Yesterday, an AI Function Inside Excel Stopped Working. If It Was in One of Your Workbooks, You Now Have a Broken Cell.
Yesterday, an AI Function Inside Excel Stopped Working. If It Was in One of Your Workbooks, You Now Have a Broken Cell.
A small event with a durable lesson: a vendor's AI call does not belong inside a formula that a reported number depends on.
Microsoft retired Excel's =COPILOT() worksheet function on 14 September 2026, as set out in Microsoft 365 Message Center notice MC1454373 on 14 August. It shipped in August 2025, reached users through the Insider and Frontier preview programs, and never made it to general availability — Microsoft had it slated for 2027 and then updated its roadmap to say it would not proceed. Excel's AI capability continues through the Copilot side pane; what has gone is the ability to call it from inside a cell.
If that sentence means nothing to you, this post is still worth five minutes, because the interesting part is not the function. It is what the function's short life says about a design decision a lot of finance teams have been making quietly all year.
What it did, and why people used it
The function let you write a natural-language instruction inside a cell and get the model's answer back as the cell's value. Classify this expense description. Summarise this free-text comment. Extract the state from this address. Pull a plausible category from this supplier name.
You can see the appeal immediately, particularly to anyone who has hand-coded two thousand rows of transaction descriptions at the end of a quarter. It was a genuinely clever idea, and the tasks people reached for it with were real tasks.
|
13 months
From first release in August 2025 to retirement on 14 September 2026 — without ever reaching general availability.
|
~4 weeks
Between the Message Center notice on 14 August 2026 and the function going dark on 14 September — the entire window anyone had to find and remove it.
|
What breaks, and what doesn't
Two things are worth separating.
The formula stops returning a value. That part is visible and, in isolation, easy to fix — you find the cell, you delete the formula, you do the work another way.
The second part is the one that matters. If that cell fed something — a category that drove a pivot, a flag that fed a summary, a classification that rolled into a management report — then the number downstream of it changes, and the thing that changes is not obviously connected to the thing that broke. A broken cell announces itself. A total that quietly moved because a broken cell sat four steps upstream of it does not.
And there is a third case that is easy to miss entirely: a workbook where the function was used once, months ago, and the result was pasted as a value. Nothing is broken there at all. But a number in your model was produced by a preview AI feature that no longer exists, and nobody wrote down that it was.
Finding every instance
This is a fifteen-minute job across a small workbook estate and a half-day job across a large one, and it is worth doing even if you are confident the answer is zero.
Search for the function name across your workbook folders rather than opening files one at a time. Be careful how much you trust the result: a Windows content search depends on indexing being switched on for that location and on the Office filter reaching formula text, and a file sitting in an unindexed folder or an offline OneDrive state can simply not appear. The reliable version is Excel's own Find with Look in: Formulas and Within: Workbook, run on each candidate file. Check shared drives and the finance team's personal folders, not only the official model library — preview features get tried in personal copies far more often than in governed ones.
Then ask the two questions in order. Does anything reference this cell? Trace dependents before you delete anything. And was this ever pasted as a value somewhere else? That one you cannot search for, which is why the honest answer is to ask the people who build your models whether they used it, rather than to assume the search was complete.
What to replace it with
Almost every finance use of this function was one of two things, and they have different answers.
A lookup wearing a costume. Categorising suppliers, mapping cost codes, assigning a region from an address — these look like language tasks but they are reference-table tasks with an incomplete reference table. The durable fix is to build the table. It is less exciting, it is deterministic, it is auditable, and it does not stop working when a vendor changes its mind.
Genuine bulk text work. Summarising hundreds of free-text comments, or classifying descriptions that really do vary. That work is legitimate, but it belongs outside the model: run it once, review the output, and bring the reviewed result in as data. The difference is not cosmetic. In the first arrangement the model recalculates against a live AI service every time somebody opens the file. In the second, the AI ran once, a person checked it, and what sits in the workbook is a fixed input with a known provenance.
The rule worth keeping after the news has gone stale
Here is the general version, and it will outlast this particular function by years:
A live call to a vendor's AI service does not belong inside a formula chain that a reported number depends on.
Three reasons, in order of how likely they are to bite you.
It is a supply risk. A formula is a dependency. A preview feature is a supplier with no contract, no notice period and no obligation to you. This is exactly what happened here: a feature that was heading toward general availability was cancelled instead, with a few weeks' notice, and the only people who found out early were the ones who read trade press.
It is not reproducible. This is not an inference about how language models behave — it was in Microsoft's own guidance for the function, which warned that results could change over time as the underlying model evolved, even with identical arguments, and told users to reach for conventional formulas where accuracy or reproducibility mattered. Everything else in a spreadsheet is deterministic. A cell that is not, sitting among thousands that are, breaks the one assumption every reviewer brings to a workbook.
It is invisible in the audit trail. Nothing about a populated cell tells you that a model produced it. Six months later, nobody can tell which numbers a person derived and which a service generated, which is precisely the question somebody asks when a figure turns out to be wrong.
None of this is an argument against using AI in finance work. It is an argument about where it sits. Use it to prepare inputs, to check finished work, to draft the tedious parts. Do not wire it into the calculation and then report the output as though the calculation were yours.
Three things to do this week
Run the search. Across every folder that holds a model anyone relies on, including personal copies. Twenty minutes, and you get a definite answer instead of a confident one.
Write down where AI touches your reported numbers. Not a policy — a list. Which workbook, which step, whether it runs live or ran once, and who reviewed the output. Most finance functions cannot produce this list today, and the exercise of trying is more informative than the list itself.
Set the boundary before the next preview feature arrives. There will be one. Decide now that AI prepares and reviews but does not calculate, and you will not have to have this conversation again on four weeks' notice.
Could you list every place AI touches a number you report?
Most finance functions have adopted AI faster than they have documented it, which is fine until a vendor withdraws something. PFL provides senior-level outsourced finance, management reporting, and AI automation for Australian NFP, NDIS, and SME organisations — including drawing the line between where AI helps and where it becomes a dependency.
Talk to PFL →The Register — Excel's Copilot function is headed for the Recycle Bin (17 August 2026)
Redmond Magazine — Microsoft Pulls Excel's COPILOT() Function After Just One Year
TechRadar Pro — Microsoft is dropping its Excel Copilot function after only a year
Windows Report — Microsoft is retiring Excel's COPILOT() function after just one year (Message Center MC1454373)
Comments
Post a Comment