One employee in Microsoft's customer and partner organization reported $28,000 of AI tool spending over 28 days. The median across the roughly 350 people who filled in the same field was about $300. The number sits in an internal employee compensation spreadsheet that Business Insider published earlier, which in 2026 gained a new column for monthly AI spend; Gadget Review picked the outlier out of it. The timing is unkind to whoever wrote it down: in early August, CoreAI executive vice president Jay Parikh sent staff a memo telling them to moderate excessive AI use, 404 Media reported.
The spread around that median matters more than the record holder. Several other employees cleared $10,000. Others, including people in the same organizations, spent a few tens of dollars. Same company, same month, same tools, three orders of magnitude apart.
Parikh's memo said the company is not trying to maximize token consumption, and that employees should focus on outcomes that are useful to customers and the business. Microsoft will update its internal rules and hold token spend to the same discipline it applies to other significant resources. It will also tighten oversight of how employees use AI and make OpenAI's GPT-5.6 Sol the standard model for internal work. According to 404 Media, that model consumes fewer tokens.
That last line is a cost decision wearing the uniform of a standards decision. The stated property of the new internal default is not that it does the work better. It is that it spends less doing it.
Two things about the $28,000 are easy to over-read. It is self-reported, in a sheet people chose to fill in, so it is a confession rather than an audit line. And the $300 median describes the middle of a self-selected sample of about 350 people, not what Microsoft spends per head. Both numbers are directionally interesting and neither is a finance figure.
What the sheet does not have is a column for what any of it produced. It records how much each person spent and nothing about what came back, which is exactly the measurement gap Parikh's memo is trying to close after the fact. A company that wanted outcomes measured would have asked for outcomes. Someone asked for spend, and roughly 350 people were happy to answer, because until August the number read as evidence of enthusiasm.
There is precedent for how that ends. Amazon ran internal leaderboards ranking employees by how much AI they used, then switched them off once the bills came in higher than anyone could justify.
The uncomfortable part is that the enthusiasm was manufactured on purpose. Microsoft told employees to use AI as often as possible, and roughly 350 of them documented how well they had listened. The correction now has to draw a line between use that counts and use that does not, and the only instrument named so far is a model that burns fewer tokens on the same work.