securely, while keeping the “Analyst” as the valuable differentiator.
Bottom-up fundamental research is a craft of synthesis: regulatory filings (10-Ks, 10-Qs, other releases); channel checks; reference to a line in the last quarterly release notes/call; a financial model that reflects these findings. You reflect on reported financials, your expectations from past publications, valuation and stock performance data, and external factors that impact your existing outlook. All of the above and more, intertwined into an analytical web to produce your final insight and recommendations.
Simply put – leveraging an LLM to give you an earnings preview note or earnings commentary note will produce vastly different results for two different people conducting the research via different prompts or accessing different models. Now add your own proprietary data set and analytical insights to this mix, and you have a whole new set of dimensions to this complexity.
Welcome to the world of Agentic RAG engineering and its impact on your world! Using RAG in your analytical framework allows you to leverage LLMs while controlling hallucinations and biases with a human-in-the-loop construct. The latter ensures your organisation remains firmly in intellectual control of your offering.
Here you are! You now have a library of agents that work together in a defined manner to aid your research process, ensuring your differentiation and ability to protect your unique insights.
What is vital is the orchestration layer. It is a choice to ensure your unique differentiation and analytical throughput. For example, you can leverage LangGraph to create your orchestration layer. The orchestration layer picks the relevant agent(s), the source of the information, your RAG dataset, your model or financial repository, your designated LLMs, news feeds, and other primary sources.
Your army of Agents (as part of your orchestration layer) can run within the Microsoft Agent Framework on Azure Foundry, or Amazon Bedrock. As a technology solution provider, we wish to be agnostic between Azure Foundry and Bedrock, and will always keep our agent contracts, prompts and evaluations portable, so we can swap models or frameworks (based on our client’s choice) without the need to rebuild the desk.
The governance and data flow controls ensuring no proprietary data ever hits any public servers or repositories, nor do they do into “training” any third-party models or platforms.
Retrieval-Augmented Generation (RAG) grounds answers in your documents rather than model memory. Your proprietary content: your universe of financial estimates and forecasts, your models, your past publications, your meeting notes, corporate presentations, and other primary sources.
The next question that needs answering – Is there a need to convert the entire RAG dataset into a vector database? Not really. Only unstructured text – published notes, earnings reviews and previews, transcripts, news feeds – where semantic search helps needs to be vectorised. For example, for a speciality chemicals company, an Agent could seek a comment on the company’s pricing power in the market based on past assessment, and then combine that with contemporary newsflow and transcripts to see if that remains relevant.
Your financial data set could remain in a SQL server and maintain its relational qualities.
The three primary RAG design choices:
Given our business context, as captured in the heading of this article, hybrid search is the right way forward. Combine keyword and vector retrieval, because tickers, segment names and exact figures are where pure semantic search fails. The financial database of estimates history should be indexed and versioned rather than vectorised.
Make retrieval point-in-time: an agent revising estimates should know which model version and which consensus were current on the date it is reasoning about, not just the latest file.
In practice, you combined the above relational database with a vectorised database of your own organization’s archive:
On the above framework, agents can ask how the “house view” evolved on margins; or what went wrong the last time a forecast was missed. External newsflow sits on top, with live news, filings, transcripts, and corporate events routed to the relevant coverage.
The research agent flags what changes the outlook; the analyst decides whether it changes the forecast!
Analysts forgive slow; clients don't forgive wrong. The rule is simple: no figure enters a draft report unless it traces to a model cell (or a financial repository) or a cited passage. In practice, the Writing agent pulls numbers from the model (or, once uploaded to a financial repository its downloaded in a report template) rather than generating them. Every claim carries a citation the analyst can click through. A reconciliation step confirms the note, the model and the valuation agree before anything moves forward.
When sources disagree (i.e., the transcript says guidance was raised, the model still shows the old range), the system should surface the conflict, not quietly resolve it. Forecast logic stays with the analyst. Agents can propose a driver change and show its sensitivity, but assumptions remain a judgement call of the analyst. This ensures the human-in-the-loop framework to drive the final assessment.
Test the system the way you would test a junior analyst. Build a golden set of past earnings updates with known answers, and score every release on citation accuracy, numeric agreement with the model and conflicts caught. A framework that cannot show those scores has not earned a place in the publishing workflow.
In the world of regulated investment research, accountability cannot be delegated to software. The framework must end in a hard approval gate, not an optional review. Before publication, the analyst works off one view: the report draft, model changes or his/her updated numbers, valuation bridge, every citation and every compliance flag.
The gate should scale with the stakes. A routine flash note may need only the author's approval. A rating change, target price move or initiation should also require a second reviewer, typically a supervisory analyst or compliance. All of these capabilities are part of the intelligent workflow approval management feature set.
Analysts can edit, reject or approve. Nothing is distributed without that sign-off, and the decision is logged alongside the agent trail that produced the draft. That audit trail also drives improvement: every analyst correction is a labelled example of where the agents fell short. Such deltas or changes in a loop could be designed to improve Agentic output in an iterative sense.
The goal is not to remove the analyst from the loop. It is to remove the drudgery, so their time goes where it earns alpha: forming the view!
At ANALEC, we bring over 24 years of expertise and customer experience to the investment research workspace. We are building the Investment Research platform for the future to empower sell-side and buy-side research organisations to think differently about their businesses. We are treating governance, entitlements, and human approval as the foundation rather than an afterthought.
If you would like to have a discussion around how we see the world of AI being enabled in your research organisation, do reach out to us!
.png)
