How to use Claude with Excel for FP&A: a practitioner's guide

Last updated: July 2026

There are three ways to use Claude with Excel for FP&A work, and they solve different problems. First, the Claude for Excel add-in puts Claude in a sidebar inside your workbook, where it can read, analyze, and modify the model with cell-level citations — available on Pro, Max, Team, and Enterprise plans. Second, Claude connected to your FP&A data layer over MCP (the Model Context Protocol) lets Claude query live, permissioned numbers instead of working from pasted exports. Third, Claude Skills and Cowork package your recurring workflows — monthly variance commentary, board-deck narratives — into repeatable procedures Claude runs the same way every time.

Most finance teams that get real value from Claude end up combining them: the add-in for in-workbook review, MCP for trustworthy data access, and Skills for the workflows you repeat every close. This guide walks through all three, with concrete walkthroughs for the jobs FP&A teams actually use Claude for — variance commentary, model auditing, and board narratives — plus the prompt patterns and security caveats that separate a useful deployment from a risky one. (For the wider landscape beyond Excel, see our guide to Claude for finance teams.)

Bottom line: Claude is genuinely strong at the review-and-narrative half of Excel work — auditing formulas, drafting variance commentary, translating model outputs into board language. It is not yet trustworthy as an autonomous model-builder, and Anthropic itself says not to ship its Excel output as a final deliverable without review. Treat Claude as a fast first-pass analyst whose work you check, and feed it governed data over MCP rather than pasted exports.

The three ways to pair Claude with Excel

Each approach connects Claude to your numbers differently — and that difference matters more than any feature list.

1. The Claude for Excel add-in

Claude for Excel is Anthropic's official add-in: Claude sits in a sidebar inside Excel (web, Windows, Mac, and iPad) and works directly on the open workbook. It can read and explain a model with cell-level citations, modify assumptions while preserving formula structure, debug errors, navigate multi-tab workbooks, and apply native Excel operations like sorting, filtering, and conditional formatting. Anthropic launched it in beta in late 2025 as part of its financial services push, and it now supports MCP connectors for external data inside the sidebar.

Two limitations matter for FP&A: it can't handle macros, VBA, or Excel data tables — so heavily automated legacy models are partly out of reach — and Anthropic explicitly does not recommend it for final client deliverables without review or for audit-critical calculations. That's the vendor telling you what we'd tell you anyway: it's a reviewer and a drafter, not an unsupervised modeler.

2. Claude + an MCP-connected FP&A data layer

MCP is the open standard Anthropic released in November 2024 for connecting AI assistants to the systems where data actually lives. Instead of pasting a P&L export into a chat window, Claude queries your data layer directly — with the same permissions, definitions, and version control your team already enforces. Aleph ships an MCP server that does exactly this: Claude (or ChatGPT) can query governed Aleph data — actuals synced from your ERP, CRM, and HRIS — and every answer traces back to one source of truth rather than whichever export happened to be on someone's desktop.

This is the difference between "Claude analyzed a file someone made on Tuesday" and "Claude analyzed the numbers." More on why that distinction dominates everything else below.

3. Claude Skills and Cowork workflows

Skills are folders of instructions, reference files, and scripts that teach Claude your way of doing a task — your variance-commentary format, your board-deck structure, your materiality thresholds. Claude invokes the relevant skill automatically when the task matches. Skills run across the Claude apps (web and desktop), Claude Code, and the API on any plan with code execution enabled — a common misconception is that they need Claude Desktop; they don't. Cowork, Anthropic's agentic desktop mode, extends this to multi-step work on local files and is available on paid plans via Claude Desktop.

For FP&A, Skills are how you stop re-explaining your conventions every month. Write the procedure once — "commentary covers every line over $25K or 10% variance, one sentence of driver, one of outlook, never speculate on unexplained gaps" — and every future run follows it.

Walkthrough: drafting variance commentary from a BvA export

The highest-ROI starter workflow, because the input is structured and the output is judged by a human anyway.

1. Get the BvA in front of Claude. Either open the workbook with the Claude for Excel add-in, or — better — have Claude pull the current budget-vs-actuals through an MCP connection so you're not screenshotting a stale export.
2. Set the frame before asking for output. Tell Claude the audience (CFO? department heads?), the materiality threshold, and the tone. Unframed, it will dutifully explain a $300 variance in office snacks.
3. Ask for drivers as hypotheses, not facts. Claude can see that T&E is 40% over budget; it usually can't see why. Have it flag each variance with a suggested driver and a confidence marker, then confirm or correct the ones that matter.
4. Edit, don't regenerate. The draft gets you from blank page to 80%. The last 20% — knowing that the marketing overage was a board-approved pull-forward — is your job, and it's the part anyone reads the commentary for.

A prompt that works:

"Here is our March BvA by department. Draft variance commentary for the CFO. Cover every line where the variance exceeds $25K or 10%. For each: one sentence stating the variance, one suggesting the likely driver marked as [CONFIRM], one on run-rate implication. Flat, factual tone. Do not explain immaterial lines."

For more on structuring finance prompts like this, we've written a full guide to AI prompting in finance.

Walkthrough: auditing a model and reviewing formulas

This is where the Claude for Excel add-in earns its seat, because cell-level citations turn "the model might have issues" into "click here, here, and here."

Open the workbook with the add-in and work through a review checklist:

• Structural scan first. "Map this workbook: what does each tab do, where are the inputs, where are the outputs, and which tabs feed which?" You get an orientation document for a model you inherited — genuinely faster than doing it by hand.
• Hunt hardcodes. "Find every cell in the forecast tabs where a constant is embedded inside a formula or overrides a formula in a calculated range." Plugged numbers are the silent killer of inherited models.
• Check row consistency. "Identify rows where the formula changes partway across the columns." A dragged formula that stops one column short is invisible on screen and catastrophic in the out-months.
• Trace the number you doubt. "Walk me through every precedent of the FY26 EBITDA cell and flag anything that looks off — sign conventions, double-counting, references to empty cells."

Claude will surface real issues and occasional false alarms. The cell citations mean each flag takes seconds to verify, which is the right division of labor: Claude generates leads, you make calls. Remember the constraint: models that lean on VBA or data tables are partly opaque to the add-in, so score coverage accordingly.

Walkthrough: board narrative from model outputs

Third workflow: turning a finished model's output tab into the narrative for the board deck or investor update.

Give Claude the outputs (via add-in, MCP query, or paste — governed beats pasted, see below), plus two things people usually forget: last quarter's narrative (so this one reads as a continuation, not a reboot) and the two or three messages you want to land. Then ask for structure, not prose-poetry:

"Using this quarter's outputs and last quarter's board letter, draft the financial performance section of the Q2 update. Lead with ARR and net burn vs. plan. Frame the gross-margin decline as the data-cost investment we flagged in Q1. Three paragraphs max. No adjectives a CFO wouldn't say out loud."

Claude is very good at this translation layer — numbers to defensible sentences. What it cannot do is decide what the board should worry about. That's the judgment call that makes it your letter; if the narrative strategy is coming from the model, something is upside down.

Where Claude is strong in Excel work — and where it isn't

Strong: review acceleration. Formula audits, hardcode hunts, tie-out checks, explaining an unfamiliar model, drafting commentary and narrative from numbers that already exist. These tasks share a shape: large surface area, mechanical checking, human verdict at the end. Claude compresses hours of scanning into minutes of verification.

Strong: translation. Model outputs to board language, BvA tables to commentary, assumptions to plain-English documentation. Finance teams underrate how much of the month is translation work.

Weak: autonomous model-building. Claude can scaffold a three-statement model or a cohort waterfall, and the scaffold is a real time-saver. But unreviewed, multi-step quantitative construction is exactly where errors compound silently — and Anthropic's own guidance says not to treat the output as final without review. Have it build the skeleton; you own every formula before the model carries a decision.

Weak: knowing what it can't see. Claude doesn't know your revenue-recognition policy changed in February unless something tells it. Which brings us to the actual bottleneck.

The governed-data problem: pasted exports vs. MCP

Here's the uncomfortable truth about most "AI + Excel" workflows: the model quality was never the constraint. The data path was.

When you paste an export into Claude, you inherit every weakness of the export: it's stale the moment it's cut, it carries no permissions (whoever holds the chat can see whatever the export contained), and there's no way to know whether it reflects the current forecast version or the one from before Tuesday's reforecast. Multiply that by everyone on the team pasting their own exports and you get confidently written commentary on top of quietly inconsistent numbers.

Querying over MCP inverts this — our full MCP guide for finance teams covers the standard, which FP&A platforms support it, and how to roll it out. Claude asks your FP&A platform for the numbers at answer time; the platform enforces who can see what; and everyone's analysis draws from the same governed source — the same reason real-time spreadsheet sync beats CSV re-uploads for the spreadsheets themselves. This is the design behind Aleph's MCP server: Aleph stays the system of record — consolidating actuals from your ERP, CRM, and HRIS with definitions and access control — and Claude sits on top as the narrative and analysis layer. Claude drafts the commentary; it is not where the numbers live.

One prerequisite worth naming: an MCP connection is only as good as the data layer behind it. If your actuals arrive by email attachment, fix that first — our data readiness checklist covers what needs to be true before an AI layer adds value instead of noise.

Reusable prompt patterns for FP&A

Four patterns that transfer across companies. Adjust thresholds and vocabulary to your own close.

1. The framed-commentary pattern (variance work):

"Draft [commentary type] for [audience]. Materiality: [threshold]. For each material item: variance, likely driver marked [CONFIRM], forward implication. Ignore immaterial lines. Tone: [flat/direct/board-ready]."

2. The audit-sweep pattern (model review):

"Review [tab/range] for: hardcoded values inside formula ranges, formulas that change mid-row, references to blank cells, and sign-convention inconsistencies. List each finding with its cell reference and why it's suspicious. Do not fix anything."

That last sentence matters — keep the read step and the write step separate, so you approve changes rather than discover them.

3. The continuity-narrative pattern (board/investor updates):

"Using [this period's outputs] and [last period's narrative], draft [section]. Lead with [metrics]. Frame [known issue] as [agreed framing]. [Length limit]. Flag anything in the numbers that contradicts the prior narrative instead of smoothing over it."

4. The assumption-challenge pattern (forecast review):

"Here are the operating assumptions behind our FY26 plan. For each, state what would have to be true for it to hold, what recent actuals say about that, and rank the three assumptions most likely to break first. Be specific; no hedging language."

Notice what all four share: audience, thresholds, output shape, and an explicit instruction about what not to do. That last one is the most commonly skipped and the most valuable.

Security: prompt injection lives in spreadsheet cells

This deserves more attention than it gets. Anthropic's own Claude for Excel documentation warns about prompt-injection attacks that "hide malicious instructions in spreadsheet content (cells, formulas, comments, etc.) to trick the AI models into taking unintended actions" — and instructs users to only work with trusted spreadsheets, not files from external, untrusted sources such as downloaded templates, vendor files, and data imports.

The threat model is simple: a spreadsheet is a document that can contain text anywhere — a hidden column, a white-on-white cell, a comment — and an AI reading the file can't always distinguish your instructions from instructions planted in the content. Anthropic's Cowork security guidance frames the risk precisely: an attack succeeds when Claude can both read content from outside your trusted boundary and take actions that could compromise you.

Practical rules for a finance team:

• Treat inbound spreadsheets like inbound email attachments. That "benchmarking template" from a conference contact goes nowhere near a Claude session that has write access or connectors enabled.
• Separate read from write. Audit untrusted files, if you must, in a session with nothing else connected; keep model-editing sessions to workbooks your team built.
• Least privilege on connectors. Scope MCP access to what the workflow needs; an analyst drafting commentary needs read access to the BvA, not write access to the forecast.
• Keep approval in the loop. Auto-approve modes trade away your chance to catch a hijacked instruction mid-task — Anthropic flags this explicitly for Cowork.

None of this is a reason to avoid AI in finance workflows. It's a reason to run them on governed, permissioned data paths instead of a folder of files of mixed provenance — the same control instinct finance already applies everywhere else.

Give Claude governed numbers with Aleph

Everything above works better when the data layer underneath is solid. Aleph keeps your actuals, forecasts, and metrics consolidated, permissioned, and synced with the spreadsheets your team already works in — and exposes all of it to Claude through MCP, so the analysis layer runs on the numbers rather than on exports of the numbers. If you're evaluating how AI fits your FP&A stack, book a demo and we'll show you what Claude can do when it's connected to a real source of truth.