The companies that deployed AI fastest are now pulling back — not because the technology failed, but because the bills arrived before the accountability frameworks did.
Walmart capped usage on its internal AI assistant. Uber burned through an entire annual AI budget within months. Microsoft told thousands of its own engineers to switch tools before the end of June. GitHub repriced its flagship AI product around consumption rather than seats.
These are not small experiments. These are some of the most sophisticated technology organizations in the world, and they are all arriving at the same conclusion: unlimited AI access was a useful strategy for phase one. Phase two is about knowing exactly what you are getting for every dollar spent.
If your organization is still in phase one — rolling out AI tools broadly and hoping the ROI emerges — you are about to get a very expensive education.
The Data Behind the Shift
A KPMG Q2 2026 Global AI Pulse survey put numbers to what many executives have been feeling. Only 7% of business leaders report having established ROI from their AI investments. Twenty-four percent face active investor pressure to demonstrate value. Forty-two percent cannot clearly see where their AI spending is going.
That last number deserves to sit with you for a moment. Nearly half of all enterprises deploying AI at scale cannot tell you, line by line, what the tools cost or what they return. That is not a technology problem. That is a governance problem.
The KPMG data also revealed something more instructive than the failure rate: organizations where the CEO actively owns AI outcomes achieve meaningful business value at 57%, compared to 21% for those without executive sponsorship. The gap is not in the tools. It is in the ownership structure.
A separate Emergn analysis found that US organizations are losing an average of 2.4% of annual revenue to AI initiatives that fail to deliver expected value. For a $5 billion company, that is $120 million walking out the door annually in projects that looked promising in a slide deck and collapsed in production.
The industry has shifted from "are we using AI?" to "where is AI actually generating returns?" That question has a very different set of answers.
Walmart: Building Accountability Into the Architecture
Walmart's story is instructive because it is not a story of failure. The retailer built an impressive in-house AI assistant called Code Puppy — an internal agent capable of helping employees write, edit, and test software, create presentations, and analyze spreadsheets. Critically, Code Puppy was designed to work with dozens of AI models from different providers, letting developers switch between OpenAI, Google, and Anthropic models depending on the task. That flexibility is sophisticated engineering.
The problem was not the technology. It was the usage patterns.
With 2.1 million employees, even modest per-person AI consumption produces enormous costs at scale. Walmart's Global CTO Suresh Kumar noted that employees were frequently asking Code Puppy to solve the same problems repeatedly — making AI requests that were more expensive than building a common solution once and distributing it. The tool was working. The habits around the tool were wasteful.
Walmart's response was measured. Rather than kill access, it introduced a token-based system that ties consumption to business value. Employees still have access to Code Puppy, Claude, and ChatGPT. The difference is that each person now has a meaningful allocation that encourages more deliberate use. Asking the same question twenty times is no longer invisible — it has a cost.
For CIOs thinking about enterprise AI governance, this is the model worth studying. The goal is not to restrict access. The goal is to make consumption legible.
Uber: When Speed Outpaces Oversight
Uber's situation is a different kind of case study. The company rolled out Claude Code — Anthropic's AI coding assistant — to approximately 5,000 engineers at the start of this year. The adoption was fast and genuinely enthusiastic. Engineers found the tool useful and used it heavily.
The problem was that Uber reportedly burned through its entire annual AI budget within months. The speed of adoption was not matched by the oversight infrastructure needed to understand what was being consumed, why, and whether it was generating returns proportionate to the spend.
This is the pattern that plays out when phase one urgency drives tool deployment without the financial management layer that should accompany it. Access gets broad. Usage gets high. The bill arrives. Finance asks for a breakdown that IT cannot provide.
For CFOs looking at AI line items that are growing faster than planned, Uber's situation is a useful reference point. The question to ask your engineering and IT leadership is not "are engineers using AI?" but "do we have token-level visibility into what is being consumed, and do we have a framework for mapping that consumption to business outcomes?"
If the answer is no, you are operating Uber's model and you will eventually face Uber's reckoning.
Microsoft: The Internal Migration Signal
Microsoft's response to AI cost pressure is the most significant of the three because Microsoft is both a major AI investor and a major AI vendor. The company spent years building its position in the AI tools market, including a deep integration with OpenAI and GitHub Copilot.
When Microsoft asked thousands of its own engineers to migrate off Claude Code and onto an internally developed alternative before the end of June 2026, the message was clear: even the organization most committed to AI adoption is applying tighter scrutiny to which tools earn their place in the stack.
This was explicitly a cost-management decision. Microsoft was not abandoning AI coding assistants — it was evaluating the cost-per-outcome of the tools it was paying for against alternatives it had built internally. The switch was not ideological. It was financial.
For enterprise buyers watching this, the takeaway is that tool loyalty in AI has a much shorter half-life than in traditional enterprise software. A tool that was the best option at the start of 2025 may not earn its licensing fee in 2026. The pace of model improvement combined with increasing pricing complexity means that AI tool selections need active review — not a set-it-and-forget-it approach.
GitHub's Market-Wide Pivot
GitHub's move to token-based pricing for Copilot is less a story about one company's cost controls and more a signal about where the entire market is heading.
Seat-based pricing for AI tools made sense when the goal was adoption. You bought seats, you encouraged use, you hoped fluency followed. The problem with seat-based pricing is that it offers no correlation between what you pay and what you get. A seat for an engineer who uses Copilot for 400 hours a month costs the same as a seat for someone who opens it twice. The economics are broken for buyers who have highly variable usage patterns.
Token-based pricing fixes that asymmetry. What you pay maps to what you consume. Finance can see the line items. Procurement can set limits. Heavy users can be identified and their productivity gains measured against their consumption. Light users do not subsidize heavy ones.
For enterprise procurement teams, this shift creates immediate leverage. As more vendors move to consumption-based models, you gain the ability to negotiate based on actual usage projections rather than seat counts. The organizations that build rigorous internal usage tracking now will be better positioned to negotiate AI contracts and cut underperformers before the next renewal cycle.
The Technical Leader's Perspective
For CIOs and CTOs, the rationing conversation is really a governance conversation that has been deferred too long.
The right architecture for enterprise AI access looks something like this: tiered access based on use case and expected ROI, token budgets allocated by team or function, usage dashboards that are visible to engineering managers and finance simultaneously, and a quarterly review process that compares AI spend against measurable output.
This is not exotic infrastructure. Most of the organizations implementing it are building on top of what they already have. The harder part is the organizational design: who owns AI usage decisions, how are budgets allocated, and who has authority to expand or contract access based on demonstrated returns?
The CIOs who have moved fastest on this typically started with a simple question: for each AI tool we are paying for, can I show the CFO a specific outcome it produced in the last quarter? If the answer is no, the tool is probably not earning its keep regardless of how much engineers like it.
Tool consolidation is also accelerating as a result of this scrutiny. Organizations that deployed five different AI coding assistants in 2025 are now picking one and optimizing for it. The breadth-first phase of AI tool adoption is giving way to a depth-first phase where fewer tools get more deliberate investment.
The Business Leader's Perspective
For CFOs, COOs, and business unit leaders, the question is simpler but the stakes are identical: are we getting dollar-for-dollar returns on our AI spend?
The KPMG data suggests most organizations cannot answer this question, and that gap is becoming a boardroom-level issue. Investors are asking. Audit committees are asking. And the organizations that built financial management infrastructure around their AI spend in 2026 will be significantly better positioned to make the case for increased investment when the business case is genuinely there.
The practical framework being adopted by enterprises that have moved past the ROI gap looks like this: define success metrics for each AI use case before deployment (not after), require that tool access is tied to training on how to use it effectively, build mandatory reporting into the deployment contract so that every AI tool has a quarter-by-quarter return review, and create a kill list — a process for sunsetting tools that do not hit their targets within a defined period.
That last piece is the most culturally difficult. Enterprise AI has been treated as a strategic priority so frequently and so loudly that admitting a specific tool is not working can feel like admitting broader failure. It is not. It is exactly the kind of financial discipline that separates organizations that get lasting ROI from those that get impressive demos.
What the 7% Are Doing Differently
The KPMG data on CEO ownership is the most actionable finding. Organizations where the CEO actively sponsors AI outcomes and is accountable for results achieve meaningful business value at nearly three times the rate of organizations where AI ownership sits elsewhere.
This is not about the CEO making technology decisions. It is about accountability living at the highest level of the organization. When the CEO is on record about what AI is supposed to deliver, it becomes impossible for individual functions to run AI experiments without connecting them to enterprise-level outcomes.
The organizations building the infrastructure to measure AI ROI now — token-level visibility, outcome tracking, quarterly reviews — are not just managing costs. They are building the measurement capability that will let them make confident bets on the next wave of AI capability, because they will actually know what the previous wave delivered.
The unlimited-access era of enterprise AI lasted approximately two years. It produced a lot of organizational fluency, some genuine wins, and a significant number of line items that no one can justify. The accountability era is here. The organizations that build the right frameworks in the second half of 2026 will be the ones positioned to scale AI confidently in 2027.
The Bottom Line
Walmart, Uber, and Microsoft are not retreating from AI. They are growing up with it. The move from open access to measured consumption is what governance looks like when a new technology moves from experimental to operational.
For technical leaders, the question is whether your AI tool portfolio has the usage visibility and outcome tracking to survive a serious CFO review. For business leaders, the question is whether you have the financial infrastructure to make AI a line item you can defend, not just a strategy slide you can present.
The 7% who have established AI ROI did not get there by accident. They got there by treating AI spending the way they treat every other significant capital allocation: with accountability, measurement, and the willingness to cut what does not perform.
That discipline is now the table stakes for the next phase of enterprise AI.
Sources: KPMG Q2 2026 Global AI Pulse Survey | MarketScale: Enterprise AI Cost Controls | UC Today: KPMG AI Cost Visibility | Bloomberg (via Economic Times)
