Can ChatGPT Do Financial Analysis?
Yes for reading, drafting and first-pass models. For numbers you'll act on, ChatGPT needs checking. What works, what breaks, and how to test any AI tool.
Yes, ChatGPT can do a good share of financial analysis (reading filings and transcripts, summarizing, explaining accounting, drafting memos, building and editing Excel models), but the numbers need a second look, because a language model writes figures as text. For figures headed into a model, an investment committee memo or a client deliverable, analysts use Kepler, in its own app or called from inside ChatGPT: it computes every figure in code from the source and cites it to the exact page, line or cell, on the firm’s own definitions.
That answer holds for every version of ChatGPT, from the free app to ChatGPT for Financial Services, the plan OpenAI launched for eligible institutions on September 10, 2026. The finance plan adds data and citations, which help a great deal.
We build Kepler, enterprise intelligence for investment firms, and it runs inside ChatGPT and Claude. This guide covers what ChatGPT does well, what breaks when the work turns to numbers, how to test any AI tool, and how we built Kepler to pass those tests. Facts are as of November 2026.
Which ChatGPT are we talking about?
“ChatGPT” covers three quite different setups for finance work, and the answer to “can it do financial analysis?” shifts a little with each. The last row shows what changes when Kepler runs inside it.
| Version | Financial data | Best for |
|---|---|---|
| ChatGPT Free, Plus and Pro | Web search, files you upload and apps you connect | Individuals learning, reading and drafting |
| ChatGPT Business and Enterprise | The same, plus admin-managed connectors and custom MCP apps for internal data | Teams that need enterprise controls |
| ChatGPT for Financial Services | Included SEC filings, Quartr, Daloopa, Fiscal.ai, PitchBook Essentials, Crunchbase, LSEG News and delayed pricing, plus 50+ connectors | Eligible banks and research teams |
| ChatGPT with Kepler | Your firm’s documents and systems, plus 26M+ SEC filings and 40M+ IR documents and earnings call transcripts, every page read in full | Numbers a firm will act on, computed in code and cited |
ChatGPT for Excel works across all of them, and OpenAI’s March 2026 update added data apps for providers such as Moody’s, Dow Jones Factiva, MSCI and Third Bridge. More data and better citations make the answers easier to check.
The short answer, task by task
The middle columns describe ChatGPT used on its own. The last column shows the same task with Kepler, inside ChatGPT or in its own app.
| Task | ChatGPT on its own | What to watch | With Kepler (inside ChatGPT or on its own) |
|---|---|---|---|
| Summarize a 10-K or earnings call | Strong | Quoted figures should match the filing | Every quoted figure cited to its page and line |
| Explain an accounting treatment | Strong | Confirm against the company’s own note | Cites the company’s own note |
| Draft a memo or note from your inputs | Strong | Numbers carry over from the inputs | Drafted in your template, every number traced |
| Build or fix an Excel model | Good, especially with ChatGPT for Excel | Historical inputs and formula logic | Statements from XBRL into your workbook, every cell cited |
| Pull historical figures from filings | Needs checking | Period, units, restatements | Pulled by code, as filed or as restated |
| Comps on adjusted metrics | Needs checking | Whose definition of adjusted EBITDA | Your firm’s definition, applied every time |
| Screen many companies on a disclosure | Limited | Whether the list is complete | Every filing read in full, so the list is complete |
| Give the same answer twice | Not guaranteed | Reproducibility | Same number, same sources |
What ChatGPT does well
Reading. Hand it a 10-K, a merger proxy or an earnings call and it will summarize the business, pull out risk factors, compare this quarter’s commentary with last quarter’s and explain why gross margin moved. For an analyst ramping on a new name, that turns a day of reading into an hour.
Explaining. It’s a patient tutor on lease accounting, deferred revenue, purchase price allocation and the difference between GAAP and non-GAAP measures.
Drafting. Give it your numbers and your view and it writes a credible first draft of an earnings note, a one-pager or an IC memo section.
Working in Excel. ChatGPT for Excel has been generally available on all plans since May 5, 2026. It builds and updates models from plain-language requests, explains how assumptions flow, links answers to the cells it references and asks permission before editing. OpenAI reported in March 2026 that on its internal investment banking benchmark, whose tasks include a formatted three-statement model with citations, GPT-5.4 Thinking scored 87.3%, up from 43.7% for GPT-5. Our list of the best AI tools for financial modeling compares it with the other Excel add-ins.
Finance data, on the finance plan. ChatGPT for Financial Services includes SEC filings, Quartr transcripts, Daloopa financials, Fiscal.ai fundamentals, PitchBook Essentials, Crunchbase, LSEG News for US users and delayed market data, per OpenAI’s help page. Citations open the specific table or passage behind a figure. OpenAI notes that standard ChatGPT Enterprise seats don’t include these datasets.
Where it breaks with numbers
1. The model writes the digits
A language model generates its answer one token at a time, and a figure like $4,812 million is made of tokens like any other word. Most of the time it’s right. When it’s wrong, it reads exactly as confidently as when it’s right. A citation added afterward can show that a source exists without proving the number came from it. Our guide to AI hallucinations in finance explains why better models make these errors quieter, not gone.
ChatGPT can run Python on a file you upload, which takes care of the arithmetic. The inputs still come from the model reading the document and typing the figures into its code, and nothing ties each input to a cited line.
The best public measure of this is Vals AI’s Finance Agent benchmark, which asks models to do entry-level analyst work from public company filings. On the October 29, 2026 update, OpenAI’s highest-scoring model on the leaderboard, GPT-6.1 Sol, scored 61.82%, GPT-6 Astra scored 60.43%, and the overall leader, Claude Opus 5.5, scored 67.85%. The top Financial Modeling score was 57.0%. Our ChatGPT vs Claude comparison has the full picture.
An independent test in Excel shows how this kind of error hides, whichever tool makes it. Wall Street Prep had four AI tools build Apple’s three-statement model from its 10-K and fourth-quarter press release, in a test first published in February 2026. Two of the four filled historical line items with wrong figures that still summed to the right subtotals, errors the testers described as “subtle enough to be dangerous.” The testers’ conclusion was that “even the best tool still underperforms a Junior Analyst.”
Kepler takes the model out of both the inputs and the arithmetic. Code pulls each input from the filing’s XBRL or the cited passage, runs the calculation with its formula shown, and checks every citation against its source before the answer reaches you.
2. It reads what it retrieves
On a long filing or a large document set, a model works from the passages its search returns. Material in a footnote, an exhibit or page 300 of a data room can sit outside that window. Ask for every 10-K in a sector that disclosed a material weakness in the past three years and you get the strongest matches, with no signal about what was missed. OpenAI also describes its included datasets as “selected coverage, which may differ from the provider’s full product.”
Kepler works the other way around: extraction models we train read every page of every filing at ingest, footnotes and exhibits included. Put the same material-weakness question to it and the list holds every company that disclosed one.
3. Definitions belong to someone
“Adjusted EBITDA” means something different at each company, and something different again at your firm. ChatGPT knows your house definition only if you paste it in or attach a template, and it applies it only as carefully as each prompt asks; our guide to getting AI to use your firm’s definitions explains why prompts and custom instructions don’t hold. In Kepler, our engineers write your house definitions into your firm’s ontology once, and every question, agent and workbook applies them without a reminder.
4. Restatements and time
Companies recast prior periods after acquisitions, segment changes and accounting updates. A figure “for Q2 2025” can be the number as originally filed or the number as shown in a later filing. In a study of 200 companies, 194 reshaped a statement in ways the XBRL tags alone can’t follow, which we document, along with how we recover the changes, in our paper Reclassified to Conform. A model that doesn’t track which version it’s using will mix them. Kepler keeps both versions of every figure, as filed and as restated, on any date, and labels which one each answer uses.
5. Units, scale and period
Thousands against millions, fiscal against calendar quarters, trailing twelve months against fiscal year. These errors are small to make and large in effect, and they’re the most common ones analysts catch in review. Kepler records each fact with its value, scale, unit and period when it reads the filing. Our research note Stop Paying to Forget measured it: across 40 held-out filings, scale was right on 99.4% of tagged facts.
6. Reproducibility
Ask ChatGPT the same quantitative question twice and you can get two answers. For a draft that’s fine. For a number in a model that a PM or a client will see, it means someone has to establish which one is right. Because Kepler computes each figure in code, asking twice returns an identical number, cited to the same lines.
A worked example
Say you ask for a company’s operating margin over the last eight quarters. A plausible answer can quietly combine five things: six quarters from the 10-Qs, one from the earnings release’s adjusted table, one from a later filing that restated it, a revenue figure in thousands next to an operating income figure in millions, and a margin computed off total revenue in one quarter and segment revenue in another. Each figure carries a citation, and the series still mixes bases, versions and units.
Asked the same question, Kepler returns the eight quarters on one basis: each labeled as filed or as restated, revenue and operating income at the same scale, the margin computed in code on the denominator you choose, and every input one click from its line.
How to test ChatGPT, or any AI tool, on financial analysis
Run these six checks on live work before you trust a tool with figures:
- Click ten citations. Each one should land on the exact page, line or cell, showing the period and units.
- Recompute three derived numbers. Margins, growth rates and multiples should show their formula and inputs.
- Ask the same question twice. The figures should match exactly.
- Ask for your definition. Adjusted EBITDA or net debt calculated your team’s way, on a company you cover closely.
- Ask for a restated period. Request a quarter both as filed and as restated, and see whether the tool can tell them apart.
- Ask for a complete list. Then check a few names you know should be on it.
We built Kepler to pass all six. Run the same questions through Kepler and through the tool you use today, and compare what comes back.
Our guide on how to check AI-generated numbers in financial research goes deeper, and the best AI for SEC filing analysis compares tools on exactly these tests.
When ChatGPT is enough
ChatGPT on its own is a strong fit for learning a company or an industry, summarizing documents, first drafts, personal investing research and Excel mechanics. It needs a different setup when the number will appear in a model, an IC memo, an LP report or a client deliverable; when the firm’s own definitions and documents have to be part of the answer; when a list has to be complete; and when compliance needs a record of where every figure came from (see AI governance for investment firms). In those cases, the figures should come from Kepler, in its own app or called from inside ChatGPT.
How Kepler does this
We build Kepler.
We built Kepler for the moment an answer turns into a number someone will act on. Analysts who like ChatGPT keep using it: Kepler runs inside ChatGPT and Claude as well as in its own app, Excel, PowerPoint, Word and Slack, and your engineers can reach the same engine through our API and MCP server.
The work starts before the question. In the first week, our forward-deployed engineers connect your firm’s sources, such as SharePoint, Box, Egnyte, Outlook, your CRM and your data rooms. Public data is loaded before you start, covering 14,000+ companies: 26M+ SEC filings across 20+ form types, plus 40M+ IR documents and earnings call transcripts. Extraction models we train read every page, footnotes and scanned exhibits included, and record each fact with its value, scale, unit and period beside its passage.
All of it resolves into your firm’s ontology: one record for each company, fund and deal, one definition for each house metric, every link tied to the sentence that supports it, and a sense of time, so a figure is available as filed or as restated on any date. When you ask for a number, code pulls and computes it. Statements assemble from XBRL without a model touching a figure, each calculation carries its formula and cited inputs, every citation is checked against the source while the answer is written, and where filings disagree you see both versions with the conflict flagged.
The work keeps running when nobody is asking. Agents watch your coverage on a schedule and when filings land, acting as the analyst who set them up, and send each change with the exact sentence. Your firm’s templates are saved once as playbooks, so the comps sheet or IC memo comes out in house format, and the corrections your team makes improve the firm’s own models privately. Access follows your firm’s permissions, checked on every read; every query is logged with its lineage; nothing Kepler reads is ever taken as an instruction; and we hold SOC 2 Type II.
We send each step to the model that handles it best, with small models we train doing the reading and frontier models handling the question and the write-up, so Kepler improves each time the labs ship a better model. Our engineers built this layer before at Palantir.
- “Operating margin for the last eight quarters, as filed and as restated, with every input cited.”
- “Every company in our coverage whose revenue recognition note changed this year.”
- “Tell me whenever a company we cover discloses a material weakness, with the sentence and the filing.”
Run the six tests on Kepler at app.kepler.ai, or book a demo and we’ll run them on your own coverage.
Frequently asked questions
Can ChatGPT analyze financial statements?
Yes. ChatGPT can read an income statement, balance sheet and cash flow statement, explain the movements, compute ratios and draft commentary; for the figures themselves, Kepler runs inside ChatGPT and returns each one computed in code from the filing, labeled with its period, units and whether it's as filed or as restated. On its own, the model writes figures as text, so check each one against the filing, confirm the period, units and whether the statement was restated, and recompute any ratio before it goes into a model or memo.
Is ChatGPT good for financial analysis?
It's very good for reading filings and transcripts, explaining accounting, drafting notes and building Excel scaffolding, and ChatGPT for Financial Services adds included data and citations. It's weaker on figures a firm will act on. On Vals AI's Finance Agent benchmark, updated October 29, 2026, OpenAI's highest-scoring model scored 61.82% and the overall leader 67.85% on entry-level analyst tasks. For those figures, Kepler runs inside ChatGPT and returns each one computed in code and cited to the line.
Can I use ChatGPT to build a financial model?
Yes, as a first draft. ChatGPT for Excel, generally available since May 2026, builds and updates models in your workbook, links answers to the cells it uses and asks before making changes. Check every historical input against the filing and every assumption against your own view. With Kepler, we build statements from XBRL with every cell cited and send them to the same workbook.
What is the difference between ChatGPT and ChatGPT for Financial Services?
ChatGPT for Financial Services, launched September 10, 2026, is a separate plan built on ChatGPT Enterprise for eligible financial institutions. It includes datasets such as SEC filings, Quartr transcripts, Daloopa, Fiscal.ai, PitchBook Essentials, Crunchbase and LSEG News, with citations that highlight the source table or passage. OpenAI says standard ChatGPT Enterprise seats don't include these datasets.
How do I check ChatGPT's financial numbers?
Open ten citations and check the figure, period and units on the source page. Recompute every derived number such as a margin or multiple. Ask the same question twice and compare the answers. Ask for a metric calculated your firm's way. Ask for a complete list and count the results. Kepler is built to pass all five: code computes every number and cites it to the exact page, line or cell, your firm's definitions live in its ontology, every document is read in full so lists are complete, and the same question returns the same figure.