Last week I ran a workshop for the investment team at a private equity firm. One associate had uploaded a financial model to a Claude chat and asked it to rebuild the model for a new company. The output was poor, and he wanted to know why. The answer is the add-in. This post covers what Claude for Excel is, what it costs, how to get it installed, and the first workflow I would hand a finance team, which I tested on a practice model before writing this.
What Claude for Excel is
Claude for Excel is a sidebar inside Excel. It reads the workbook you have open, including the formulas, the cell references and the tab structure. You ask it a question and it answers with cell-level citations you can click. You ask it to change an assumption and it keeps the formula relationships intact so the downstream cells recompute. It can trace where a number came from, find the source of a #REF! error, populate a template, and sort, filter or format without you leaving the sheet.
Skills run inside it too. A skill you built in Claude Desktop shows up when you type a forward slash in the Excel sidebar, and Anthropic ships a few Excel-specific ones of its own.
The same add-in covers PowerPoint and Word. You install it once and it covers all three, and a conversation can span your open workbook, deck and document.
Which plans include it
It is included with the Pro, Max, Team and Enterprise plans at no extra charge. It is not on the Free plan. Usage counts against the same limits as your Claude account, so a long session in Excel draws from the same allowance as a long session in the app. As of October 2026 it is generally available, no longer a beta.
How to install it
For yourself: open the Claude for Microsoft 365 listing on Microsoft AppSource, select Get it now, then open Excel and sign in with your Claude account. On Windows the add-in is under Home, then Add-ins. On a Mac it is under Tools, then Add-ins.
It runs on Excel for the web, on Windows with a Microsoft 365 subscription (build 16.0.13127.20296 or later) and on Mac 16.46 or later. It does not run on Excel 2016 or 2019 perpetual licenses, on iPad or on Android.
If you open Add-ins and the store is empty, your IT team has turned off a Microsoft setting called "Let users access the Office Store". That is a Microsoft 365 admin setting, and your admin can either turn it on or deploy the add-in centrally to named users. At the firm I worked with last week, each add-in is enabled per person through an IT ticket, and that is the normal pattern at a regulated firm. Ask before you assume it is broken.
Why uploading a model to a chat fails
Back to the associate's question. Uploading a workbook to a chat and asking for a rebuild is the hardest version of the request. A chat upload flattens the workbook. Claude sees the values, not the live formulas, so it is reconstructing dozens of interlinked calculations from a picture of the answers. That is where the errors come from.
The add-in reads the live workbook. So the request to make is "hollow out a copy, and prove the copy still works". Claude does the mapping and the clearing, and Excel does the proving. You keep every judgment call.
The workflow: map, hollow out, backtest, run
Work on a copy with the deal material taken out, never on the live file. Then run these in order in the Excel sidebar.
Step 1: map it
Map this workbook. For each tab, give its purpose, its inputs, and what it feeds. Then list every number typed inside a formula, every external link, every hidden tab, every merged cell, and every formula that breaks the pattern of its row. Give me the tab and cell reference for each one. Change nothing.
This costs you nothing and it is the documentation the model never had. It also surfaces the things that will bite you in step 4: a growth rate typed into a formula instead of pulled from the inputs tab, a hidden tab with an old assumption on it, a year where someone typed over the formula.
Step 2: hollow it out
Turn this workbook into a blank template. Before you clear anything, copy the current input values into a block below them on the same tab, labeled Previous company, and leave that block alone. Then keep every formula exactly as it is and clear only the live input cells. List every cell you cleared with its tab and its row label, so I can check the list before I accept it. Where you are unsure whether a cell is an input or a calculation, leave it alone and tell me which ones you left.
Read the list it gives you against the inputs tab, one row at a time. This is the one step where your judgment is required, and it is where a finance person decides whether to trust the method.
Step 3: backtest it, in Excel, with no prompt
Paste the old company's inputs back into the template. On a Checks tab, put the original model's key outputs in one column as values, the template's live outputs beside them, and a difference column. Every difference has to read zero. If one does not, the template lost something and the map from step 1 tells you where to look.
Claude does not get to grade its own work here. Excel proves the template.
Step 4: run the new company
Paste the new company's inputs in, then:
These inputs are for a different company from the one this model was built on. The previous company's inputs are in the block labeled Previous company. Flag any new input that looks out of range against them and say why. Then check the workbook for anything left over from the old company: a label, a note, a date, a hardcoded figure.
Once it has worked twice, ask Claude to turn steps one to three into a skill called model-to-template, with the rule that it clears only input cells and lists everything it cleared. Tell it not to write any company name, deal name or figure into the skill, so it can run on any model.
What happened when I tested it
I built a synthetic five-year three-statement model with a Checks tab and planted eight faults in it: three rates typed inside formulas, one year of capex typed over its formula, one formula written with SUM where the rest of the row used addition, a merged title cell, an unlabeled row, and a hidden tab holding a superseded gross margin. Every fault was set to equal the input it bypassed, so the model still balanced. A test that changes its own answers cannot also demonstrate step 3.
Step 1 found all eight, with cell references, and invented nothing. It also pointed out that the backtest compares the model with itself and so cannot catch a typed-in number while the inputs are unchanged. That is true. It is why step 1 exists.
Step 2 cleared the eighteen input cells and nothing else, listed every one with its row label, and told me which typed values it had left alone and why. It re-read all 191 formulas afterward to confirm none had changed.
Step 3 read zero on all six outputs.
Step 4, with a second company's inputs in, named all four of the old company's typed-in numbers and what each one was doing to the new figures: Year 4 revenue understated by about 2,000, Year 2 inventory inflated by about 3,300, Year 5 capex frozen at the old value. The balance checks still read zero with those four numbers in place, because the model balances either way. The map is the only thing that catches them.
Anthropic's own cautions
Anthropic's own guidance is to use Claude for Excel only with trusted spreadsheets. A file from outside the firm can carry hidden instructions that steer the add-in, and Anthropic has documented cases where that led to data being extracted or changed. So this is for your own models. A banker's file or a target's data room export does not qualify. When Claude proposes a risky change, the add-in asks you to confirm. Read those.
And Anthropic does not recommend it for calculations someone else will rely on without a person checking them. The backtest in step 3 is that check, done in Excel, before anyone sees the output.
The same list makes a model Claude can read
Someone in the same workshop asked for a way to keep Excel formulas easy to check. It turns out to be the same list. A model built this way is quicker for a colleague to pick up and it is the kind Claude reads most reliably, because both of them are looking for the same thing.
- Inputs live on their own tab, shaded, one value per cell, each labeled with its units.
- No typed numbers inside formulas. =B5*1.05 hides an assumption, so the 1.05 belongs in an input cell.
- One formula per row, copied straight across every column.
- Time runs left to right, one period per column, with the same columns on every tab.
- Every row has a label.
- Avoid merged cells, hidden tabs, external links and large array formulas.
- Every hardcoded number carries a note saying where it came from.
- A Checks tab that reads zero when the balance sheet balances and the cash ties, and a README tab saying what each tab does.
Where this fits
This workflow came out of a Role-Specific Workshop for an investment team, the part of a program where each team brings its own work and leaves with something built on it. If your deal team is already on Claude and the Excel add-in is the piece nobody has set up, that is a short conversation. Start with how we work with private equity firms, or book 30 minutes and bring a model.
Nicole Patten is the founder of Elevate Online and runs a Claude-specific training practice. She spent 7 years at Google as a Senior UX Engineer before dedicating her career to helping teams use AI responsibly and effectively. 100% of her business runs on Claude.