AI is still moving deeper into corporate work, but the bill is starting to challenge the promise that software agents are a cheaper substitute for people.
Microsoft's pullback from direct Claude Code access has turned a quiet budgeting problem into a larger question for the AI market. Companies spent the past two years telling employees to use generative AI everywhere. Now some of the most committed buyers are discovering that enthusiasm has a cost, and that cost does not always look like a clean productivity gain.
The issue is not that AI tools are useless. Far from it. Coding assistants, research agents and workflow copilots are already changing how work gets done. The problem is that enterprise adoption is moving from pilot projects to heavy daily use, and the economics become less forgiving when thousands of employees start sending prompts, running code checks, chaining agents and asking premium models to do work that was never priced like ordinary software.
Microsoft has reportedly canceled most of its direct Claude Code licenses and is steering employees toward GitHub Copilot CLI instead. That move came only months after the company encouraged broad use of Claude Code across engineers, designers, project managers and other teams. It also sits awkwardly beside Microsoft's wider AI ambitions, including its Azure relationship with Anthropic through Foundry and its central role as OpenAI's most important cloud partner.
The uncomfortable lesson is simple. A tool can look cheap at the seat level and expensive at the usage level. Old enterprise software economics were built around licenses, renewals and predictable usage bands. AI adds compute every time a worker asks a model to think, revise, test, summarize or call another tool. With agents, that can happen many times inside a single task.
As Axios recently noted, corporate leaders are starting to question whether fast-rising AI spending is producing returns they can actually measure. The same report said Microsoft cut back on Claude Code partly because of cost, while Uber executives have also raised concerns about the price of AI tools. Fortune previously reported, citing The Information, that Uber used its full 2026 AI coding tool budget in the first four months of the year after pushing internal adoption through usage leaderboards.
That matters because many companies have treated AI adoption as a cultural race. Teams were rewarded for using more of it, not necessarily for proving that the work became more valuable. In that environment, employees do what incentives tell them to do. They try the tool on everything. Some tasks are worth it. Others are not. The invoice does not make that distinction.
Nvidia's Bryan Catanzaro gave the market the cleanest version of the concern when he said his team's compute costs are far beyond employee costs. That is not an argument against AI. It is a warning about scale. If the best use cases need expensive models, repeated inference and large amounts of cloud infrastructure, replacing human labor is not automatically cheaper. In some cases, the cheaper answer may be a person using a limited amount of AI with discipline.
Cheaper models do not guarantee cheaper work
The next phase of AI spending will be shaped by a contradiction. Token prices are expected to fall, but overall usage may rise much faster. Goldman Sachs has projected that agentic AI could drive a 24-fold increase in token consumption by 2030, reaching roughly 120 quadrillion tokens per month. Gartner has also forecast that inference for a one-trillion-parameter model could cost nearly 90% less by 2030 than it did in 2025, while warning that falling unit costs may not translate into lower enterprise bills.
That is the part executives need to understand. Agents do not behave like a cheaper search box. They plan, call tools, check outputs and sometimes loop through steps repeatedly. A simple prompt might be inexpensive. A multi-step agent that writes code, tests it, explains it, revises it and checks documentation can consume far more. Multiply that by a large workforce and the savings story becomes much harder to prove.
Microsoft's own workplace research still shows why companies are not walking away. Its 2026 Work Trend Index, based on 20,000 workers across 10 countries and analysis of Microsoft 365 productivity signals, found that nearly half of Copilot interactions involved analysis, decision-making and problem-solving. It also reported a 15-fold year-over-year increase in active agents on Microsoft 365, with even faster growth in large enterprises.
That points to the practical answer. The question is not whether AI should be used. It is where the work is important enough to justify the compute. Coding, data analysis, customer support triage and document-heavy workflows can still make sense when time saved is visible and quality improves. Letting every employee use frontier models for casual questions, duplicate drafts or low-value experiments is harder to defend.
For startups, this creates both a risk and an opening. Founders selling AI tools will need to prove ROI with more than a polished demo. Buyers will ask about usage caps, model routing, audit trails and whether cheaper models can handle routine work before premium systems are called in. The winners will not be the companies that simply add agents to everything. They will be the ones that make AI feel useful without making the finance team feel trapped.
The market is entering a more disciplined period. AI will keep spreading through the enterprise, but the easy story that automation always cuts labor costs is losing strength. What comes next is more practical: fewer unlimited experiments, more measured deployments and a sharper focus on the work where machines clearly earn their keep.
Also read: AI agents are starting to do real research math • AI still has not solved software pricing, and Snowflake knows it • Rivian says AI will make CarPlay less important in its EVs
