How to Use Claude for Financial Analysis

To use Claude for financial analysis, give it the reading, reasoning and drafting, connect it to the data your firm already licenses, write your house conventions into projects and skills, and send every number headed for a model or a memo through a tool that computes it in code. That last step is where Kepler fits: Kepler runs inside Claude as an Anthropic-verified connector, so each figure comes back pulled from the filing, calculated by code and cited to the exact line, while Claude keeps the analysis and the write-up.

This guide covers what Claude for Financial Services includes as of November 2026, how to set it up, prompts that work for filings, models and memos, where to check the output, and how to put recurring work on a schedule. We build Kepler, and Anthropic published a case study on how we built Kepler with Claude. Facts about Anthropic’s products come from Anthropic’s own pages, linked inline.

What Claude brings to financial work

Anthropic launched Claude for Financial Services in July 2025 and has added to it twice in a big way since. The October 2025 update brought Claude into Excel as a beta, added connectors to LSEG, Moody’s, Aiera (with a Third Bridge data feed), Chronograph, Egnyte and MT Newswires, and shipped six finance skills: comparable company analysis, DCF models, due diligence data packs, company teasers, earnings analyses and initiating coverage reports. The May 2026 release made the Excel, PowerPoint and Word add-ins generally available and published ten agent templates, among them a pitch builder, an earnings reviewer, a model builder and a month-end closer.

On the benchmark analysts cite most, Claude is the leader. In the October 29, 2026 update to Vals AI’s Finance Agent benchmark, Claude Opus 5.5 placed first at 67.85%, with Claude Sonnet 5.5 fourth.

Where Claude works	What analysts use it for	Status, November 2026
Claude app (chat and Cowork, now merging into one)	Reading filings and transcripts, research, drafting, scheduled briefs	All paid plans
Excel	Explaining a workbook, changing assumptions, building multi-tab models	Generally available
PowerPoint	Slides in your master, native charts and diagrams	Generally available
Word	Memo edits with tracked changes, in company styles	Generally available
Outlook	Inbox triage, draft replies, scheduling	Beta
Claude Code, API, Managed Agents	Firm-built agents with per-tool permissions and audit logs	Varies by product

The Microsoft 365 page says context carries from Outlook to Word to Excel to PowerPoint, so a brief can become a model and then a deck without re-explaining the work. Firms can run Claude on AWS, Google Cloud or Azure, and Anthropic says it doesn’t train on Team or Enterprise data by default.

Two things Claude can’t know on its own: your firm’s conventions (how you define adjusted EBITDA, your peer sets, your templates) and whether a figure it wrote matches the filing. The steps below handle the first. The checks in step 4, or Kepler, handle the second.

Step 1: Set up the workspace

Pick the surface for the job. Use the Claude app for reading and research, Excel for models, PowerPoint and Word for deliverables, and Cowork, which Anthropic is folding into the main Claude app, for longer tasks that touch files and connected tools. Engineers who want firm-specific agents start in Claude Code or the API.

Create a project per coverage area or deal. Put the reference material Claude should always have in the project: the coverage list, your definitions sheet, the model template, the memo template and the style guide. Project instructions apply to every chat inside it.

Connect licensed data through your own subscriptions. Anthropic’s solutions page lists pre-built connectors to LSEG, FactSet, S&P Global (Capital IQ), Morningstar, PitchBook, Moody’s, Dun & Bradstreet, Daloopa, Box, Databricks and Snowflake, and the May 2026 release added Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C Intralinks, Third Bridge and Verisk. Each premium connector works on the license your firm already holds. On Team and Enterprise plans, an owner adds a connector under Organization settings, then each analyst signs in individually, which keeps Claude to the data that analyst is entitled to see (Anthropic support). A connector set up with one shared API key gives everyone that key’s access, so ask for per-user sign-in where the provider offers it.

Install the finance plugins. Anthropic’s financial services repository packages the agents, skills and connectors as plugins. In Cowork (or the combined Claude app, where it has rolled out), open Settings, then Plugins, then Add plugin, and paste the repository address. The core financial-analysis plugin adds slash commands such as /comps, /dcf, /earnings and /ic-memo. Our guide to MCP for financial data covers what to check before you connect any server.

Step 2: Write your firm’s conventions down

Claude applies what you give it. Most bad AI-generated analysis traces back to an unstated convention: a fiscal year treated as a calendar year, thousands read as millions, a press-release figure used where the 10-K differs.

A skill is the place to write those conventions once. Anthropic describes Agent Skills as folders with a SKILL.md file that teach an agent a task, plus any scripts or references it needs; the standard is now maintained under the Linux Foundation’s Agentic AI Foundation. Skills work the same way in Excel, PowerPoint and Word. A useful first skill for an investment team covers:

• Definitions. Adjusted EBITDA with your add-backs listed, net debt with or without leases, your peer sets by name.
• Periods. How to calendarize fiscal years and whether to use LTM or the last fiscal year by default.
• Source order. Filed 10-K and 10-Q first, then the earnings release, then the transcript, then anything else, and to flag any disagreement between them.
• Units and signs. State the scale on every figure, show costs as negatives or positives consistently.
• Gaps. When a filing doesn’t disclose a figure, leave the cell empty and say why.
• Format. Your memo headings, your color code for inputs and formulas, your chart style.

Step 3: Prompts that work

Good prompts for financial work name the document, the period, the basis and the output, and ask Claude to show its source for every figure.

Filings and transcripts

• “From the company’s two most recent 10-Ks, list every risk factor that was added, removed or materially reworded. Quote both versions and give the page for each.”
• “List every commitment and contingency in the notes to the latest 10-K, with the note number, the amount and its unit. If the 10-Q since then updates any of them, show both.”
• “Summarize what management said about pricing on the last two calls. Quote each executive by name, and separate prepared remarks from Q&A.”

Models in Excel

• “Build a three-statement model for FY2022 to FY2029. Put historicals on their own tab with a source column (filing, page, line) for every hardcoded number. Projections are formulas only, inputs in blue.”
• “Walk me through how the revenue growth assumption flows to free cash flow in this workbook, cell by cell.”
• “Before you change anything, list every cell you plan to edit and why.”

Memos and decks

• “Draft the IC memo in our template using only figures from the Model tab. Mark any number you didn’t take from the workbook.”
• “Turn the comps tab into two slides in our master, with each multiple footnoted to its cell.”

For the comps workflow end to end, see how to build trading comps with AI.

Step 4: Check the numbers before they leave the desk

Claude’s benchmark lead is real, and the same leaderboard explains why checking still matters. Vals AI describes its test as the work of an entry-level analyst on public filings, the leader earns about two-thirds of the available credit, and the best Financial Modeling score is 57.0%. Anthropic frames its own reference agents the same way: the repository says they “draft analyst work product” for “review by a qualified professional.”

Run these checks on any figure that drives a decision:

1. Open the citation. The exact number should be on the page, for the right period, in the right unit, on the right basis (GAAP or adjusted, as filed or as restated).
2. Recompute anything derived. A margin, a multiple or a growth rate should rebuild from its inputs.
3. Ask again in a fresh chat. A figure that changes between runs was written by the model rather than looked up.
4. Ask for a complete list and count it. “Every 8-K the company filed last year” against EDGAR tells you whether Claude read everything or the top results.
5. Follow the number out. Check that the source survives into the workbook and the deck.

Our checklist for AI-generated numbers goes deeper, and ChatGPT vs Claude for financial analysis compares the two assistants on the same points.

Step 5: Put recurring work on a schedule

Claude’s scheduled tasks run on all paid plans, hourly, daily, weekly or on weekdays, and they now run remotely, so they continue when your laptop is closed (tasks that need local files or apps still run locally). Each run has the same connectors, skills and plugins as a normal task. Good first schedules for an investment team: a pre-market brief on the coverage list, a weekly summary of new 8-Ks for holdings, and a Friday digest of guidance changes.

Engineering teams can go further with Claude Managed Agents (in beta), which Anthropic describes as long-running sessions with per-tool permissions, managed credential vaults and a full audit log. Whichever route you take, treat everything an agent reads as data. Anthropic’s support pages warn that a malicious MCP server “may include hidden instructions,” and the same risk applies to a document or an email an agent opens.

How Kepler does this inside Claude

Kepler is an Anthropic-verified connector in Claude. Find it in the connector directory, connect once with your Kepler account, and it’s available in every chat on desktop, web and mobile, with no API key to manage. Claude keeps the reasoning; Kepler does the retrieval and the arithmetic.

Ask in the same chat (“Use Kepler: calculate Palantir’s fully diluted share count and enterprise value from the latest 10-Q at today’s price”) and the answer arrives as a Kepler card. It holds the full analysis, a sources block with a badge showing whether every figure was cited, and an Open in Kepler link to the citation view and workbook. Claude’s summary comes underneath. Our post Kepler in Claude: a working guide for analysts walks through three real runs.

What changes is underneath each number. Kepler’s deterministic engine pulls every figure from its source and computes every derived value as code over cited inputs, so the model never writes a digit. Statements assemble from the filings’ own XBRL, as filed or as restated, and every citation is checked against its source before the answer reaches you. Ask the same question next week and you get the same number with the same sources. Ask for a model and it comes back as a workbook with every historical cell sourced; where a filing doesn’t disclose a line, the cell stays empty and the run says why.

Kepler also knows your firm. Our forward-deployed engineers, who built this layer before at Palantir, build your firm’s ontology from your documents and systems, so “our adjusted EBITDA” and “the Fund III portfolio” mean one thing in every answer. Every document is read in full before you ask: 26M+ SEC filings, 40M+ IR documents and earnings call transcripts, and whatever you connect from SharePoint, Box, Egnyte, Outlook or your CRM. Agents run on schedules and on new filings, acting as the person who set them up.

Connect Kepler in Claude, or start at app.kepler.ai. Engineers can build on the same engine through our API and MCP server, and we’re happy to run a demo on your own coverage list.

Frequently asked questions

Can Claude do financial analysis?

Yes: Claude reads filings and transcripts, reasons through multi-step questions, builds and edits Excel models, and drafts memos and slides in your templates. For figures that go into a model or an IC memo, add Kepler from Claude's connector directory: Kepler computes every number in code from the source and cites it to the exact page, line or cell, while Claude keeps the analysis and the writing.

What is Claude for Financial Services?

It is Anthropic's package for banks, asset managers, insurers and fintechs, launched in July 2025. It includes Claude in Excel, PowerPoint and Word, data connectors to providers such as FactSet, S&P Capital IQ, PitchBook, Morningstar, LSEG and Moody's, finance skills for comps, DCFs and earnings reviews, ten agent templates released in May 2026, and deployment on AWS, Google Cloud or Azure.

How do I connect Claude to FactSet, Capital IQ or PitchBook?

On Team and Enterprise plans, an owner adds the connector under Organization settings, then each analyst signs in with their own provider account. Premium connectors need your firm's existing subscription with that provider. Anthropic's financial services repository on GitHub also lists the MCP server address for each provider, for firms that run Claude through Claude Code or the API.

Can Claude build a financial model in Excel?

Yes. Claude in Excel is generally available on paid plans and can build multi-tab models, explain any cell and change assumptions without breaking formulas. Ask it to keep a source column for every hardcoded historical and to list the cells it will change before it edits. Through Kepler, the historicals arrive from the filings' own XBRL with every cell cited.

How accurate is Claude on financial questions?

Claude leads Vals AI's Finance Agent benchmark: Claude Opus 5.5 ranked first at 67.85% in the October 29, 2026 update, and the top Financial Modeling score was 57.0%. That is the strongest public result, on tasks Vals describes as entry-level analyst work, so check every figure that drives a decision against its source before it leaves your desk.

How do I use Kepler inside Claude?

Search for Kepler in Claude's connector directory, connect it and sign in with your Kepler account. It then appears in every chat on desktop, web and mobile, with no API key to manage. Ask your question in the same chat and the answer arrives as a Kepler card with every figure linked to its source, followed by Claude's summary.