Microsoft has begun quietly swapping OpenAI and Anthropic models for its own MAI systems inside Excel and Outlook, according to reporting cited across US outlets this week. Tens of thousands of prompts in the two applications now run weekly on Microsoft’s in house models, a move framed unambiguously as cost management rather than technical preference.
Also read: Microsoft Cuts 4,800 Jobs, Xbox Division Hit Hardest
The rationale is straightforward. Microsoft has enjoyed heavily discounted computing through its long partnership with OpenAI, but that arrangement will not hold indefinitely. AI chief Mustafa Suleyman’s team is reportedly trying to get ahead of the moment when frontier labs start charging market rates. Suleyman put it plainly in June: Microsoft pays Anthropic a lot of money, and the internal goal is to reduce, then eliminate, that cost. It follows an earlier signal in May, when Microsoft began winding down most of its internal Claude Code licenses, six months after rolling them out enterprise wide.
Taken alone, this could read as one company optimising its stack. Read alongside the rest of 2026, it looks like the start of a correction. Uber’s CTO went viral for burning through the company’s entire annual AI coding budget in four months. Uber’s chief operating officer has since said publicly that the link between AI spending and measurable customer benefit is not there yet. Amazon quietly retired an internal usage leaderboard after engineers began burning tokens on meaningless tasks purely to climb the rankings, a habit some at the company had taken to calling tokenmaxxing. Meta ran something similar. Starbucks pulled an AI inventory system from more than 11,000 stores after staff found it faster to count stock by hand.
The pattern has a name now: the AI spending reckoning. For roughly two years the operating assumption inside large enterprises was simple, spend more, get more. Budgets set in the autumn of 2025, before agentic tools like Claude Code reset expectations of what a single employee could consume in tokens, are reportedly now obsolete almost everywhere. Uber exhausted an entire year’s AI budget in three months. One company is said to have spent half a billion dollars in a single month after failing to cap employee usage.
None of this amounts to companies abandoning AI. MIT research cited widely this year traced most failed AI deployments not to weak models but to a simpler problem: teams adopting tools before deciding what those tools were meant to achieve. The correction under way is about discipline. It means matching the model to the task, measuring cost per completed outcome rather than cost per token, and routing simple requests to cheaper systems while reserving frontier models for work that genuinely needs them.
For a company the size of Microsoft, in housing a slice of Excel and Outlook traffic is a modest technical shift with a clear signal. It tells every finance team watching that even the biggest AI spender in the room is now treating token consumption as a line item to be managed. The free for all phase of enterprise AI adoption, it seems, is closing. What replaces it is less dramatic and considerably more boring: procurement discipline, usage caps, and a harder question than whether a company is using AI. The new question, increasingly, is whether it is paying for itself. The age of free AI lunches seems to be dwindling away.
