47% of Surveyed Companies Are Over Their AI Budget. Few Cut Projects, Many Look for Money Elsewhere

47% are already over budget — but few companies cut initiatives

Futurum's "2H 2026 CIO & Technology Buyers Global Enterprise Decision Maker Survey" splits surveyed companies' AI spending into three main groups: 46.9% over plan (35.6% moderately over plan + 11.3% substantially over plan), 31.8% approximately on plan, 5.6% under plan. A fourth group matters just as much: 10.0% of companies say they have no formal AI budget to measure against at all — a governance gap in its own right.

The pattern repeats when it comes to where the money is found. When the overrun is covered by reallocating within the IT budget, most companies cut external vendors first, not the AI project itself. And more telling still: nearly a quarter of over-budget organizations are already moving money out of other business units' budgets, not just IT — evidence, Futurum says, that AI funding authority has begun to move outside the IT organization.

For a CEO or CFO, the practical takeaway isn't "AI is expensive". That part was already known. The takeaway is that when the budget is exceeded, the company usually doesn't stop the project. The most common responses are asking for more budget or absorbing the overrun. If that money systematically comes from cut contractors or budgets borrowed from other departments, controlling AI cost stops being a purely technical conversation — it becomes a capital-allocation one.

Why agents can amplify the AI cost problem

Futurum's survey measures AI spending in general — it doesn't prove that AI agents specifically are the cause of the budget overruns it already found. There is, however, a structural reason the risk grows where usage is expanding: from chat toward autonomous agents.

A chatbot spends tokens (the units of text AI models bill for) when a human sends it a prompt: consumption is tied directly to a human action, easy to anticipate. An agent can trigger spending without direct supervision at every step: repeated tool calls, subagents launched in parallel, automatic retries when a step fails, goals pursued across multiple stages without a human checking in at each one.

McKinsey synthesizes the mechanism from recent research on coding agents, cited in its own article: in a programming task, an agent can consume roughly 1,000 times more tokens than a chat dialogue about the same coding problem, because models are stateless, meaning they don't remember prior calls, so the agent often resends the entire context as it progresses. Roughly 60% of an agentic task's cost comes from checking, repairing, and re-verifying the work, not the initial answer. And in programming, the same task can vary by a factor of up to 30 in cost between two runs, because the agent may choose different paths, tools, or numbers of retries each time. Cost behaves as a wide distribution, with large swings from one run to the next — a budget built on average cost per task won't be enough.

That doesn't mean the overruns in Futurum's survey come from agents. The survey doesn't measure that. It means agents add a new source of cost variation, on top of the pressure that already pushed nearly half of surveyed companies over budget.

McKinsey manages AI consumption permanently. In EY's survey, only 64% monitor tokens

McKinsey processes roughly 5 trillion AI tokens a month internally (a figure from May 2026). The firm placed no usage limits at first; it's now shifting to guidelines and guardrails, ahead of the next wave, in which autonomous agents will consume intelligence on behalf of users rather than simply responding to prompts. Managing AI demand has become, McKinsey says, a permanent capability, not a one-time exercise. Separately, the firm has drawn a clear conclusion from its own experience: roughly 10% of users generate roughly 65% of total token consumption. It explicitly recommends model routing (expensive models only for the steps that actually need them, cheaper models for the rest of the workflow) and pooling: grouping consumption spend at the organizational level, instead of capacity locked into individual licenses. McKinsey also performs prompt optimization and has invested in an internal AI gateway that routes requests to model vendors.

There's also a thematic link to what we've already written on massai.ro about the AI usage gap between leading companies and typical ones: according to McKinsey's "State of AI 2026" survey (1,719 participants, fielded between May and June 2026, cited by McKinsey ahead of publication), roughly one-fifth of companies say they've already constrained AI use because of operating costs — a further sign the problem isn't confined to Futurum's survey.

EY's survey of 534 senior US leaders confirms from another angle just how widespread the concern is: 82% of senior leaders whose companies invest in AI are concerned about token consumption, and 98% of those already using token-billed tools say those costs have made them reconsider their approach. In the same survey, only 64% of leaders at organizations investing in AI say they actively monitor token consumption and have clear budget guardrails in place — a gap between concern and control that takes the same shape as what we've already documented on AI agent permissions: concern runs high, verifiable control lags behind.

What your company should measure

The recommendations below are MassAI's own, built on the mechanisms McKinsey and EY document above — translated into practice for an ordinary company, not an organization processing trillions of tokens a month.

• Cost per accepted outcome, not cost per token or per prompt — an answer you have to fix by hand isn't "cheap" just because it used few tokens.
• Cost per workflow, measured separately from cost per model — the same process can cost differently depending on how many steps, tools, and retries it involves.
• Identify your heaviest users — if 10% of users generate most of the consumption, that's where control needs to concentrate, not spread evenly across the organization.
• Model routing with a clear owner — who decides when an expensive "frontier" model gets used and when a cheaper one is enough, and who reviews that decision periodically.
• Caching for repetitive steps — a practical MassAI recommendation, not a research finding: many agentic workflows repeat the same queries or checks, and caching cuts cost without cutting functionality.
• A named AI budget owner, not "IT in general" — with real authority over whether to ask for more budget, absorb an overrun, or stop an initiative.
• Alert thresholds before the overrun happens, not just after-the-fact reporting. If nearly half of companies end up over budget, the realistic goal isn't "zero overruns" — it's catching the overrun before it turns into an emergency reallocation.

None of this removes the risk that AI will cost more than you planned. As usage expands from simple chat interactions toward autonomous-agent workflows, the structural risk grows. But the difference between a company that can afford to scale AI and one that stalls at the first budget overrun rarely comes down to how much it spends. It comes down to how fast it sees the overrun and how clearly it knows who decides what happens next. It's also the natural follow-on to the question we already raised in our article on the cost-per-task of AI models: once you've picked the right model for each task, the harder question follows — how do you control consumption at the organizational level, not just at the level of a single task.

Sources: ↗ Futurum Research — 46.9% of Enterprises Report AI Spend Over Budget in 2H 2026 · ↗ McKinsey — Is that AI agent worth it? Agentic economics and the modern operating model · ↗ EY — C-Suites Pivot from AI Adoption to Unlocking Value as Escalating Token Costs Trigger Fiscal Scrutiny