The recurring parts of analysis got faster: monthly commentary, deck narrative, the SQL for a data pull, the summary of a competitor's earnings call. A chat assistant does each in minutes, and a tool that runs code builds a scenario table from your CSV. What it cannot do is know your business: why the margin moved, what the board will ask.
The failure modes are specific. Language models approximate arithmetic unless they run code, confuse thousands with millions, mix fiscal and calendar years, and write a convincing explanation for a variance they cannot see. They also do not know today's rates, prices, or multiples.
The rules around your data are strict. Unreleased results, forecasts, and deal information are material nonpublic information; customer financial data may fall under GLBA; anything feeding external reporting sits inside SOX controls. Use approved tools, and review the output as you would a new analyst's work.
Quick wins this week
- Paste last month's budget vs actual (units stated) and ask for draft commentary on every line over your threshold, with DRIVER UNKNOWN wherever the model has no basis.
- Upload a competitor's earnings call transcript to NotebookLM and ask for guidance changes, pricing comments, and capex plans, each with a citation you can click.
- Describe the report you need from the data warehouse in plain English and ask for the SQL, then run it on a small date range before trusting the joins.
- Ask Copilot in Excel or ChatGPT to explain a formula you inherited, cell by cell, and to list every hard-coded number inside it.
- Turn a finished analysis into a one-page memo: headline, three drivers, risks, the ask. Then cut what the model padded.
What AI can do for financial analysts, task by task
Budget vs actual commentary
Give the comparative table with units, the cost-center names, and the drivers you already know. Ask for two-sentence explanations per line and a flag wherever the model has no basis. Check each driver against the ledger or source system; the model will assert a plausible cause for a variance it cannot see.
Auditing and documenting Excel models
Paste the formulas from a tab (Copilot can read the sheet directly) and ask for hard-coded inputs, sign errors, circular references, and formulas that break the pattern in their row. Then ask for a one-paragraph description of each tab for the model documentation. Rebuild one output by hand before trusting the audit.
Reading filings, transcripts, and research
Upload the 10-K, earnings transcript, or broker note to NotebookLM or Claude and ask for a table: guidance, segment revenue, margin commentary, capex, each with a page reference. Open every cited page. Extracted numbers are wrong often enough that an unchecked table is a liability.
Scenario and sensitivity tables
Upload a CSV of your driver assumptions to a tool that runs code and ask for base, downside, and upside cases with the code shown and every added assumption stated. Recompute one scenario in your own model; if they disagree, the model's assumptions differ from yours somewhere.
Board and executive narrative
Give the analysis, the audience, and the decision being requested. Ask for a storyline: governing thought, three supporting points, evidence for each, then action titles for the slides. Draft in Gamma or your template. Every number on a slide traces back to a cell you can name.
SQL, Power Query, and data cleanup
Describe the tables, join keys, and output columns you need. Ask for the query plus a plain-English explanation of each join and filter. Run it on a narrow date range and reconcile the total to a report you already trust before widening it.
Prompts for financial analysts
Replace the bracketed placeholders, paste into any chat assistant, and iterate on the result.
Variance commentary with flagged unknowns
You are an FP&A analyst writing commentary for the monthly reporting package. Amounts are in [UNITS]. Audience: [CFO OR DIVISION LEAD]. Below is the budget vs actual table for [MONTH] and last month's commentary for context. For each line where the variance exceeds [THRESHOLD], write two sentences: the variance, then the driver. Use only drivers supported by the context; where none exists, write DRIVER UNKNOWN. Note any variance that reverses last month's explanation. Output a table: Line, Variance, Commentary, Confidence. Table: [PASTE TABLE] Last month: [PASTE PRIOR COMMENTARY]
Tip: Including last month's commentary catches explanations that quietly contradict each other.
Audit a spreadsheet model
You are a senior financial modeler reviewing a colleague's Excel model. Below are the formulas from the [TAB NAME] tab, listed by cell. Report: (1) every hard-coded number inside a formula, with its cell, (2) formulas that differ from the pattern in their row or column, (3) sign or unit inconsistencies, (4) circular logic, (5) three questions you would ask the builder. Do not rewrite the model; return findings as a table with cell references. [PASTE FORMULAS WITH CELL REFERENCES]
Tip: Run it on a tab you know well first and count real findings versus noise.
Extract guidance and key figures from a filing or transcript
You are an equity research associate. Using only the document I uploaded, [DOCUMENT NAME AND PERIOD], build a table with rows for revenue guidance, margin commentary, capex plans, pricing actions, headcount, and any change from prior guidance. For each row give the figure or quote and the page or timestamp where it appears. If the document does not state something, write NOT STATED rather than inferring. Finish with three questions the document leaves open.
Tip: The NOT STATED cells are where a model would otherwise guess.
Build a scenario table with code
You are a financial analyst using Python. I uploaded [FILE NAME] with columns [LIST COLUMNS]. Build base, downside, and upside scenarios for [METRIC] over [PERIOD] by flexing [DRIVER 1] by [RANGE] and [DRIVER 2] by [RANGE]. Show the code, state every assumption you added, and return a table plus a short note on which driver the result is most sensitive to. Do not round until the final table. If a column is ambiguous, ask before computing.
Tip: Confirm the code actually executed; a table with no code block behind it is a guess.
Turn analysis into a one-page executive memo
You are a finance director writing for [AUDIENCE, E.G. THE CEO AND BOARD]. Turn the analysis below into a one-page memo: a one-sentence headline that states the conclusion, three drivers with one supporting figure each, two risks, and the specific decision required by [DATE]. No hedging language. Keep every figure exactly as given; do not compute new ones. [PASTE ANALYSIS OR NOTES]
Tip: Read the headline last. If it could have been written before the analysis, it is not a conclusion.
Want a prompt for something else? Use the Prompt Builder or browse data and spreadsheets prompts, finance and analysis prompts, presentations and reports prompts, research and learning prompts.
Skills to build
Stating units, periods, and definitions up front
Why: Most numeric errors in AI output trace back to an ambiguous input: thousands vs millions, fiscal vs calendar, gross vs net.
How: Keep a snippet with your fiscal calendar, units convention, and metric definitions, and paste it at the top of every prompt.
Knowing when the model is computing versus guessing
Why: Text-only answers approximate arithmetic. A scenario table produced without code is a guess dressed as a table.
How: Ask for code and check that it ran. In ChatGPT, Claude, and Julius, look for the executed code block; if it is missing, ask again.
Tracing every figure to a source
Why: A number in a board deck that came from a chat window and cannot be traced is a career risk, and under SOX it is a control gap.
How: Nothing AI-generated goes into a deliverable until it is linked to a cell, a page, or a query. Keep the link in your working notes.
Red-teaming your own storyline
Why: AI is good at generating the objections a CFO or board member will raise, before the meeting instead of in it.
How: Paste your storyline and ask for the five hardest questions and the evidence that would answer each. Build a backup slide for the two you cannot answer.
Reusable analysis templates
Why: Monthly commentary, model audits, and filing summaries repeat. A saved prompt with your definitions and format keeps output consistent.
How: Put your best prompts in a Claude Project or custom GPT with your units convention, fiscal calendar, and house style in the instructions.
Tools worth knowing
ChatGPT
The general-purpose AI assistant most of your coworkers already use.
Claude
A careful writing and analysis assistant that shines on long documents.
Microsoft Copilot
AI inside Word, Excel, Outlook, and Teams, with your company's data protections.
Julius AI
Chat with your spreadsheets and data files and get charts, stats, and answers back
NotebookLM
A research notebook that only answers from the sources you give it, with citations.
Perplexity
An answer engine that cites its sources, built for research rather than chat.
Cautions for financial analysts
Unreleased results, forecasts, M&A details, and customer contracts are material nonpublic information. Pasting them into a consumer AI tool can breach company policy and, if the tool retains data, create disclosure and insider-trading exposure. Use only tools your company has approved for confidential data, never a personal account.
If your analysis feeds external reporting, AI-generated figures or commentary fall inside internal control over financial reporting. Keep the same preparer and reviewer sign-off, record that AI was used and what was checked, and expect internal audit to ask for that record.
At a bank, insurer, or broker-dealer, customer financial data is protected under GLBA and its privacy rules, and FINRA supervision and recordkeeping rules can reach AI-assisted communications with the public. Check with compliance before using any AI tool on client-related work.
Models miscount, drop a zero, mix thousands and millions, and confuse fiscal quarters. They do not know current rates, prices, or valuation multiples and will produce a plausible number anyway. Use code execution for math, citation tools for market facts, and re-foot every table.
Your 30-day plan
- Week 1: Confirm which tools are approved for confidential data and whether a business tier with no-training terms exists. Write your definitions snippet: units, fiscal calendar, key metrics.
- Week 1: Run last month's commentary through the variance prompt on a de-identified extract and compare with what you wrote. Note every invented driver.
- Week 2: Have the model audit one tab of a model you know well and count true findings versus noise.
- Week 2: Summarize one filing or transcript with citations and open every cited page. Time it against your usual method.
- Week 3: Build one scenario table with code from a CSV of your drivers and reconcile a scenario to your own model.
- Week 4: Save your prompts and definitions in a Project or custom GPT, and use the red-team prompt before your next leadership review.
Frequently asked questions
Will AI replace financial analysts?
Can I put company financials into ChatGPT?
Can ChatGPT build a financial model?
Is AI accurate with numbers?
Which AI tool is best for FP&A?
Terms used on this page
Related roles
- AI for AccountantsAI drafts the flux commentary, writes the XLOOKUP, explains the K-1 to your client in plain English, and reads the 40-page lease for you. You still own the numbers, the judgment, and the signature.
- AI for Data AnalystsAI writes the SQL, the pandas, the DAX, and the sentences around the chart. It cannot know your data, so your job shifts to the right question, the checked join, and keeping regulated data out of the wrong tool.
- AI for ConsultantsAI synthesizes twelve interview transcripts into themes with quotes, builds the issue tree, argues against your recommendation before the client does, and drafts the SOW. The judgment, the relationship, and the confidentiality obligations are yours.
- AI for Operations ManagersOperations managers live in SOPs, incident reports, staffing plans and spreadsheets, which is the material AI handles best. Here is how to use it to get your week back, and where the line is for safety and people decisions.