An internal anonymous spreadsheet at Microsoft reveals that roughly 350 employees spent a median of $300 on AI tools over the past 28 days, with one person hitting $28,000—nearly a 100x gap. CoreAI has the highest departmental median at $975, four times that of Azure, while the record single bill came from CPS, the team that helps customers deploy AI in production. Company-wide AI costs have been laid bare in a single spreadsheet.

It's like the company handing out unlimited meal cards to everyone, hoping people would eat well—but at the end of the month, the bill tells a different story: most employees spend $30–$40 on a normal workday lunch, while a few colleagues entertaining clients rack up $20,000 a month on Japanese cuisine alone. The problem isn't the meal card itself—it's that a handful of people quietly send dining costs through the roof. Microsoft's current approach: default everyone to the company cafeteria (GPT-5.6 Sol), require an extra approval to eat elsewhere (Claude), and cap each card while reconciling monthly. That's where the analogy ends, though—the real difference is this: AI tokens are a variable bill measured by the word, whereas a traditional meal card has a hard ceiling no matter what. One call to a model, one chunk of code generated, and the bill jumps another notch. That's where the "out of control" problem really lives.
Incident

An Anonymous Spreadsheet Sets Off a Bomb: $28,000 Burned in 28 Days

Microsoft's internal anonymous compensation spreadsheet added an AI spending column this year. Why did that column cause an uproar?

Every year, an anonymous compensation spreadsheet circulates inside Microsoft, where employees voluntarily fill in their raises, bonuses, and stock. In 2026, a new column appeared: how much did you spend on AI in the past month?

Only about 350 US employees out of the global workforce voluntarily submitted data, enough to sketch a steep curve. The median across all self-reporting employeesmedian(line everyone up from lowest to highest—the person right in the middle is the median), CoreAI's median, and the biggest single bill are all on this card.

$28,000
Highest single-employee 28-day token spend
Source: IT Home
$975
CoreAI monthly median spend
Source: IT Home
$300
All self-reporting employees' monthly median spend
Source: IT Home
223,000
Microsoft global headcount
Source: IT Home

Not just because the numbers are big.

Huge spending doesn't necessarily translate to huge output. The spreadsheet was meant to promote transparency, but it moved AI costs from the company ledger to the personal bill for the first time—and put a question on the table: is the top spender using AI efficiently, or just burning money?

Why It Matters

The Highest Median and the Wildest Outlier Aren't in the Same Department

How much you spend is strongly correlated with how much your job depends on AI—but the single most expensive bill doesn't come from the model team. It comes from the customer deployment team.

Break it down by department and the gap opens up immediately. CoreAI's monthly median is more than 3x the company-wide figure, with personal extremes running even higher; Microsoft AI is roughly half of that, and Experiences and Devices plus Azure are even lower. The distribution tells a clear story: the people building models and platforms are themselves steeped in models.

The counterintuitive part is CPS (Customer & Partner Solutions). Its median is far less dramatic than CoreAI's, yet it produced the company's all-time high—people doing customer deployments have the least budget awareness when using AI.

The logic behind it isn't complicated: CPS employees spend their days running AI pilots for customers, building demos, and running large-scale evaluations, so token(the smallest unit large models bill on, based on input and output text volume—the more words, the higher the bill) costs are naturally tied to project scale.

The median itself shows that AI is already a routine tool. The extremes tell a different story: someone has a stack of automated agent tasks running in the background—typing prompts by hand couldn't burn through that much.

How do you clamp it down? See the next section.