AI for Equity Research: Tools and Workflows for Analysts

AI for Equity Research: Tools and Workflows for Analysts

AI is now a practical part of the equity research workflow, not a future promise. Analysts use large language model (LLM) tools such as ChatGPT, Claude, Perplexity and NotebookLM to screen ideas faster, digest earnings calls in minutes, accelerate model maintenance and produce first drafts of research notes. The value is speed on the repetitive 70 per cent of the job; the judgement, the differentiated view and the accountability for accuracy remain entirely human.

This guide walks through the workflows that work in practice, tool-agnostic, and the limits and compliance cautions, including material non-public information (MNPI) risks, that every analyst and their compliance team should understand before adopting AI.

Where does AI actually help in equity research?

Four workflows deliver most of the benefit today: idea screening, earnings call and filings analysis, financial model upkeep, and drafting. Each works best when the analyst supplies the source material and a tightly framed prompt, rather than asking an AI to "know" facts from memory.

1. Screening and idea generation

LLMs are useful for structuring a screen, not running it. Use AI to:

• Translate an investment thesis into screening criteria ("companies with improving returns on capital, falling capex intensity and net cash") that you then run in your data platform.
• Summarise sector dynamics, value chains and competitor maps as a starting point for coverage initiation.
• Stress-test a thesis by asking the model to argue the bear case against your bull case, which surfaces blind spots quickly.

Search-grounded tools such as Perplexity help collect recent public news flow across a watchlist, but every claim should be traced to its underlying source before it enters your work.

2. Earnings call and filings analysis

This is the highest-value use case. Feed the transcript, press release and presentation into the tool and ask for:

• A structured summary: guidance changes, segment performance, pricing commentary, margin drivers, capital allocation.
• Tone and language shifts versus the prior quarter's transcript (paste both and ask for a comparison of management language on key topics).
• Every question asked in Q&A, with a one-line summary of management's answer and any non-answers flagged.
• Extraction of all quantitative guidance into a table you can check against your model.

NotebookLM-style tools, which ground answers strictly in documents you upload, are particularly suited to working across a stack of annual reports and transcripts because they cite the exact passage behind each answer, which makes verification fast.

3. Model updates and data work

AI does not replace your spreadsheet, but it compresses the drudgery around it:

• Extract reported figures, KPIs and segment tables from PDFs into structured form for entry into your model, then reconcile the AI's extraction back to the source before relying on it.
• Draft the list of model changes implied by results ("revenue guidance raised 2%, tax rate guidance cut to 21%") as a checklist for your update.
• Generate or debug spreadsheet formulas and scripts for repetitive data transformations.
• Sense-check outputs: ask the model whether your implied margins, growth rates or multiples look internally consistent, treating its answer as a prompt for your own review rather than validation.

4. Drafting notes and client communication

LLMs produce competent first drafts of results comments, initiation sections and morning meeting bullets when given your model output, your view and your house style. The discipline that matters:

• Provide the facts and the conclusion; let the AI provide structure and prose. Never let it supply numbers from memory.
• Keep your differentiated view front and centre. AI-drafted research converges on consensus phrasing, which is precisely what clients do not pay for.
• Fact-check every figure, date and attribution in the draft against your sources before publication. You sign the note; the model does not.

What are the limits of AI in equity research?

• Hallucination: LLMs generate plausible but false figures, quotes and citations, especially when asked about specific companies without source documents. Grounding in uploaded documents reduces but does not eliminate this.
• Stale knowledge: a model's training data lags reality. Anything time-sensitive must come from live sources you provide or a search-grounded tool, then verified.
• Numerical weakness: LLMs are unreliable at multi-step arithmetic and valuation maths. Calculations belong in your model; the AI explains or checks logic at best.
• Consensus bias: models reproduce the average of their training data. They will not hand you an out-of-consensus insight; they can only help you build and test one.
• Confidentiality of inputs: consumer-tier AI tools may retain or train on what you type. Assume anything pasted into an unapproved tool has left your control.

What are the compliance and MNPI risks?

Compliance considerations should be settled before workflows are adopted, not after. The key issues for analysts in the UK and Ireland:

• MNPI and inside information: never input material non-public information, or anything that could constitute inside information under the Market Abuse Regulation, into an external AI tool. Uploading it to a third-party service is an unauthorised disclosure risk in itself, regardless of what the model does with it. This includes draft research with unpublished recommendations or target prices, deal knowledge, and non-public issuer communications.
• Approved tools only: use enterprise deployments with contractual commitments on data retention, no-training clauses and audit logging, approved by your firm. Shadow use of personal accounts is a common and serious policy breach.
• Research integrity: regulated investment research carries analyst certification and supervisory approval requirements. AI-assisted drafting does not dilute the analyst's responsibility for substance, fairness and the basis of recommendations, and firms should document where AI was used in the production process.
• Records and supervision: retain prompts and outputs where they form part of the research production record, and ensure supervisory review covers AI-assisted content like any other.
• Vendor and copyright issues: respect licence terms when feeding paid data, broker research or paywalled content into AI tools; redistribution restrictions usually apply.

Which AI tool should analysts use for what?

The major tools have overlapping but distinct strengths, and most analysts end up using more than one:

• ChatGPT and Claude: strong general-purpose assistants for summarisation, drafting, thesis stress-testing and working through uploaded documents. Claude's large context window suits long filings and multi-quarter transcript comparisons; both offer enterprise tiers with no-training commitments that compliance teams can approve.
• Perplexity: search-grounded answers with citations, useful for rapid news sweeps across a coverage list and for collecting recent public information on a company or theme. Always click through to the cited source before relying on a claim.
• NotebookLM: document-grounded by design, answering only from sources you upload and citing passages, which makes it well suited to deep work across a defined document set such as five years of annual reports.
• Platform-embedded AI: the established market data platforms are embedding AI assistants over their own curated data, which reduces hallucination risk for factual queries and sits inside infrastructure your firm has already approved.

A sensible mental model: use search-grounded tools to find information, document-grounded tools to interrogate it, and general assistants to structure, compare and draft, with your own model and sources as the single point of truth for every number.

A worked example: results day in 30 minutes

On results morning, a practical AI-assisted sequence looks like this. Drop the press release and presentation into your approved assistant and request a guidance-versus-consensus table and segment summary (five minutes). While the call runs, capture the transcript; afterwards, ask for a Q&A digest and a language comparison against last quarter (ten minutes). Update the model yourself from the source documents, using the AI's extracted figures only as a cross-check (ten minutes). Finally, give the assistant your updated numbers, your view and your house template, and edit its first draft into your results comment (five minutes plus your editing). The analyst's judgement enters at every step; the AI removes the typing.

How should an analyst start?

Begin with one low-risk, high-volume workflow, typically earnings transcript summarisation on public documents, using a firm-approved tool. Build a small library of tested prompts, verify outputs rigorously for the first month, and measure the time saved. Expand to drafting and data extraction once trust and controls are established. Analysts who develop these skills deliberately, including through structured CPD on AI for finance professionals, will simply cover more ground per week than those who do not.

Study with Learnsignal

Learnsignal's AI for Finance CPD courses teach accountants and finance professionals how to use AI tools effectively and responsibly, from prompting fundamentals to governance and compliance. Build practical AI skills for your research and analysis work at Learnsignal CPD.