Claude for Excel: What It Does, Where It Fails, How to Use It

Claude for Excel is an add-in that runs in a sidebar inside your workbook. It reads what is open, answers questions with citations you can click through to specific cells, changes assumptions without flattening the formulas underneath, and traces errors back to their source. This guide covers installation, the builds it runs on, what it will not do, and the review discipline that has to sit around it before a workbook goes anywhere important. It is the spreadsheet chapter of Claude for Data Analysis.

Install it, or get it deployed

Individual install is three steps. Open the Claude for Microsoft 365 listing on Microsoft AppSource, select "Get it now," then open Excel, activate the add-in, and sign in with your Claude account. The add-in ships as part of Claude for Microsoft 365, so the same install puts the PowerPoint, Word, and Outlook add-ins in place too.

For an organization, an admin deploys it from the Microsoft 365 Admin Center: turn on "Let users access the Office Store" under Settings, Org Settings, User owned apps and services, then go to Settings, Integrated apps, Add-ins, search for Claude for Microsoft 365, and assign it to the org or to specific groups. Two wrinkles are worth knowing before you file a ticket. Tenants that block the Office Store can deploy with a custom manifest XML file instead. And if your organization uses Microsoft Entra Privileged Identity Management for admin roles, the Integrated apps page does not recognize PIM-activated roles and deployment fails; Anthropic documents this as a known Microsoft issue and the workaround is to deploy from an account whose role is permanently active.

Organizations that route AI traffic through Amazon Bedrock, Google Cloud Vertex AI, Azure AI Foundry, or an LLM gateway can deploy the add-in without individual Claude accounts, with models coming from that platform.

Supported builds, and the exclusions that catch people

As of September 2026, Anthropic lists three supported environments: Excel on the web, Excel on Windows with a Microsoft 365 subscription on build 16.0.13127.20296 or later, and Excel on Mac version 16.46 or later (build 21011600 or later).

The exclusions are the part people hit. Excel 2016 and 2019 perpetual or volume licenses are not supported. Excel on iPad is not, because the add-in requires SharedRuntime and iPad does not provide it. Excel on Android is not. Older Microsoft 365 builds below the SharedRuntime threshold are not. If you work somewhere that standardized on a perpetual license, this is a blocker rather than a bug.

Plan gating is simple: generally available on Pro, Max, Team, and Enterprise, not on Free. Which tier you want for spreadsheet-heavy weeks is in Which Claude Plan (and Model) Should a Data Analyst Pay For?, and the prices are on Claude plans compared (2026).

Cell-level citations are the whole point

Most AI spreadsheet features answer in prose. This one answers with references. Ask "walk me through how the revenue number in cell C42 is calculated" and the reply cites the cells it read, each one clickable to navigate straight there. That turns a claim into something you can check in about ten seconds, which is the only reason it is usable on a model you are responsible for.

Use it that way deliberately. On a workbook you inherited, the first three prompts should all be questions, never edits:

Walk me through how the terminal value on the DCF tab is calculated,
citing cells. Then list every hardcoded number in the calculation chain
and tell me which tab it came from.

If a citation points somewhere you did not expect, you have found either a modeling error or a misreading, and both are worth knowing before you change anything.

Building and debugging models

Beyond questions, the add-in does four kinds of work. It updates values while keeping formula relationships intact, so "change the discount rate to 8% and update dependent calculations" edits the input and lets the model recompute. It debugs, tracing a #REF! in a summary tab or a #DIV/0 in a single cell back to a root cause across a multi-tab workbook. It builds, either populating an existing template ("populate this LBO template with a $500M purchase price and 6x leverage") or generating a model from a description such as a three-statement model from a trial balance. And it performs native Excel operations on request: sorting, filtering, editing pivot tables, conditional formatting, and data-validation dropdowns.

For auditing formulas rather than building them, the spreadsheet-formula-auditor skill gives you a repeatable pass over a sheet you did not write, and it works the same way in the add-in as it does in chat.

What it will not do

Two capabilities are explicitly unsupported: data tables, and macros or VBA operations. If your sensitivity analysis is built on data tables, Claude cannot drive it. If your workbook's logic lives in VBA, that logic is invisible to the add-in.

There are also three situations Anthropic says the tool is not recommended for: final client deliverables without human review, audit-critical calculations without verification, and models containing highly sensitive or regulated data without proper controls. Read that as a scope statement rather than boilerplate. It is a fast, cited assistant for a model in progress, and it is not a signer of anything.

Overwrites and the review pass before a deliverable

The add-in warns before overwriting existing data. That protection is real, but the thing to check after an editing session is subtler than a lost tab: it is a cell that used to hold a formula and now holds a number. Scan the range Claude touched, confirm the formulas are still formulas, and re-derive one changed output by hand.

Build the rest of the review into a habit before anything ships:

• Re-derive one number end to end, without the add-in.
• Check that every citation in the explanation resolves to the cell you expected.
• Ask Claude to list every assumption it made and every cell it changed, then verify that list against the workbook rather than against its summary.
• Sanity-check the direction of the result. If a higher discount rate raised the valuation, something is wrong regardless of how confident the explanation reads.

The general version of this checklist, including the failures that survive a casual read, is How to Check an AI Data Analysis.

Connectors, skills, and the other Office apps

Connectors pull external context into the workbook; Anthropic names S&P Global, LSEG, and Daloopa as examples. Skills you have enabled in your Claude settings apply automatically while you work, which means a skill that encodes your team's modeling conventions travels into Excel without being installed separately there. Persistent instructions are per app: open Settings in the sidebar and set formatting conventions like "format numbers with thousand separators," and they apply to every Excel conversation but not to Word or PowerPoint.

Cross-app mode extends one conversation across Excel, PowerPoint, Word, and Outlook, which is how a model becomes a summary deck or a memo without copy and paste. It only reads and writes files that are already open, it cannot create, open, close, or switch files, and it is unavailable when you connect through a third-party platform.

Two operational details round it out. Long conversations are auto-compacted into new conversations so you do not run out of context, and usage counts against your normal Claude limits. Chat history lives locally in your browser via IndexedDB rather than on Anthropic's servers, so it does not sync across devices, and add-in activity is not included in Enterprise audit logs or the Compliance API.

If your organization already pays for Microsoft Copilot seats, run both against a model you know well before choosing. Copilot is broader and already licensed; the citation trail is what Claude for Excel is selling.