Six months ago, you might have asked Excel's Copilot to do a calculation and thought, "Well, I guess that's it." Have you stayed with that impression until today?
Let me give you the conclusion first: Today's Excel Copilot is a completely different tool from six months ago. It now has a place to remember your preferences. It has a place to write rules for specific files. You can now choose the AI brain that runs inside it. And with a simple, vague request like "Analyze the whole thing," it can now build everything from aggregations to graphs.
As someone whose job is to promote Copilot within my company, the words of those who come to me for advice have changed over the past six months. Before, it was "What can it do?" Now, it's "There are so many features, I don't know where to start." The change is so fast that for those who stepped away for a while, the big picture is no longer visible. That's why this article summarizes everything Excel's Copilot can do as of September 2026 in one place. It includes copy-pasteable prompt text for everything. You don't have to read it all now. Save it and open only the parts you need when you need them. That's how I've written it.
I also promise these changes by the time you finish reading:
- You'll be able to handle data you're handed on the spot with a simple "Analyze this."
- The time spent manually fixing the appearance of tables and graphs will disappear.
- You'll be able to draw the line yourself on how much to leave to Copilot and where to verify with your own eyes.
Some of the features listed here have been released as individual articles before. This one piece rearranges them into a complete picture so you can see at a glance what is possible now. Features will continue to increase, so I've included the date in the title. By the time you are reading this, there may be more things not written here.
The target audience is those using a paid Microsoft 365 Copilot license at work. I am an AI promotion manager at a major Japanese company. Everything written here has been confirmed to work on my actual machine. Conversely, even if something has been announced, if it doesn't work in my environment yet, I haven't included it.
No programming is required at all. Everything is just asking in plain Japanese (or English).
Before the main topic, a 30-second check. Open the Copilot window in Excel and see if the word "Auto" appears at the top. If it does, you can use everything up to Part 3 (Choosing the AI Brain). If you don't see it, Parts 1, 2, and 4 will still work. If the Copilot button itself isn't on the ribbon, it's likely a license hasn't been assigned, so check with your company's IT department.
Part 1: 5 Things You Can Press Today (Zero Setup)
First, things you can use today without any preparation. I think the biggest misconception about Excel's Copilot is that "it won't work unless you give it perfect instructions." In fact, it's the opposite; there are even situations where Copilot performs better when you ask vaguely. These five are representative of that.
The logic is simple, like a hospital. A patient doesn't need to explain symptoms in medical terms. They say "I feel unwell," and the doctor narrows it down through questioning. Data analysis and formula decoding are the same; you just bring the problem. Sorting it out is their job.
Just remember one location first: the "Allow editing" toggle at the top of the Copilot window. This is the switch for whether to let Copilot take action. If it's "Chat only," it's a consultant; if it's "Allow editing," it becomes a worker that adds sheets and builds tables. All of Part 1 in this article is used in the "Allow editing" state.
1-1. Outsourcing Data Analysis
Have you ever frozen in front of data you were handed with "Analyze this," not knowing where to start? The answer is: you can just throw it while frozen.
Raw survey data
1Please analyze the entire data in this sheet. Tell me what you noticed.
With this one sentence, a summary table and graph are built on a new sheet, and the chat returns insights like "While this item is highly rated, there are several points noted here." Surveys, sales lists, attendance, inquiry history—anything with rows and columns works. The premise is that the workbook is saved in OneDrive or SharePoint, and it's enough if the first row is a header.
When I tried it with a training survey (50 respondents, 8 evaluation items, free text at the end), a sheet called "Training Survey Analysis" was added in a few minutes, arranged like this: First, 4 lines of conclusions/key points. "Materials, instructors, and practical utility were high, with all ratings at 4 or above. The biggest improvement point is exercise time." Next, key KPIs (Overall satisfaction 4.38, sufficiency of exercise time 2.94), average and distribution from 5 to 1 for each item, and comparison tables by department and experience. Then classification of free text—like "Lack of exercise/operation time: 23 cases, 46%, top priority improvement request," in four columns: theme, count, composition ratio, and interpretation. Finally, a graph of averages by item. The sheet header even included a note it added itself: "Original data has not been changed."
Analysis results
All I did was send one sentence. If I had started by designing pivots myself, I would still have been struggling with the angles at that point. If your environment doesn't output free text classification, just follow up with "Summarize the common opinions from the free text column into three points and tell me how many cases there were for each."
The trick is the order. Analysis has two steps: "exploration" and "finishing." In the initial exploration, it's better not to narrow down the order so that angles you didn't think of get mixed in. Look at what comes out, and if you want to see more of something, only then order it to "re-aggregate so the difference between departments is clear." Throw first, narrow down later. Just remember this order. If you're unsure how to show a graph, ask "Make it a graph suitable for a report to a supervisor," and Copilot will even handle the choice of presentation. It's the same as delegating to a capable subordinate; rather than tying them down from the start, it's better to let them work freely once and then correct the direction.
To be honest, not all the angles that come out are correct. It remains your job to judge whether it's okay to talk about the average of a department with only 3 respondents. Still, the distance to a report is incomparable between having a draft in front of you and having a blank page.
1-2. Decoding Mysterious Formulas
A long formula with triple-nested IFs and VLOOKUPs left by a predecessor. No one touches it because they're afraid of breaking it. You can turn that into plain language.
1Please explain step-by-step what the formula in (cell address) is doing. Please also include the role of each function used.
The response is a step-by-step explanation: "First, it looks for the corresponding unit price from another sheet; if not found, it displays zero; if found, it multiplies it by the quantity." The formula itself doesn't change. What changes is that it goes from an "untouchable spell" to "readable calculation." If you can read it, you can fix it. If you can fix it, you can hand it over. You can finally work on workbooks that were frozen due to a predecessor's transfer.
There's also a reverse use. Have it explain a formula you made and leave that explanation as a note for your successor. One piece of handover documentation is added without you writing it. Additionally, if you ask "Can you make this formula simpler?", it will even suggest a shorter alternative using new functions.
1-3. Prepping Dirty Data
Most causes of aggregation errors are dirty data. Duplicate rows, extra spaces, mixing "Company Inc." and "Co., Ltd." You can also ask for this cleaning.
1Please clean the data in this sheet. Look for duplicates, extra spaces, and notation inconsistencies, and show me a list of the problems found before fixing them.
How many times a month do you spend on this for name lists or sales data arriving from other departments? If you have 10 minutes of prep 4 times a month, that's 40 minutes. Just by eliminating this cleaning, Part 1 pays for itself.
The important part is the phrase "before fixing them, show me a list." Let it fix things only after you've seen with your own eyes what will change and how. Because cleaning is powerful, if you leave it all to the AI, it might unify notations you didn't want changed, so please work after taking a copy of the original file.
1-4. Asking for the Current Status of a Long-Unopened Management Sheet
Opening a task management sheet in a shared folder after two weeks. Over 100 rows, red text additions everywhere. What is going on now? Instead of tracing it with your eyes, ask this:
1Summarize the recent changes in this workbook.
On my machine, following the last update date and a count like "Recent changes: 8 red text additions, 4 new items," the main changes were bulleted at the project content level, each with the source cell address attached. You just jump to the addresses of items you're interested in to check the original text. Preparation 10 minutes before a meeting changes from a reading task to a thinking task.
To be precise, Copilot is reading the content of the table, not the edit history recorded by Excel. It knows the creator and last saver, but not who wrote which cell. Remember it's a tool to know "what happened," not to track down "who did it."
1-5. Asking for Advanced Analysis Without Knowing Python
Anomaly detection, statistical trend analysis, simulations. Until now, these were "jobs for people who can write Python." Today's Excel Copilot writes and executes the Python code needed for the requested analysis itself and returns only the results to you. It worked in my environment with zero Python knowledge.
1Use Python to find anomalies in this data. Please also include the reason why you judged it as an anomaly.
When I tried it with the previous survey, a sheet called "Python Anomaly Analysis" was added. The content had three parts. First, the number of cases per judgment method: 0 cases by IQR, 0 cases by Z-score, 1 response with the same score for all items, 18 responses with extreme high/low differences, 12 responses deviating from the overall trend, and 21 candidates for confirmation excluding duplicates. Next, the mean, standard deviation, and quartiles for each item. And a list of the 21 confirmation candidates with a "Reason for Judgment" column for each—like "All 7 items are 4 points" or "High utility and satisfaction but low exercise time and operation understanding." At the end, there was even a "Note on Interpretation."
Results of anomaly search with Python
What I liked was that when statistical outliers were zero, it didn't end with "no anomalies" but suggested confirmation candidates from other perspectives. It doesn't jump to conclusions but presents materials for humans to make the final judgment. With this design, I can leave it to the AI with peace of mind.
Just by adding "Use Python" at the beginning, Copilot switches its analysis method. "Statistically analyze sales trends," "Simulate next term's numbers in three scenarios," "Check if there is a correlation between these two items." Analysis that used to take days waiting for a specialized department now takes minutes at your desk. The results come back in the form of an Excel table, so you don't need to read Python code. However, before using the numbers in a report, check the number of original data rows and one value with your own eyes. The more advanced the analysis, the higher the value of this habit.
Part 2: Making It Remember Your Preferences (5 Minutes)
After trying Part 1, the next frustration arises: "The content is right, but I'm fixing the appearance every time." Graph colors, total positions, cell merging. This "final touch-up" takes about 2 minutes, but if it's 5 times a week, that's 40 minutes a month, 8 hours a year. Deciding "it's faster to fix it myself" is correct for a one-off, but it's the most expensive decision if done every time.
That manual work disappears if you teach it your preferences once. An operation where you state every order every time you ask is like telling your family "light seasoning, less rice" every single time; it won't last. The right way is to have it remember, like "the usual" at your favorite restaurant.
The location is "Customize Copilot" (or "Personalization" in some environments) from the three dots (...) at the top right of the Copilot window. If you write your preferences in the input field and save them, from then on, no matter which workbook you open, results will be returned according to those preferences. Start by pasting this set and rewriting it to your own style.
1Use blue tones for graphs and limit colors to three.2Do not merge cells.3Add bold headers to tables.4Place the total row at the top of the table, not the bottom.5Keep explanatory comments short and only about what was changed.
The change after teaching it is subtle but certain. Even with the same "Aggregate by location and make a graph" request, before teaching it might be a 5-color graph with a bottom total; after teaching, it's a 3-color blue graph with a top total and bold headers. Without saying anything, a table that looks like you finished it yourself comes back from the start. Not only does the 2-minute fix disappear, but more importantly, you become able to "show Copilot's output to others as is." Whether you can think "I can project this table on the screen as it is" right before a meeting—that 5 minutes makes that difference.
The trick to writing is a few short, clear lines. Bullet points of short lines work more reliably than long paragraphs. If you can't think of what to write, remember the last time you fixed a Copilot output. The content of that fix becomes one line of your preference.
One thing I want you to know as a specification: This setting is linked to your account, is not saved in the file, and is not shared with others. This is a design, not a flaw. Table preferences vary by person, and your "total at the top" shouldn't be forced on everyone. So, personal preferences go to the individual, and team rules go to the file (which I'll write about in Part 4). This distinction will be effective later.
Part 3: Choosing the AI Brain
From here, this is for those who saw "Auto" in the 30-second check at the beginning. The identity of that "Auto" is the model selection menu. You can choose the AI model itself that runs inside Copilot.
When you open it, the choices are not from just one company. On my screen, four latest models from two camps, GPT-based and Claude-based, are lined up. This means you can swap the AI brain inside Excel. A situation where brains from multiple companies coexist in one tool was unthinkable a year ago. The lineup will change with updates, so you don't need to remember the names. Just remember the mechanism.
I've received reactions like "Can you really choose the AI inside Excel?" many times within the company. You can. And this fact itself is hardly known even to people who use Copilot every day.
Usually, you can leave it on "Auto," which is the "leave it to Copilot" setting. If you select a model name from the menu, it will be fixed for that session and return to standard when you close Excel. You don't have to worry about forgetting to change it back.
As a practical feeling, the effect is easy to understand when consulting on the "design" of an aggregation. Even with the same consultation like "What's a good angle to show this budget vs. actual data?", the depth of the proposal differs visibly depending on the model. If you feel the answer is shallow, try changing the brain first before refining the question. That often solves it.
The guideline for usage is simple. Leave light work on Auto. Swap to a high-performance model only for heavy monthly analysis, complex aggregation design, or data processing mixed with long text. And remember, when you feel "Copilot's answer is shallow," you have the option to change the model before changing how you ask.
Part 4: Making the File Remember Rules with ".Rules" (10 Minutes)
What you taught in Part 2 was "your preference." What you teach here is "the rules of that file." Amounts in yen, negatives in red, don't delete columns without permission. You can write your department's "common sense" into the file itself.
People who give formatting orders verbally every time will receive a table with bare numbers without yen marks in the month they forget the order. If there is a place to remember rules on the file side, this accident structurally won't happen.
The method is to create one sheet named ".Rules" in the workbook and write one rule per line in column A. Start with this set.
1Display amounts with a yen symbol and 3-digit separators.2Unify percentages to one decimal place.3Display negative numbers in red with parentheses.4Do not delete or reorder existing rows or columns unless instructed.
The beginning of the month after setting this up looks like this: Update the numbers and say "Execute .Rules." The yen notation, the red negatives, and the percentage digits all come back arranged according to the rules. The formatting orders you were saying verbally every month are now permanently resident in the file.
Execution is just one sentence: "Execute .Rules." If you hide the sheet, it won't be read, so please leave it as is. The fact that rules are in a visible place itself becomes handover material for the team.
And the greatest value of .Rules is that it is shared with the file. The rules you set will be effective for the Copilot of everyone who opens that workbook. In other words, to summarize: Your personal taste goes to personal settings (not shared). Rules everyone follows go to .Rules (distributed with the file). If you're unsure which to write in, just ask yourself once, "Who is this for?" The answer will come in 3 seconds.
This "distributability" is effective the moment the quality of the team's tables is unified. If you distribute a workbook with a rule set you spent 30 minutes refining as a template, it will come back in the same format no matter who does the work. It doesn't end with individual time-saving; it becomes a department standard.
If you're unsure what rules to write, it's fastest to just transcribe the orders you give verbally every time you fix a subordinate's or junior's table, line by line. The more times you've said an order verbally, the better the rule. Starting with 4 lines and adding one line every time a manual fix occurs is sufficient.
Part 5: Safety Devices Before Delegating
By now, Excel has become a partner you can delegate a lot of work to. Finally, here are the devices to prevent accidents when delegating. There are two.
5-1. Have It Show the Work Plan Before Execution
When letting it touch an important workbook, the biggest fear is "having something broken without permission." Then, have it produce the steps before it moves.
1Before executing, please show me the plan of the steps, what you will do to which cells. Proceed with the work only after I approve it.
When used in practice, a plan like "I will change the format of columns B and C, add a month-over-month column to column D, and finally update the graph" comes back first, and the hands move only after you say okay. When the fear of it proceeding on its own disappears, the range of work you can delegate expands by one level. It's the same as "share the steps before starting" when delegating work to a subordinate; the basics of delegation don't change even if the partner is AI. A dedicated mode for showing the plan first is being rolled out sequentially, but even in environments where it hasn't been deployed, you can get the same effect just with this way of asking.
5-2. One Sentence to Protect Original Data and a 60-Second Double-Check
When asking for analysis or cleaning, add this one sentence.
1Do not change the original sheet; output the results to a new sheet.
Another most reliable safety device is to switch the toggle mentioned in Part 1 back to "Chat only." In this state, Copilot will not touch the sheet at all, no matter what it's asked. Make it a habit to look here first when you just want to consult or when you're opening someone else's workbook.
And before using the numbers that come out, do a 60-second double-check. Calculate the average or total of one question yourself and see if it matches. See if the number of rows in the original data matches the population of the aggregation. With just these two points, the reliability of the report will be on a different level.
You might feel double-checking is a hassle, but it's the opposite. When a number is pointed out in a meeting, whether you can say "It has been double-checked" directly links to the credibility of your report. Just avoid the situation where you can only answer "The AI produced it" when asked about the aggregation method.
You don't have to doubt everything. Delegate the work, but don't delegate the responsibility. This is the single rule of etiquette common to all the features introduced in this article.
Part 6: 3 Common Stumbling Blocks
Finally, I'll leave three questions that are always asked within the company.
Q1. The features in the article don't appear on my screen.
That's not your operational error. Not everyone is looking at the same Copilot. New features are distributed to each environment in order, and even within the same company, the screen might be different from the person in the next seat. Some features reached my environment first, and there are features that have been announced but haven't come to me yet. I've only included what worked on my actual machine in this article, but even so, there's a possibility it might be off by a few weeks from your environment. For features that don't appear, check the menu again next month. The features will wait for you.
Q2. Is it okay to let the AI read company data?
The Copilot your company has contracted for is provided on the premise that the input content will not be used for external learning, and it only works within the scope of your access rights. In that sense, it is a completely different act from pasting internal materials into a free AI service used personally. However, whether it's information you yourself are allowed to handle is a different story. For materials with sensitivity labels or personnel information, being able to view them is different from being allowed to process and share them. If in doubt, think about whether you could leave that material printed on your desk.
Q3. Does it work even if the data is dirty?
It works. Even if the formatting is messy or column headers change halfway through, you can throw it first. If you're concerned, just run the cleaning from Part 1 once. The most wasteful pattern is starting to prepare by thinking "I'll tidy it up before giving it to the AI" and running out of steam before handing it over. Throw it dirty, and tidy it up after it moves. That order is fine.
Summary: Back to the First Question
What can Excel's Copilot do? Here is the answer as of September 2026, line by line.
With one vague sentence, it can do analysis, formula decoding, data cleaning, status updates, and even advanced analysis using Python (Part 1). It remembers your preferences (Part 2). You can choose the brain (Part 3). The file remembers rules (Part 4). And you can make it move after showing the plan (Part 5).
The identity of the "is that it?" from six months ago was not performance, but the lack of a place to make it remember. Now there is a place. In other words, Excel has changed from a tool you order each time to a subordinate that works by remembering what you taught it.
I'll also leave the time required for setup. Part 1 is 0 minutes (just ask). Personal settings in Part 2 take 5 minutes. .Rules in Part 4 take 10 minutes. Parts 3 and 5 are just things to remember when using them. With a total of 15 minutes of setup, everything in this article starts moving.
That said, you don't need to set up everything today. There is only one thing to do.
Open any one piece of data you have on hand now and throw Part 1's "Please analyze the entire data in this sheet." With zero setup, you'll see today's capability. Those who thought "is that it?" six months ago should be surprised by the difference. If you feel a response, save this article and set it up in order starting from Part 2.
Once you work with an Excel that remembers what you taught it, you won't be able to go back to the way of repeating the same explanation every time.
I haven't included features that have been announced but don't work in my environment yet. This is everything as of September 2026, built only with things that worked. For the time being, you can fight sufficiently with the five parts written here.
Copilot evolves really fast. It's overwhelmingly faster to learn from someone who is practicing than to keep chasing the latest alone.
Thank you for watching until the end.
I post useful information about Copilot every day.
If you want to know techniques you can use from today, please follow me.
[Open Chat for Copilot Users Started!]
🎁 Participation Bonus 🎁
Copilot Work Style Starter Kit
- 100 Copilot Work Techniques
- List of No-Gos when using Copilot at work
- Quick Reference Table for Copilot Work Techniques
Copilot beginners and those who just want to watch are also welcome!
First, please feel free to take a look!
👇 Join here.
Copilot Work Techniques You Can Try from Today
