Microsoft has shipped new Copilot features into Excel almost every month in 2026, from workbook rules to Python-powered analysis to finance-specific “skills.” For any HR manager or team lead deciding how much to lean on it, and how much Excel training the team still needs, the real question is not whether Copilot is impressive. It is what it can actually be trusted to do without supervision, and where it still needs a human who understands Excel.

This article breaks down what Copilot in Excel handles well in 2026, where it still falls short, what it actually costs once the promotional pricing is stripped away, and what that means for how much training your team still needs alongside it.

What Is Copilot in Excel, Exactly?

Copilot in Excel is Microsoft’s AI assistant built directly into the spreadsheet, sitting in a side pane rather than a separate app. Unlike a general chatbot, it reads your actual worksheet: your column headers, your data types, your real formulas, so its suggestions are grounded in the workbook you are looking at rather than a generic answer.

It works in three modes: chat, where it answers questions and offers suggestions without touching the sheet; edit, where it makes changes you can accept or reject; and plan, where it breaks a bigger request into steps first. Recent updates have added Python-powered analysis for more advanced data work, workbook rules that let a team define formatting and formula conventions that stick to a specific file, and a file picker that can ground Copilot in emails, meetings, and other Microsoft 365 data beyond the spreadsheet itself.

What Copilot in Excel Actually Does Well

Used on the right kind of task, Copilot genuinely saves time. A few areas stand out.

Formula Writing and Explanation

Asking Copilot to build a SUMIFS, an XLOOKUP, or a nested IF and explain what each part is doing is one of its strongest use cases. It saves the time spent digging through forums or half-remembered syntax, and the explanation helps less experienced users actually understand the formula instead of just copying it.

Data Cleanup and Structuring

Copilot is reliable at removing duplicates, fixing inconsistent formatting, and standardizing messy entries before analysis starts. This is unglamorous work that eats a disproportionate amount of time in most reporting workflows, and it is one of the lowest-risk places to let AI take the first pass.

Charts, Summaries, and Quick Analysis

You can ask Copilot to summarize a table, flag outliers, or generate a chart in plain language instead of manually configuring one. For a quick first look at a dataset, this is faster than doing it by hand and generally accurate, as long as the underlying data is already clean.

Repeatable “Skills” for Recurring Tasks

A June 2026 update lets teams define repeatable Copilot “skills” for tasks like month-end close, variance analysis, or building a financial model, so the same instructions apply consistently every time instead of being rebuilt from scratch. Microsoft has said it tests these against benchmarks built specifically around finance workflows, which is a meaningful step up from a generic AI feature bolted onto a spreadsheet.

Where Copilot in Excel Still Falls Short

The same updates that make Copilot useful also make its limitations easy to miss, because the output usually looks polished even when it is wrong.

Large or Messy Datasets

Copilot reliably handles tables up to roughly two million cells. Beyond that, it either truncates its context or refuses to run at all. It also struggles with merged cells, blank rows inside a table, and inconsistent headers, the kind of formatting issues that look fine to a person but confuse the model. A few minutes of cleanup before asking Copilot to work on a file saves a lot of frustration later.

Numbers and Facts It States With Confidence

Independent testing of Excel’s COPILOT() function, which pulls AI-generated data directly into cells, found missing entries, incorrect facts, and inconsistent results depending on exactly how a prompt was worded. The output did not look uncertain. It looked like a normal spreadsheet result, which is exactly the problem for a team that does not have the underlying knowledge to spot when a number is wrong.

Complex or Judgment-Heavy Formulas

Copilot can still produce broken formulas, unexpected circular references, or logic that technically runs but does not do what was actually asked for, especially on multi-step or highly specific requests. Excel’s calculation engine is not the issue here. A poorly specified prompt still produces a formula that looks plausible, and only someone who understands what the formula should be doing will catch that it is not.

Platform and Access Inconsistencies

Feature availability is not always consistent across Excel for Windows, Mac, and the web, and 2026 has seen real service disruptions, including a period where “Edit with Copilot” stopped working in the desktop app entirely. What one employee’s Copilot can do in a given week does not always match what a colleague’s can do, which makes it a hard tool to standardize training around without a plan.

Why Copilot Doesn’t Replace the Need for Excel Skills

The pattern across every limitation above is the same. Copilot works best when someone on the receiving end already knows enough Excel to check its output, and it is genuinely risky when nobody does. An employee with a solid Excel foundation can glance at a Copilot-generated formula and know within seconds whether the logic is sound. An employee without that foundation has no real way to evaluate what they are looking at, and tends to either trust it blindly or avoid the tool altogether out of caution.

AI and Excel Skills Gaps: Why Teams Fall Behind goes deeper into why this gap forms across a team, and covers why the underlying skill still matters even as the tools around it keep changing.

How Teams Should Actually Use Copilot in Excel Right Now

Clean Your Data Before You Ask Copilot to Touch It

A few minutes fixing headers, removing merged cells, and standardizing formats before starting will avoid most of the confusion Copilot runs into on messy files.

Treat Every AI-Generated Number as a Claim to Verify, Not a Fact

This applies especially to anything pulled in through the COPILOT() function or used in a report that leaves the team. A quick manual spot-check catches most errors before they become a bigger problem downstream.

Use It to Speed Up Familiar Work, Not to Learn Excel From Scratch

Copilot is a strong accelerant for someone who already knows what they are doing, and a poor substitute for structured learning for someone who does not. How to Build an Excel Training Plan for Employees in the Age of AI covers how to build AI orientation into a training plan without turning it into the whole plan.

Why Nordic Teams Feel Confident Choosing Learnesy Alongside Copilot

Rolling out Copilot licenses across a team solves the access problem. It does not solve the skills problem, and for HR managers running training across departments in Sweden and Norway, that gap is exactly where a structured program earns its place.

Learnesy delivers Excel, data analysis, and AI training in Swedish and Norwegian, built around real industry workflows in logistics, procurement, finance, and real estate, with an HR admin dashboard that shows who has actually built the foundation Copilot assumes they already have. A few things tend to matter most to the HR managers evaluating that decision:

• 12 years working with Nordic business teams, since 2014
• 500+ companies trained, with a 4.9 rating on Google
• A dedicated Customer Success Manager and a kickoff meeting for every client, not a support ticket queue
• Courses built around logistics, procurement, finance, and real estate workflows, not generic office scenarios

For HR managers weighing what a training platform needs to cover alongside a Copilot rollout, 5 Factors HR Managers Should Look for in a Team Excel Training Platform is a useful place to start.

Summary: What Copilot in Excel Can and Can’t Do

Copilot in Excel is genuinely good at formula writing, data cleanup, quick summaries, and increasingly at repeatable finance workflows. It is still unreliable on large or messy datasets, on facts and figures it states with unearned confidence, and on complex or judgment-heavy formulas, and it comes with a real licensing cost most teams underestimate. None of that makes it a bad tool. It makes it a tool that is only as good as the Excel skills of the person using it, which is the part most rollouts skip.

Explore Learnesy Courses