AI In Excel, Tool By Tool (And The License You Need)

There is a model picker in the top right of Excel’s Copilot pane, and Anthropic’s models are in the list. Which means the Claude versus Copilot argument finance teams have been having for two years is, on the right subscription, a dropdown.

Most people will never see it. It needs a commercial Copilot subscription and an admin toggle buried in the Microsoft 365 Admin Center called “AI providers for other large language models”, which is not a phrase anybody goes looking for on a Tuesday.

That is the pattern with AI in Excel. Every route in works, and each one is gated behind a different subscription that nobody writes down, so you find out at the point you were about to demo it.

So this page goes tool by tool and names the gate each time. Copilot, the =COPILOT function that stops working on 14 September 2026, Claude for Excel, ChatGPT, Python in Excel and Power Query. Then a list of what to reach for depending on the job in front of you.

The five ways AI gets into Excel

Two of them ship inside the product. Copilot sits in a side pane and can edit the workbook rather than only talk about it, and Power Query has had pattern detection working on your columns since long before anyone called it AI.

The other three you bring in yourself. Claude for Excel is an add-in from Anthropic that reads the open workbook. ChatGPT and its peers work in a browser tab through copy and paste, which sounds primitive and is the route most people get real work out of. Python in Excel runs Microsoft-hosted Python inside a cell, and it is the one that surprises people when they find it.

They are not interchangeable. Each one is good at a different part of the job, and three of them will refuse to open at all on the license your employer probably issued you.

Copilot in Excel, and the model picker

Copilot is a chat pane that understands both your data and how Excel works. Ask it in plain English and it writes the formula, builds the PivotTable, or restructures the sheet. It is generally available across Excel for Windows, Mac, the web, iPad, iPhone and Android, though what it can edit varies by platform.

The part worth your attention is the model picker. Microsoft’s own documentation says Copilot in Excel supports the latest Anthropic and OpenAI models, switched from a menu in the upper right of the pane. Two conditions apply and both catch people out. You need a commercial Microsoft Copilot subscription, and your admin has to allow “AI providers for other large language models” in the Microsoft 365 Admin Center.

Your choice also lasts one session. Close Excel and Copilot goes back to the Microsoft default, so if you tested something on a Friday with a different model and it worked, check the picker before you blame the prompt on Monday.

The agentic side of Copilot, which plans a multi-step change and then makes it, is generally available in Excel for the web, Windows and Mac. It needs a Microsoft 365 Personal or Family subscription with AI credits, a Microsoft 365 Premium subscription, a commercial Microsoft 365 Copilot subscription, or a Copilot Chat-eligible business or enterprise plan. I put it through real work in Agent Mode in Excel, and the longer verdict on Copilot generally is in Copilot for Excel. Microsoft also shipped Copilot finance skills into Excel in June 2026, behind a settings button most people never click.

Analyze Data, the free one everyone forgets

Analyze Data has been on the Home ribbon for years, it costs nothing extra, and it will surface trends and outliers in a clean table without a single prompt. It is worse than Copilot at everything except the price. Point it at a tidy range before you go asking for a license.

The =COPILOT function stops working on 14 September 2026

For a while you could drop a prompt straight into a cell. =COPILOT("Categorize this expense", A2), filled down a column, was a genuinely useful trick for classification work.

Microsoft ends it on 14 September 2026. Their support page states plainly that starting September 14, 2026, the COPILOT function will no longer be available, and points everyone at Copilot in Excel instead. While it lasts it needs a Copilot add-on license on a work or school account, or a Microsoft 365 Premium subscription on a personal one.

If you built anything on the formula version, that is a migration with a fixed date on it. Anything in a saved workbook that a colleague refreshes in October will come back as an error, and the workbook will not tell you why.

Claude for Excel

An add-in that sits in a side panel and works on the workbook you have open. Anthropic lists it for the Pro, Max, Team and Enterprise plans, checked 28 August 2026. If you tried it in the spring and hit a tier you were not on, it is worth another look.

Where it wins is explanation. It writes the formula and then tells you what it did in language a controller can review, which matters when the number ends up in a board pack and somebody asks how it was derived. It also chews through plan usage faster than anything else here, so I save it for the model that has to be right rather than the quick tidy-up. Full walkthrough in Claude for Excel, and the head-to-head is in ChatGPT vs Claude vs Copilot for Excel.

ChatGPT, and the route that costs nothing

Paste the block, describe the sheet, read the answer back. No add-in, no license, no admin ticket. The catch is that the model cannot see your file, so you have to hand it the shape yourself: sheet names, the header row, roughly how many rows, and one example of a value that is messy.

Do that and it performs about as well as the paid routes on formulas and cleanup logic. Skip it and you get a confident answer built on columns that do not exist. More in ChatGPT for Excel, and if you want the wider comparison across nine of these, the best AI for Excel has the scoring.

Python in Excel

Type =PY in a cell and you get Python running on Microsoft’s servers against your worksheet data, with pandas and matplotlib available. For anything statistical, or for a chart Excel refuses to draw properly, it is the cleanest option on this page. Ask a model to write the block and paste it in.

Check your license before you plan around it. Microsoft supports it on a paid Microsoft 365 subscription that includes the desktop apps, and free or perpetual consumer licenses do not qualify. The walkthrough is in Python in Excel.

Power Query, where the unglamorous wins are

Nobody films a demo about column type detection. It is also where most of the hours go, because the model that writes a beautiful formula still cannot do anything with a file where the header row starts on row 7 and the amounts are text.

The pattern that works is to let AI write the transformation logic and let Power Query hold it, so it runs the same way next month without anybody retyping a prompt. AI in Power Query covers the setup, Power Query data cleaning is the full workhorse guide, and the Power Query tricks I use most are worth ten minutes on their own.

What to reach for, by the job in front of you

The tool matters less than matching it to the task. Here is how I split it.

Cleaning a file somebody else built

Power Query, with a model writing the steps. Do it once and it repeats. If the file is a one-off and you will never see it again, paste the mess into a chat window and move on with your day. Start at automating data cleaning.

Writing a formula you half remember

Any of them. The free chat window is fine for this. Describe the columns and ask for the formula plus one sentence on why it works, so you can review it rather than trust it. Creating formulas in Excel covers the mechanics, and LAMBDA functions are where this gets genuinely powerful once you can read what comes back. Checking AI Excel formulas takes about ninety seconds, and it is the step people skip.

Building the chart

Every model here is poor at chart formatting and confident about it. Ask for the data shaped correctly, then format it yourself, or use Python in Excel and skip the chart engine. I go through my whole process in building Excel charts with AI.

Writing the commentary

Claude, by a distance, and it is the highest-value use of AI on this list because it is the task you repeat every single month. Give it the variance table and your own two-line summary from last month as the style reference. The starting point is variance analysis using rate and volume, because a model cannot write commentary on a split it was never given.

Making it run without you

This is where you leave the chat window behind. Excel automation with Office Scripts is the route I use, because a script runs the same way in August when somebody else is covering for you. Excel automation is the wider view, and if the sheet lives in Google instead, Google Sheets automation is the equivalent.

Where this breaks

Check one row by hand. Pick the biggest number, work it out on a calculator, and see whether the output matches. If it does, the rest of the column is almost always fine, and if it does not, the setup was wrong rather than the arithmetic.

The failure mode nobody warns you about is a plausible wrong answer. A model asked to categorize 400 expense lines will categorize all 400, including the twelve it had no basis for, and it will not flag which twelve. Ask it to mark anything it was unsure about and it usually will.

Keep the prompts you reuse somewhere findable. When somebody asks in March how the October classification was done, “I asked an AI” is a bad answer and a saved prompt is a fine one.

And know what leaves the building. Copilot and Claude for Excel both operate under enterprise terms your organization has signed, and a free consumer chat window has not. Pasting a customer-level P&L into a personal account is the version of this that ends up in a policy meeting. If yours has no policy yet, that conversation lives under AI, careers and governance.

Where to go next

If you are choosing between tools rather than tasks, AI tools for finance has the comparisons with a verdict attached. Finance automation is where the spreadsheet work meets the rest of your month, Power BI picks up when the file outgrows Excel, and AI agents for finance is the part that keeps working while you sit in a meeting. If you want the model working on files you are not sitting in front of, Excel MCP servers is the step past the add-ins. And if you cannot tell whether a job belongs in Excel or Power BI, Excel and Power BI techniques draws the line.