{"id":72282,"date":"2026-06-12T22:27:12","date_gmt":"2026-06-12T22:27:12","guid":{"rendered":"https:\/\/www.europesays.com\/ai\/72282\/"},"modified":"2026-06-12T22:27:12","modified_gmt":"2026-06-12T22:27:12","slug":"why-ai-spending-keeps-outrunning-budgets","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ai\/72282\/","title":{"rendered":"Why AI Spending Keeps Outrunning Budgets"},"content":{"rendered":"

\n Agentic AI<\/a>
\n ,
\n Artificial Intelligence & Machine Learning<\/a>
\n ,
\n Next-Generation Technologies & Secure Development<\/a>\n <\/p>\n

Analysts Say AI Cost Problem Is Actually a Consumption Problem<\/p>\n

Jennifer Lawinski<\/a> \u2022
\n June 12, 2026 \u00a0 \u00a0 <\/p>\n

\"Why
\n Analysts and CIOs say the real driver in AI cost overruns is uncontrolled token consumption, not pricing. (Image: Shutterstock) <\/p>\n

As artificial intelligence proliferates across the enterprise, tech leaders are finding out they all have one thing in common: AI is expensive.<\/p>\n

See Also: AI Agents Introduce a New Insider Threat Model<\/a><\/p>\n

Examples of cost overruns and budgets gone wild are becoming commonplace. By mid-April, Uber spent<\/a> its entire 2026 AI budget. An unnamed enterprise told<\/a> Axios that its decision to give employees unrestricted access to Anthropic’s Claude with no usage caps or spending controls in place cost it $500 million in a single month. Microsoft quietly wound<\/a> down most internal Claude Code licenses for a major engineering division, having its developers use its own Copilot CLI instead. Amazon shut<\/a> down its leaderboard for internal AI usage because employees were gaming the system to rack up points, driving up costs.<\/p>\n

Taken together, these tell a story that’s reflected in recent research from Bain & Company, which polled 951 global companies for its 2026 Automation and AI Pathfinder Survey. AI is costing companies more, and saving them less, than they had expected. Nearly 40% of companies that measured cost savings from AI projects landed below 10%, despite having targeted 11% to 20%. At the same time, 90% of those companies are increasing their AI budgets again, even if the savings didn’t materialize.<\/p>\n

This misalignment, said Greg Zorella, principal analyst at Forrester, comes from how enterprises are using it.<\/p>\n

“Most of the pitfalls we see right now are consumption out of control,” Zorella said. “Even if you had a really tight, firm pricing model, you’d still have those events, because consumption just isn’t under control. If an autonomous agent is out of control, that becomes a recurring, compounding impact, and who knows when a human actually steps in to do something about it?”<\/p>\n

As enterprises shift from pilot to always-on production systems, they’re learning how many tokens these new systems are burning through, and it’s demanding a whole new strategy. CIOs are now building the governance and forecasting models that arguably should have existed before the bills arrived.<\/p>\n

“AI often starts as an experiment, but ends up behaving like infrastructure,” said Ha Hoang, CIO at Commvault.<\/p>\n

What’s Driving Up Spend<\/p>\n

Debo Dutta, chief AI officer at Nutanix, said that part of the reason the cost of AI has gone up so dramatically is structural. Until about a year ago, he said, vendors priced AI based on the number of questions and answers, and each exchange generated perhaps 2,000 tokens.<\/p>\n

Now, the math has changed, and more sophisticated systems that write code, run tests and check their own work through sub-agents cost much more. “What used to consume a couple thousand tokens per question now consumes 20,000 or more,” Dutta said. “That’s forced every vendor to change its pricing.”<\/p>\n

And as AI takes on more of the work that developers used to do manually, like triaging, unit testing and refactoring, the workload being put on AI is growing. “People are doing exactly what they should be doing,” Dutta said. “They’re using AI to automate their entire software engineering practice, but the pricing model assumed they were only doing a fraction of that.”<\/p>\n

Goldman Sachs predicts<\/a> that token consumption could increase 24 times by 2030, reaching 120 quadrillion tokens a month, driven primarily by agentic AI workloads.<\/p>\n

Dutta said the rough industry benchmark for engineering token spend now runs from $10,000 to $50,000 per employee annually, with power users at the high end. “Nothing beats coding in terms of spend,” he said. And what drives that spend up is something many enterprises and developers have a choice on: model tier.<\/p>\n

“By default, a lot of people are using the most expensive model, and that’s an order of magnitude more expensive than the lower-end models, which may be good enough for the task,” Dutta said.<\/p>\n

The other factor driving up costs is that often enterprises think they’re buying one thing, when in reality, AI projects drive up cost across several line items, from tokens to personnel.<\/p>\n

Bain’s survey found that only 7% of companies are running fully autonomous agents in production today. The dominant model, reported by 38% of respondents, is “human approval required,” and another 32% operate with guardrails and exceptions, meaning a person steps in whenever the agent encounters something it can’t handle confidently.<\/p>\n

The gap is wider among companies that missed their cost-savings targets: only 38% of those have agents operating at guardrails-level autonomy or above, compared with 50% of companies that hit their targets.<\/p>\n

That gap is exposing flaws in the way companies are building their AI business cases, said Michael Heric, partner and global leader of corporate support and service operations solutions and automation at Bain & Company.<\/p>\n

“For a lot of companies, the business case focuses on token spend and stops there,” Heric said. “They’re not factoring in the other one-time and recurring costs beyond the tokens, like the data engineering costs or the governance cost of keeping a human in the loop. Those should all be part of the business case.”<\/p>\n

Everything Old Is New Again<\/p>\n

Heric said that the cycle he sees playing out with AI is similar to the ways emerging technologies have rolled out in enterprises in the past. The same companies that are succeeding in AI today are those that were able to scale robotic process automation and cloud computing. The C-suite prioritized the technology, and IT built scalable technology and change management programs.<\/p>\n

The ones who are struggling now are the ones who struggled then, Heric said. “It’s just the stakes and the cost of those mistakes are way greater now.” A failed RPA pilot might cost a company a few hundred thousand dollars, maybe a few million at most, but for when AI goes wrong, “half a billion dollars is gone. That’s huge.”<\/p>\n

One factor that exacerbates the problem is the way technology spending is allocated across the enterprise budget. “All the benefits of AI tend to accrue to the business, and all the cost accrues to the IT organization,” Heric said. “It’s hard to draw a straight line between the cost of an AI use case and what goes directly into a budget.” That’s why, he said, “the partnership between the CFO and the CIO has never had to be greater.”<\/p>\n

Zorella said he also sees a gap between how quickly companies are deploying AI across the business and how quickly they’re building systems to manage the cost.<\/p>\n

“AI familiarity and sophistication are coming before spend management,” Zorella said. And the problem is compounded when business units decide to procure their own AI licenses outside of IT. “Bring your own AI” makes total spend even harder to track, Zorella said.<\/p>\n

Using the Right Model for the Job<\/p>\n

The answer isn’t capping token usage, as that creates tradeoffs companies may not want to make. If an agent hits a cap while executing a process, work could be cutoff midstream, and it may be difficult to know how important that process was, Zorella said.<\/p>\n

Instead, Zorella said enterprise technology teams need to put rules in place to make sure that the right models are being used for the job, a discipline he called “model tiering.” For example, not every employee needs tools that can build new agents just because the license includes them. “You don’t need agent-building capabilities if you’re not an agent builder,” he said.<\/p>\n

Nicholas Merizzi, principal at Deloitte Consulting, leading the firm’s AI infrastructure practice, said that model tiering is understood in principle, but under-implemented in practice. “The focus over the past couple years has been to show value, and it was not focused on cost. Organizations are hitting spend thresholds now that are forcing the question of value per token,” he said.<\/p>\n

There are blueprints and companies that offer model routing so the technology capability is there, Merizzi said, but internal politics can create challenges when it comes to setting policy. “Oftentimes the bigger obstacle is that model tiering can cause organizational friction as each product team might have a preferred model they want to use,” he said. “The practical unlock is treating model selection as an architecture decision with financial consequences, not a developer preference – and building routing logic that enforces it.”<\/p>\n

Unit costing is another strategy teams should use, Zorella said. Rather than setting an overall AI spending target, some companies are setting token budgets by agent and by task. “If you have 10 agents,” he said, “one agent can go crazy and use up the budget for the other nine.” Planning at the unit level, tied to a specific outcome, makes it possible to catch that before it happens.<\/p>\n

Calculating your total AI spend can also get messy when different teams, some using shadow AI products, are generating bills which could have different pricing structures – per user, flat subscription fee or pay-as-you-go per token. Companies need to aggregate that into a single picture before they can even get to regulating tiers and usage caps, Merizzi said.<\/p>\n

Zorella recommended that AI spend and the value it generates should sit in the same P&L. “That’s where you’re able to tell a holistic story about the value AI is generating for the enterprise.”<\/p>\n

Hoang said that token spend should be governed differently depending on how the tokens are being used to drive business value. “For experimentation, organizations can tolerate higher costs because they’re buying learning,” she said. “For operational efficiency initiatives, there should be clear expectations around cost savings and return on investment. Strategic capabilities may justify higher spending, but they still require visibility and accountability.”<\/p>\n

Zorella said he advises clients to start with the business goal and work backward. “Map the spend to an area of strategic importance for the company, not just to the cost of an individual agent,” he said. “If you’re spending heavily on use cases that aren’t aligned to company priorities, that’s a recipe for costs getting out of hand.”<\/p>\n

Heric said he expects it will take another year or two before CIOs and CFOs can forecast AI budgets with real confidence, because AI spending is now large enough to move the needle on company earnings. “It’s not going to be a rounding error,” he said. “It’s going to be significant.”<\/p>\n","protected":false},"excerpt":{"rendered":"Agentic AI , Artificial Intelligence & Machine Learning , Next-Generation Technologies & Secure Development Analysts Say AI Cost…\n","protected":false},"author":2,"featured_media":72283,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[6],"tags":[179,7493,39929,30374,7495,9727,39931,30553,39930,20047,39928],"class_list":["post-72282","post","type-post","status-publish","format-standard","has-post-thumbnail","category-agentic-ai","tag-agentic-ai","tag-agentic-artificial-intelligence","tag-ai-budget-governance","tag-ai-cost-management","tag-ai-roi","tag-ai-spending","tag-cfo-cio-collaboration","tag-enterprise-ai-costs","tag-model-tiering","tag-token-consumption","tag-token-costs"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/72282","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=72282"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/72282\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/72283"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=72282"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=72282"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=72282"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}