AI was sold as a way to reduce labor costs, but the bill is starting to move in the opposite direction. Microsoft and Uber are now showing why token consumption is becoming a boardroom problem, not just an engineering detail.
The uncomfortable part of enterprise AI is no longer whether employees will use it. They are using it heavily. The problem is that some companies are discovering that the more successful the rollout looks on an adoption dashboard, the harder it becomes to explain the cost.
That is the lesson now coming through from Microsoft and Uber. According to Fortune, Microsoft has been moving many developers away from direct Claude Code licenses and toward GitHub Copilot CLI, while Uber has had to revisit its AI tooling budget after engineering usage grew faster than expected. The story is not that AI tools failed. It is almost the opposite. They became useful enough that people kept using them, and the token bill followed.
This changes the economics executives thought they were buying. A normal software subscription is easy to model. You count seats, negotiate a rate and build it into the annual budget. AI agents do not behave that way. They read, write, reason, retry, inspect files, generate code and sometimes repeat the same work in ways that are hard to predict from the outside. The cost is tied to activity, not headcount.
For years, companies talked about AI in terms of productivity. The promise was simple enough: give employees better tools, automate repetitive work and eventually reduce the amount of human labor needed for certain tasks. That may still happen in some places, but finance teams are now learning that productivity tools can also become consumption engines.
Uber is a clean example because software engineering is one of the areas where AI adoption has moved fastest. Developers can use tools such as Claude Code, Cursor and GitHub Copilot to draft code, inspect bugs and work through multi-step changes. Those workflows are more valuable than autocomplete, but they are also far more expensive because agentic tools can consume large volumes of input tokens before producing a useful answer.
That matters because a token is not a metaphor. It is a unit that can be priced, metered and billed. When thousands of engineers start running coding agents across large codebases, the company is no longer paying for a helpful sidebar in an editor. It is paying for a new layer of compute that sits inside daily work.
This is why the word tokenmaxxing has moved from internet joke to management headache. Some workers and teams have treated heavy AI usage as a signal of seriousness, almost as if a larger token bill proves deeper adoption. But consumption is not the same thing as output. A worker can burn more tokens without shipping better software, answering more customer issues or improving margins.
The pricing reset is coming
The bigger question is whether this is a temporary cost problem or the real shape of enterprise AI. Model providers will argue that prices will fall as chips improve, inference becomes more efficient and smaller models handle more routine tasks. There is truth in that. Every major cloud cycle has started with expensive capacity and moved toward cheaper, more optimized infrastructure.
But AI agents add a complication. If the price per token falls while each task consumes far more tokens, the total bill can still rise. A recent academic study on agentic coding tasks found that these workflows can consume vastly more tokens than ordinary code chat, with repeated runs on the same task varying widely in cost and higher spending not always producing better accuracy. That is the part CFOs will notice.
Microsoft is especially important in this debate because it sits on both sides of the table. It sells AI productivity through Copilot, pays for infrastructure, works closely with OpenAI and also has to manage its own internal AI usage. When a company with that much leverage starts steering employees toward tools it controls more directly, other enterprise buyers will pay attention.
For vendors, this creates a pricing dilemma heading into 2026 renewals. If they push too hard on usage-based pricing, customers may introduce caps, route work to cheaper models or bring more workloads onto private infrastructure. If they keep prices too low, the economics shift back onto the AI companies and their investors. Neither side can pretend the bill is invisible anymore.
There is also a management lesson here. Companies that treat AI as a general instruction to use more tools will get messy economics. Companies that tie AI usage to measurable work will have a better chance of defending the spend. That means budgets by team, model choice by task, limits on expensive agent runs and a clearer answer to a basic question: what did this token bill actually produce?
The next phase of AI adoption will be less glamorous than the last one. It will involve procurement reviews, internal dashboards, cost controls and arguments over whether a workflow belongs on a frontier model at all. That may sound dull, but it is where the real market will be decided. AI does not need to be cheap to win. It needs to be worth what it costs.
Also read: Ericsson is leaving Kista for a new Stockholm technology base • AI researchers are testing life after the Transformer • Prosus is asking Brussels to rethink its Delivery Hero sell-down
