For two years, the gap between “AI is impressive” and “AI is useful at my job” had a very specific shape for most business teams. Claude could describe the analysis beautifully. It could not hand you the spreadsheet.
You’d get a lovely markdown table, copy it into Excel, watch the formatting collapse, fix it manually, and quietly conclude that the demo was better than the reality. That was a real limitation, and it’s the one that’s changed.
Two different things, often confused
There are two separate capabilities here and people conflate them constantly, which leads to buying or expecting the wrong one.
1. File output — Claude produces a real Office file. Built-in file skills let Claude generate genuine binary .xlsx, .pptx, .docx, and .pdf files. Not markdown pretending to be a table, not CSV you’ll have to reformat — the actual file, which you download and open in Excel or PowerPoint like anything else. There’s no code execution to set up and no middleware. You describe what you need; Claude works out that a file is the right answer and produces it.
2. Claude for Excel — Claude works inside your workbook. A separate add-in, available on Pro, Max, Team, and Enterprise plans, running in Excel on the web, on Windows, on Mac, and on iPad. This one isn’t about producing a file; it’s about interrogating and editing one that already exists.
The rough division: output when you’re making something new, the add-in when you’re working on something that exists. Most people need both, at different moments.
What file output changes day to day
The honest test of this feature is how boring the examples are. Good.
- “Take these four regional exports and build me one workbook with a summary sheet, a tab per region, and the totals cross-checked.”
- “Turn the findings in this document into a twelve-slide deck with a summary up front.”
- “Produce the client report as a PDF using the structure in last quarter’s version.”
- “Build me a template for the team to fill in — dropdowns on the status column, and lock everything except the input cells.”
That last one is worth dwelling on. Dropdown validation, protecting sheets, locking specific cell ranges — these are the fiddly, genuinely tedious things you do when producing a template for non-technical colleagues to fill in. They’re also exactly the kind of task where you know what you want and can never remember which menu it’s under. Describing it is faster than doing it, which is the whole promise of this technology arriving in a form you can actually use.
Pair this with Cowork and it gets more interesting still: point Claude at the folder holding your raw exports, and the output file is built from your real data rather than from what you pasted in.
What Claude for Excel changes
The add-in is aimed at a different problem, and one specific feature is the reason it exists: cell-level citations.
Ask a question about your workbook and Claude answers and tells you which cells the answer came from. If you’ve ever inherited a financial model and spent an afternoon tracing where a number originated, you already understand why this matters more than it sounds. The recurring, genuinely expensive question in spreadsheet work is not “what’s the number?” — it’s “how did this number get here, and can I trust it?”
Beyond that, it can:
- Update assumptions while keeping formulas intact — change an input and have the model’s structure survive
- Debug errors — find why a formula is returning something wrong
- Build financial models from a description
- Apply native Excel operations — sorting, filtering, conditional formatting — as operations in the workbook rather than as instructions for you to follow
For finance, FP&A, and anyone who lives in a model, this is the closest thing to a colleague who has read the entire workbook and remembers all of it.
The limits, stated plainly
This section matters more than the capability list, because deliverables are involved.
No data tables, macros, or VBA. If your workbook’s logic lives in VBA, that layer is out of scope. Claude can reason about the spreadsheet around it; it won’t write or modify the macro code.
Not for client deliverables without review. Anthropic’s own documentation says this explicitly, and it’s the right posture. The output is good enough to look finished, which is exactly what makes unreviewed output dangerous.
Not for audit-critical calculations without verification. Same principle, higher stakes. If a number is going to be audited, a human verifies it. Cell-level citations make that verification much faster — they don’t remove it.
Not for highly sensitive regulated data without proper controls. If you’re in a regulated sector, this is a conversation with your compliance team and your account team before it’s a workflow, not after. Our security guide covers the controls available.
None of these are surprising. They’re the same limits any competent person would put on a fast new colleague: excellent, quick, and not yet the one who signs off.
The security point most people skip
Here is a genuine risk that no setting fixes.
A spreadsheet is not inert content. Text inside a workbook can be crafted to manipulate an AI reading it — to extract information from elsewhere in its context, or to steer it into modifying records. Anthropic advises directly against applying Claude for Excel to spreadsheets from untrusted external sources.
So the working rule is simple: treat external files like email attachments. A workbook a vendor sent you, a template downloaded from a forum, a file from a counterparty — don’t point Claude at it directly. Copy out the data you need into your own file and work there. It takes a minute and removes the whole class of problem.
How this reshapes the Copilot question
The strongest argument for Microsoft 365 Copilot was always proximity: our work happens in Excel, and Claude isn’t in Excel.
That’s now partly answered. Claude for Excel is in Excel, and cell-level citation is a capability Copilot’s side panel doesn’t match for model-tracing work.
But be fair about the rest of it. Copilot has broader in-app presence — Word, Outlook, PowerPoint, Teams, all of it — and if your team’s bottleneck is email volume and meeting load rather than analysis, that breadth is worth more than depth in one application. The comparison is genuinely closer than it was, and it now turns on which Office app your pain actually lives in. Full comparison here.
A sensible way to start
- Produce something new first. Ask for a real file — a deck from a document, a workbook from a few exports. It takes two minutes and it’s the moment most people stop thinking of this as a chat toy.
- Then interrogate something existing. Install the Excel add-in and point it at a model you didn’t write. Ask “where does this figure come from?” and watch the citations. That’s the moment the finance people get interested.
- Automate the recurring one. Whatever file you rebuild every month — that’s the third step, and it’s where the hours actually come back. Cowork can now run that on a schedule without your laptop being open.
- Name a reviewer. Before anything produced this way goes to a client, a board, or an auditor, one named human checks it. Decide that now, while it’s easy, rather than after the first embarrassing number.
The shift here is smaller than the hype and bigger than the scepticism. Claude didn’t become an accountant. It became able to hand you the file — which was, for an enormous number of people, the only thing standing between “interesting” and “useful.”