Someone on your team is using a $200-a-month AI subscription to plan a vacation to Portugal.
I know this because I hear it from every CIO I talk to. You approved the AI budget. You rolled out the tools. You gave the speech about transformation. And now someone in accounting is generating packing lists while the operations team that could actually redesign your supply chain workflow is waiting for compute availability.
The allocation decision belongs to the organization. It has provided an expensive, scarce resource without explaining what useful work it should support or helping people learn to do that work. Blaming the employee planning a vacation misses that responsibility.
NVIDIA cut gaming GPU production by 30-40% in the first half of 2026 to prioritize AI chips. Memory now accounts for 70-80% of the bill of materials for high-VRAM GPUs. The Economist reported in April 2026 that OpenRouter, one of the most-watched model-routing platforms, saw weekly token usage quadruple between January and March alone. The GPU generating beach itineraries is not a metaphor. It is an allocation decision someone should have made deliberately and didn’t.
The instinct is to restrict access. That instinct is wrong. But doing nothing is also wrong.
The Price of a Token
Here is the part most executives skip when they think about AI costs: the compute itself is getting more expensive, not less.
GDDR6 prices rose roughly 60% in late 2025. GDDR7 climbed even more sharply. NVIDIA’s RTX 5090 carries a $1,999 MSRP but retails between $3,500 and $4,000. These are not temporary blips in a supply chain recovering from disruption — they are structural reallocations as the entire semiconductor industry pivots toward AI data center products. J.P. Morgan Asset Management’s 2026 analysis adds the part most executives miss: many AI services are priced well below cost today, subsidized by providers competing for adoption, and a single user running newer agentic tools can demand 10-100x more compute than a year ago. The economics of compute have changed, and they are not changing back to where most budget models assume they will.
Enterprise AI spending is projected to approach half a trillion dollars by 2026. Eighty-six percent of enterprises are increasing their budgets. The average organization now spends $1.2 million annually on AI-native applications — a 108% year-over-year increase. The money is flowing freely. The compute to absorb it is not.
For budget holders, the gap between daily use and advanced use deserves attention: 88% of employees use AI daily, but only 5% use it in advanced ways that transform their work. A 2026 Gartner survey put a sharper edge on the same gap from the project side: only 28% of AI use cases meet their ROI expectations, 20% fail outright, and the remaining 52% underdeliver against the business case that approved them. Forrester now expects roughly a quarter of planned 2026 AI spend to slip into 2027 as boards reassess. That gap — between what organizations are spending and what they are getting — is not a rounding error. It is the central leadership problem of the year.
Every token has a real cost. The question nobody is asking loudly enough is whether that cost is buying something.
The Productivity Paradox, Again
An NBER study of 6,000 executives found that nearly 90% of firms reported no measurable impact from AI on either employment or productivity over the prior three years. Executives forecast only a 1.4% productivity increase over the next three. A separate PwC-linked survey reported by VBeyond found 56% of CEOs see zero financial return from AI investments, while 22% report higher overall costs once the tools moved past pilot.
Robert Solow’s ghost is having a moment. “You can see the computer age everywhere but in the productivity statistics.” That was 1987. We are running the same paradox with better GPUs and worse excuses.
But here is where the story turns — and it turns on a measurement problem, not a technology problem.
Only 23% of enterprises can accurately measure their AI return on investment, despite 89% having adopted AI tools. The organizations in that 23% — the ones who actually built the instrumentation to track what AI produces — report a 27% average productivity improvement and $8,700 per employee annually in efficiency gains.
These findings put measurement at the center of the argument. Without it, a budget holder cannot distinguish useful work from activity, even when both involve AI.
42% of companies abandoned most of their AI projects in 2025, up from 17% the prior year. They did not fail because AI is broken. They failed because they never built the infrastructure to know whether it was working. When you cannot see value, every cost looks like waste — and the default response is to kill the project.
Is measurement genuinely hard? Yes. Measuring knowledge-work productivity has always resisted clean quantification, and AI makes attribution harder, not easier. You cannot tag the business value of every prompt. But you do not need to. Portfolio-level measurement — tracking outcomes at the team and workflow level rather than the individual interaction level — is achievable and sufficient. The 23% who do it are not counting API calls. They are measuring what changed in the work.
The FinOps parallel makes this concrete. 98% of FinOps teams now manage AI spend, up from 31% in 2024. That is the fastest adoption of a cost-management discipline in enterprise history. Yet 53% still cannot see the full scope of what they are spending, and 40% cannot quantify the value derived from it. They built the spending dashboards before they built the value dashboards.
That is like knowing your grocery bill to the penny while having no idea whether you are eating well.
Enable Broadly, Train Deliberately
The reflex to restrict AI access has already been tested at scale. It fails.
98% of organizations report unsanctioned AI use. 45% of AI adoption happens outside formal IT procurement. When you restrict access, you do not reduce usage — you make it invisible. Shadow AI thrives in the gap between what organizations sanction and what employees need. Restriction does not solve the waste problem. It guarantees you cannot see it.
The leverage point is not access. It is capability.
Employees who receive 81 or more hours of annual AI training gain 14 weekly productivity hours — nearly double the 8-hour median. Organizations with formal training programs achieve 2.7 times higher proficiency and 4.1 times higher user satisfaction. Companies with mature enablement programs report 3.8 times higher ROI on their AI investments compared to ad-hoc approaches. The data on this is not ambiguous. Training is the multiplier.
A Harvard Business School field experiment with Kenyan entrepreneurs sharpens why. Giving everyone the same AI assistant lifted profits and revenues for already-strong performers by 10-15% — and dropped them by about 8% for weaker ones. Uniform access produced non-uniform outcomes. The variable was not the model. It was the user’s ability to convert outputs into decisions.
Companies are missing up to 40% of potential AI productivity gains due to gaps in talent strategy. Not gaps in tools. Not gaps in budget. Gaps in the investment that actually converts AI spending into AI returns.
I wrote about how AI access is becoming a talent play — organizations that restrict it lose their best people. That argument still holds. But access without training is like handing someone the keys to a Formula 1 car and pointing at the highway. You have enabled an expensive accident, not a transformation.
The vacation planner is not a bad employee. They are an untrained employee using a powerful tool for the only thing they know how to do with it. That distinction matters, because it changes your response from punishment to investment.
Does some personal use build genuine AI fluency? Yes — and not every experiment needs a business case. But there is a line between deliberate practice and mindless offloading. The organizations with the best fluency outcomes are not the ones with unmonitored access. They are the ones with structured enablement.
Walmart committed nearly $1 billion to AI upskilling for its 1.6 million US workers — an eight-hour Google AI certification for every employee, partnerships with OpenAI for customized training experiences. Their CEO explicitly stated they expect “roughly the same number of people” despite full AI deployment. Walmart bet that trained workers are more valuable than fewer workers plus more AI. That bet, at that scale, is a thesis about where value actually lives.
JPMorgan Chase took a parallel path. The bank rolled out its internally built LLM Suite to all 200,000 employees as opt-in, then tracked every use case against business outcomes. The result: 200,000 daily users within eight months, an 83% reduction in research time for analysts, and $2 billion in annual benefits matching their $2 billion AI investment costs. A 1:1 return that most enterprises cannot demonstrate — not because their AI is worse, but because they never built the measurement system to show it.
They did not ration compute. They instrumented it.
The Workslop Tax
There is a word for what happens when untrained people use AI for professional work. Researchers at Harvard Business Review call it workslop: “AI-generated work content that masquerades as good work, but lacks the substance to meaningfully advance a given task.”
41% of workers encountered it in the last month. Average rework time per instance: one hour and fifty-six minutes. For a 10,000-person organization, that translates to over $9 million annually in lost productivity. That is not an AI cost. That is a training cost wearing an AI mask.
The G-P 2026 AI at Work Report found 88% of executives are now worried employees are using AI to “perform productivity” — producing visible activity without underlying value — while 69% say the time their teams spend monitoring and reworking AI outputs has gone up over the past year. The productivity tax is real, and it is moving upward into managers’ calendars.
The reputational damage compounds the financial one. 50% of recipients viewed workslop senders as less creative, capable, and reliable. 42% saw them as less trustworthy. A colleague who sends one obviously AI-drafted email without editing passes the checking work to the recipient and spends credibility in the process.
Klarna learned this the expensive way. The Swedish fintech replaced approximately 700 customer service agents with an AI assistant, measuring only cost per interaction. Customer satisfaction cratered. CEO Sebastian Siemiatkowski publicly admitted “We went too far” and the company reversed course, rehiring human agents. Compute was cheap per unit but expensive in aggregate when it drove away customers and destroyed institutional capability.
That is the workslop tax at the organizational level. Not the cost of AI. The cost of AI without accountability.
The Leader’s Balance Sheet
I keep coming back to the same tension in every conversation with executives about AI budgets.
Be generous enough with access to attract and retain talent — because the organizations that restrict AI are losing people to the ones that don’t. Be demanding enough with accountability to justify the investment — because half a trillion dollars in annual AI spending with no measurement infrastructure is not strategy. It is faith.
Neither generosity alone nor austerity alone works. The balance is four commitments, and the sequence matters:
Enable broadly. No access gatekeeping. Restriction breeds shadow AI, and shadow AI is ungovernable by definition. Give everyone the tools. Remove procurement friction. Make sanctioned AI the path of least resistance.
Train deliberately. The 81-hour annual threshold from the EY research is the target, not a nice-to-have. Build it into performance expectations. Protect calendar time for it. Walmart’s $1 billion is not a cost center — it is the primary lever for converting AI spending into AI returns.
Measure rigorously. If you are in the 77% who cannot track AI ROI, you are flying blind. Build the value dashboards, not just the spending dashboards. Track what AI produces, not just what it consumes. The organizations doing this report 3.8 times higher returns.
Cut ruthlessly — but only after the first three have had a fair runway. Where investment produces nothing after enablement and training, stop spending. Not quietly — publicly. Show what you spent, what it produced, and why you are changing. Cutting before enabling and training is just rationing with extra steps. The transparency justifies continued generosity by proving you are not wasting money.
This is not new leadership. It is the same resource management discipline that tamed cloud cost explosions, governed software sprawl, and built FinOps practices from scratch in two years. The tool is different. The discipline is the same.
Ask what AI costs alongside what your people can do with it. Give them time to learn, measure what changes in the work, and use those results to decide where to keep spending.
References
[1] “Why GPU Prices Are Rising in 2026.” BuySellRam, 2026. https://www.buysellram.com/blog/why-gpu-prices-are-rising-in-2026-how-memory-economics-and-ai-are-reshaping-the-graphics-market/
[2] “AI Productivity Paradox: CEO Study.” Fortune, 2026. https://fortune.com/2026/02/17/ai-productivity-paradox-ceo-study-robert-solow-information-technology-age/
[3] Niederhoffer, K., Rosen Kellerman, G., Lee, A., et al. “AI-Generated ‘Workslop’ Is Destroying Productivity.” Harvard Business Review, 2025. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
[4] “EY Work Reimagined Survey.” EY, 2025. https://www.ey.com/en_gl/newsroom/2025/11/ey-survey-reveals-companies-are-missing-out-on-up-to-40-percent-of-ai-productivity-gains-due-to-gaps-in-talent-strategy
[5] “EY Work Reimagined Survey.” EY, 2025. https://www.ey.com/en_gl/newsroom/2025/11/ey-survey-reveals-companies-are-missing-out-on-up-to-40-percent-of-ai-productivity-gains-due-to-gaps-in-talent-strategy
[6] “The State of Enterprise AI in 2025.” Larridin, 2025. https://larridin.com/blog/state-of-enterprise-ai-in-2025
[7] “State of FinOps 2026 Report.” FinOps Foundation, 2026. https://data.finops.org/
[8] “Shadow AI: Navigating the Growing Threat of Ungoverned AI Adoption.” Help Net Security / Delinea, 2025. https://www.helpnetsecurity.com/2025/11/12/delinea-shadow-ai-governance/
[9] “The State of Enterprise AI in 2025.” Larridin, 2025; “EY Work Reimagined Survey.” EY, 2025.
[10] “2025: The State of Generative AI in the Enterprise.” Menlo Ventures, 2025. https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/
[11] “Is AI Running Out of Compute?” J.P. Morgan Asset Management, 2026. https://am.jpmorgan.com/us/en/asset-management/liq/insights/market-insights/market-updates/on-the-minds-of-investors/is-ai-running-out-of-compute/
[12] “The AI Supply Crunch Is Here.” The Economist, April 2026. https://www.economist.com/leaders/2026/04/30/the-ai-supply-crunch-is-here
[13] “Uncover AI ROI and Cost Optimization Strategies That Work in 2026” (citing Gartner 2026 and Forrester). Svitla, 2026. https://svitla.com/blog/ai-roi-and-cost-optimization-in-2026/
[14] “Why 56% of CEOs See Zero AI ROI - And How to Reverse It in 2026.” VBeyond Digital, 2026. https://www.vbeyonddigital.com/blog/why-56-of-ceos-see-zero-ai-roi-and-how-to-reverse-it-in-2026/
[15] “The AI Reckoning: 73% of Executives Report Underwhelming ROI” (G-P 2026 AI at Work Report). G-P / PR Newswire, 2026. https://www.prnewswire.com/news-releases/the-ai-reckoning-73-of-executives-report-underwhelming-roi-from-ai-efforts-as-focus-shifts-from-hype-to-high-stakes-pressure-testing-302768787.html
[16] “AI Won’t Make the Call: Why Human Judgment Still Drives Innovation.” Harvard Business School, 2024. https://www.hbs.edu/bigs/artificial-intelligence-human-jugment-drives-innovation