How to Use AI as a Financial Analyst: What Works, What Does Not, and the Rules

A recommendation has a name on it. Whether that is a buy rating, a valuation in an investment memo, or the number in a model that a committee allocates capital against, someone is accountable for it being reasonable — and that someone is not the model. That single fact should decide where AI belongs in a financial analyst's work and where it does not, and it holds regardless of which tool gets faster next year.

Start with the obligations, because they narrow the options

This is not a hypothetical concern the profession has yet to address. CFA Institute's Standard V(A), Diligence and Reasonable Basis, requires that a recommendation rest on a reasonable and adequate basis, supported by appropriate research and investigation — a standard that predates generative AI but applies to it without modification. An AI-drafted output is a source you investigated, not a substitute for having investigated. CFA Institute has since published a dedicated framework on ethics and AI in investment management that applies the existing Code and Standards to exactly this question, rather than writing new rules from scratch.

Confidentiality does the same double duty. Standard III(E), Preservation of Confidentiality, covers client and firm information the same way whether it is emailed, printed or pasted into a chat window — a general consumer AI tool is not a confidential channel unless your firm has specifically vetted it as one. If you work at a broker-dealer, FINRA Regulatory Notice 24-09 is explicit that existing rules — including Rule 2210 on communications with the public — apply to generative AI output exactly as they apply to anything a person wrote by hand. Using AI does not lower the bar; it just changes where the checking has to happen.

Practically, that gives you three tiers of work: things fine to do with anonymised or public inputs in any tool, things that need a properly vetted tool with no training on your data, and things — anything touching material non-public information, in particular — you do not put into a general AI tool at all, full stop. Other professions bound by similarly specific standards have reached the same three-tier shape independently — how to use AI as an auditor covers how the PCAOB's updated evidence standards land on nearly identical ground.

What it is genuinely good at

The pattern across all of these: the model produces a draft or a first pass, and you supply the analysis and the number-checking that turns it into something you would put your name on.

1. 01Reading long documents against a question. A 10-K, an earnings call transcript, a credit agreement — "where does this say anything about covenant step-downs" is a search problem, and it is fast and generally reliable when the document is directly in front of the model rather than recalled from training.
2. 02Turning a finished model into a narrative. You have already built the sensitivity table; ask for a 150-word summary for the investment committee that leads with the variable that actually moves the outcome. You did the analysis. The model wrote the prose.
3. 03Comps and formula construction. Describing the screen you want in words and getting back an Excel formula or a Python snippet to pull it together — easy to verify, because you run it and check the result against a company you can compute by hand.
4. 04First-pass industry orientation. Getting up to speed on an unfamiliar sector's structure, players and terminology before you start real research. Treat this as background reading to be verified, not as the research itself.
5. 05Rehearsing the other side of a thesis. Ask it to build the strongest case against the position you are leaning toward. It is a genuinely useful sparring partner precisely because it does not care whether your thesis holds up.
6. 06Drafting the first version of a memo skeleton. The structure of an investment memo is fairly generic even when the content is not, so let the model handle the shape and put your analysis into it.

A worked example

The bad version: "summarise this company's prospects" with a 10-K attached and nothing else. You get fluent, generic optimism, a plausible-sounding growth number that may not trace to any line in the filing, and no way to tell which sentence is grounded and which is invented.

The version that works: ask a narrow, checkable question — "quote the exact sentence in this filing that discusses customer concentration risk, and give me the page reference" — then open that page yourself before it goes anywhere near a memo. Once you have the grounded fact, a second, separate request to draft a paragraph around it is low-risk, because you are no longer asking the model to also be the source.

Ask it to find and quote, verify what it found, then ask it to write. Collapsing those three steps into one request is where the wrong number gets into a document with your name on it.

Where it will quietly get you into trouble

• —Anything with a stale knowledge cutoff. A model's training data has a date. A share price, a most recent quarter's results, a rate decision, or "the current CEO" can all be wrong in a fluent, confident way if the honest answer required a live lookup it did not do. Insist on a tool that shows its sources for anything time-sensitive, and check the date on the source itself.
• —Arithmetic across a large model. It is a language model, not a calculator. It can produce a plausible total for a multi-step calculation that is simply wrong, and unlike a spreadsheet it will not throw an error — it will just be off. Compute in the tool that computes, and use AI to write about the result, not to produce it.
• —Citations and figures you have not opened yourself. A specific percentage, a named study, an exact multiple — these are exactly where fabrication concentrates, because specificity reads as credibility. Treat every number without a source you personally opened as unverified.
• —Material non-public information, in any form. A general consumer AI tool is not a data room and not covered by your firm's information barriers. This is the one line that should never move regardless of how useful the shortcut looks.
• —A judgement call dressed as a question. "Is this a buy" produces an answer with the tone of conviction and none of the accountability. The recommendation, and the reasonable basis behind it, stays yours under Standard V(A) no matter which tool helped you draft the memo.

The connecting thread is that these systems fail confidently rather than visibly, which is a harder failure mode to catch than an error message. What AI Is Actually Bad At covers the general pattern, and How to Check an AI Answer When You Are Not the Expert covers the checks that work even outside your specific coverage universe.

Getting started without a project plan

1. 01Confirm which tool, if any, your firm has actually vetted for confidential or client information, and use nothing else for anything beyond public documents. If nothing has been vetted, that is itself the first thing to raise.
2. 02Pick one recurring task where you can judge the output instantly — a first-pass industry summary or a comps formula is the usual right answer, not a valuation conclusion. Find the Repetitive Part is a short test for choosing it if the answer is not obvious.
3. 03Do that task both ways for two weeks — your way and with a draft — and keep the time. You want evidence, not a feeling that it is faster.
4. 04Build a small library of the prompts that actually produced grounded, checkable output rather than confident prose. The value compounds through reuse.
5. 05Expand to a second task only once the first is boring, and only ever into tasks where you can still catch a wrong number quickly.

The part that actually protects your job

The tasks most exposed here are the ones furthest from judgement: pulling a first-pass comp set, summarising a document, drafting a memo skeleton. Those are also the tasks that were already partly automated by tools that predate any of this. What is left — the recommendation, the reasonable basis behind it, and being the named analyst who is accountable for both — is not automatable in the same way, because part of its value is that a specific, accountable person stands behind it.

The useful stance is neither refusing the tools nor deferring to them. It is being the analyst who can produce a first draft in a fifth of the time and then reliably catch the one number in it that is wrong. That second half is the actual skill, and it is worth practising deliberately rather than assuming it comes free with the tool — the same conclusion reached from the accounting side in how to use AI as an accountant, and from the hiring side in AI-proof careers.

The short version

Use AI for reading long documents against a specific question, drafting narrative around numbers you already checked, and rehearsing the other side of your thesis. Do not use it for arithmetic at scale, time-sensitive facts, material non-public information, or the recommendation itself. Confirm what your firm has actually vetted before you put anything confidential near a general tool, check every number against its primary source, and start with one task you can verify instantly.

Coursium teaches exactly this layer — practical use of the tools plus the checking that has to go with it. Stay ahead of AI by learning them on your phone.