Claude in Excel is good with formulas and blind to your macros

What does Claude in Excel actually do?

The tool is an add-in, so it opens as a panel down the right hand side of Excel. It can see the workbook you have open, and that includes every tab in it. You type a question or an instruction in plain English. It works directly on the file in front of you instead of a copy you pasted somewhere.

The citations are the part that matters for real work. The answer carries a cell reference you can click when it answers a question about a number, and Excel jumps to that cell. It might tell you the gross margin line is driven by an assumption three tabs away. You can go and look at the assumption instead of taking its word for it.

Anthropic’s documentation groups its work into four jobs. These match the four things people ask a spreadsheet person to do. It explains how a number was derived, and it changes a value while keeping the formula relationships intact. This means the cells downstream recalculate properly instead of being flattened into hard numbers. Tracing errors is the third one, so it finds the root cause of something like a #REF! or a #DIV/0! rather than just telling you the cell is broken. It also builds a new model from a description, or it fills in a template you already have.

It also does ordinary Excel operations directly. It handles sorting, filtering, editing PivotTables, applying conditional formatting, and creating data validation dropdowns. You ask for those in words and it performs them. That is the difference between an assistant that writes you a formula and one that operates the application.

What stops it from wrecking the file?

Two things are worth knowing before you let it edit anything. It has overwrite protection, so it warns you before it writes over existing data. It also has auto-compaction. This folds a long conversation into a new one automatically, so the session does not run out of room part way through a job.

There is a settings field called Instructions. This is where you put the preferences you want applied to every conversation in Excel, such as thousand separators or bold column headers. Those instructions apply to that application only. What you set in Excel does not carry into Word or PowerPoint, which is easy to forget when you wonder why the formatting came out differently.

Why will it not install, and which Excel do you need?

Most people stop here when the setup fails, because the failure is quiet. The add-in either appears in Excel or it does not, and Excel does not explain why. Anthropic publishes the exact requirement. You can check it in one minute instead of guessing.

The tool runs on Excel on the web. It runs on Excel on Windows with a Microsoft 365 subscription at build 16.0.13127.20296 or later, and on Excel on Mac at version 16.46 or later, build 21011600 or later. You can read your own build on Windows under File, Account, About Excel. The number is written on the first line.

The list of what it does not run on is more useful. Excel 2016 and Excel 2019 on a perpetual or volume licence are out. This covers a large number of office machines that were bought once and never moved to a subscription. Excel on iPad is out, because Anthropic says the add-in needs SharedRuntime. That is the Office feature that lets an add-in keep running in the background while you use the app, and the iPad does not provide it. Excel on Android is out. Older Microsoft 365 builds below that SharedRuntime threshold are also out, even though they are subscription copies.

Installing it yourself takes three steps. You open the Claude for Microsoft 365 listing on Microsoft AppSource and choose Get it now, then you open Excel. You activate the add-in and sign in with your Claude account. The add-in lives under Home, Add-ins on Windows, and under Tools, Add-ins on Mac.

An admin usually deploys it in a company through the Microsoft 365 Admin Center under Settings, Integrated apps. There is one documented trap there. The organisation might use Privileged Identity Management, which is the Microsoft system where an admin activates their admin role only when they need it. The Integrated apps page does not recognise a role activated that way, so the deployment fails. Anthropic names it as a known Microsoft issue with the tracking ID 11126536, and the workaround is to deploy from an account whose role is permanently assigned.

What will it not touch, even once it is installed?

Anthropic lists two unsupported capabilities, and these are things you usually discover halfway through a job. Macros and VBA operations are not supported. Data tables are also not supported. Data tables here means Excel’s what-if Data Table feature under the Data menu, which is the grid that recalculates a formula across a row and a column of input values.

The macro limitation decides it for most people. A long lived business workbook usually has a button somewhere that runs a routine, and that routine does something nobody has written down. The sidebar will happily discuss the cells, but it will not see the code that fills them. This is the answer if that is the workbook you hoped to hand over, and it is better to know now than after wasting an afternoon.

There is a second list, and Anthropic is direct about it. It says the add-in is not recommended for final client deliverables without human review, for audit-critical calculations without verification, or for models holding highly sensitive or regulated data without proper controls. That is a vendor telling you not to ship its output unchecked. This is unusual, and I take it at face value.

The checks I described in checking whether ChatGPT or Copilot got your Excel numbers right apply here without a single change. A sheet that computes a wrong answer with no error on screen is the failure mode I keep coming back to in spreadsheet audits and silent errors. An assistant that edits cells for you makes that easier to create.

Where does Microsoft’s Copilot in Excel still win?

It would be easy to call this the better assistant, and I think it often is for the specific job of reasoning through a financial model. Microsoft’s Copilot in Excel wins on the plainest thing of all. It wins on where it runs and what you have to pay for it.

Microsoft’s FAQ page for Copilot in Excel says it is available across Windows, Mac, the web, iPad, iPhone and Android. The edit, chat and plan modes are available on Windows, Mac and the web. It comes with a Microsoft 365 Personal or Family subscription on an AI credits plan, with Microsoft 365 Premium, or with a commercial Copilot subscription. That is a subscription a lot of people already pay for.

Read the last row twice, because both vendors say the same thing in their own words. Neither of them claims the output is correct. That is the honest position, and it decides how you use either tool. You use it as a fast first draft that a person checks.

The pattern is familiar if you have used the Copilot inside Power BI. I went through where that one helps and where it stops in the limits of Power BI Copilot, and the shape is the same here. It is strong on explaining and drafting, but weak the moment the model depends on something the assistant cannot see. Microsoft has also moved these features around before, and that is worth remembering when you build a process on one. The retired COPILOT function in Excel is the example I keep pointing at.

What do connectors and Skills change for an analyst?

A connector is a link to something outside the file. It lets the assistant use information that is not in the workbook you have open. You click the plus button under the chat box in the add-in and pick Connectors. Anthropic names S&P Global, LSEG and Daloopa as the common ones. That tells you plainly who this was built for, along with whatever custom connectors your own organisation has turned on.

Skills are reusable recipes for a task. You enable them in your Claude settings and they apply automatically when they are relevant. You can also type a forward slash in the sidebar to pick one directly. The useful part for an analyst is that a Skill holds your house rules. This includes the way your company lays out a model, the checks you always run, and the format a client expects. This is the same instinct behind the AI tools I actually keep for analyst work, which is to set a process up once rather than retype it.

There are two warnings here, and both are in Anthropic’s connectors and Skills page. The first is that a custom connector can introduce security risks, and Anthropic says this directly. Someone has to look at what a connector can reach before it is switched on. The second is about how your company connects. Connectors and Skills may not be available at all if your company routes AI traffic through Amazon Bedrock, Google Cloud Vertex AI, Azure AI Foundry or an LLM gateway rather than through individual Claude accounts.

That second limit catches people out because the add-in still installs and still works. The capability you read about is simply missing in that mode, and the reason sits with how your admin wired it up.

What should you not put in it?

I did not expect to find this next part on a vendor page. It is the reason I would think carefully before using the sidebar on files that arrive by email. Anthropic warns that you should only use Claude for Excel with trusted spreadsheets, because a file from an external source can carry hidden instructions.

Hidden instructions mean ordinary text sitting in the workbook in a cell or a note. The text is written to be read by the assistant rather than by you. The assistant reads the workbook, so it reads that text too, and it can be worded to look like a request from you. This is called prompt injection. It is hard to fix because the assistant has no reliable way to tell your instruction apart from text that is pretending to be one.

Anthropic says its own testing found scenarios where the add-in could be manipulated. It could extract sensitive information, modify critical data, or perform destructive actions if it was allowed to act without verification. The defence it ships is a confirmation step, so it asks you first when it proposes something risky. The confirmation box is not a formality. It is the control, and it only works if you read it.

What happens to the workbook contents afterwards?

Inputs and outputs are deleted on Anthropic’s side within 30 days of being received or generated. They keep a cache for a few hours after deletion, so you can still reach context from a workbook you just closed. Your conversation history is different. It is kept locally in your browser using IndexedDB, which is the browser’s own on-disk storage. Anthropic says those conversations are not stored on its servers and they are not synced across devices.

That is good for privacy, but it is awkward in practice. The thread does not follow you if you move to another machine. Clearing the browser’s site data takes your history with everything else. Reinstalling the add-in does not remove it, and that is the one case where it survives.

I point out the governance gap to anyone handling client or regulated data, and Anthropic states it plainly. Claude for Excel does not inherit custom data retention settings an organisation may have set, and its activity is not included in Enterprise audit logs. Enterprise organisations with the Compliance API enabled do get these sessions covered, and that coverage is in public beta. You should check that before the file is opened if your answer to a client about where their numbers went has to come out of an audit log.