Where AI actually helps an analyst
Financial analysis is a research-heavy discipline drowning in text: earnings transcripts, 10-Ks, 10-Qs, broker notes, regulatory filings, and management commentary. Large language models are good at exactly one half of the job — reading, compressing, and reformatting that text — and dangerous at the other half: producing numbers. The skill is knowing which side of the line each task sits on. Used well, an LLM turns a two-hour transcript slog into a ten-minute review; used carelessly, it puts a hallucinated EPS figure into your model.
How analysts use AI in practice
Earnings transcript summarisation. Paste a full earnings call transcript and ask for the three things that changed versus last quarter, every forward guidance statement, and any hedging language from management. The model surfaces tone shifts (“cautiously optimistic” → “monitoring closely”) that signal a change in confidence. You still read the raw quotes it pulls.
Filing Q&A with RAG. Instead of Ctrl-F across a 200-page 10-K, load the filing into a retrieval-augmented setup and ask plain questions: “What are the disclosed customer concentration risks?” or “How did deferred revenue change?” A good RAG pipeline returns the answer with the exact passage quoted, so you can verify in one click.
Narrative drafting. AI writes a competent first draft of the qualitative sections of a research note — business overview, risk paragraph, summary of the quarter — which you then sharpen with your actual thesis. It removes the blank-page tax, not the thinking.
Model augmentation. Ask for the Excel formula for a XIRR, a Python snippet to pull and clean a price series, or a sensitivity table structure. The model is strong at the syntax and structure of analysis; you supply and check the inputs.
Guardrails that keep you out of trouble
Never let an LLM be the source of a number that goes into a model or a published note — verify every figure against the primary filing. Never paste material non-public information or client data into a consumer chatbot; use an enterprise zero-retention tier or strip the detail. Treat any unsourced claim as a hallucination until proven. And keep an audit trail: which prompt produced which draft, and which figures you independently confirmed. The analyst’s name is on the report, so the analyst owns every digit in it.