Frontier AI models cannot accurately predict their own token consumption before a task runs, and identical tasks can vary by up to 30x in total token spend, according to Microsoft Research’s May study, the first systematic analysis of token consumption patterns in agentic coding tasks. Enterprises are already hitting this constraint through blown budgets. Forrester’s July analysis cites four named examples: Uber burned its annual AI budget in four months, Microsoft ended Claude code licenses after exhausting its yearly AI allocation, Tesla capped AI spending at $200 per week, and Priceline faced unexpected surges in AI development renewal costs.
CIOs, finance leaders, and IT operations teams are the ones left holding those invoices. The operational consequence is concrete: organizations that treat AI agent spend as a fixed procurement cost, rather than a dynamic consumption signal requiring runtime controls, will continue exhausting budgets ahead of schedule with no warning mechanism in place.
Why AI agent infrastructure costs resist forecasting
The Microsoft Research study analyzed trajectories from eight frontier models on SWE-bench Verified and tested whether models could predict their own token usage before execution. They cannot. Self-predicted token usage correlated with actual usage at no better than 0.39, a weak-to-moderate relationship, and models systematically underestimated real costs.
Agentic tasks are expensive in ways that standard AI usage is not. They consume up to 1,000 times more tokens than code reasoning or code chat, and input tokens drive the majority of that cost. Those inputs accumulate through context windows, tool-call outputs fed back into the model, and multi-step reasoning chains that expand during autonomous execution. Every additional tool call adds tokens that were not in the original task estimate. The more autonomous the agent, the more opportunities it has to generate costs without any human checkpoint, Lumenova noted this month.
Two findings from that research complicate budget modeling further. Task complexity as rated by human experts only weakly aligns with actual token consumption a task that looks simple to a developer may cost far more to execute than one that appears complex. Separately, accuracy peaks at intermediate cost levels and plateaus before spend peaks. Higher token usage does not produce proportionally better outcomes. Budgeting by volume rather than by task design wastes money without improving results.
Model selection carries its own spread. On identical benchmark tasks, Kimi-K2 and Claude Sonnet 4.5 consumed an average of more than 1.5 million more tokens than GPT-5. Provider rate cards from OpenAI, Google, and Anthropic describe per-unit cost but say nothing about how many units a given workflow will consume, ZDNET reported in May. The per-token price is only half the equation; how many tokens a chosen model burns on a given workflow is the variable no rate card reveals. One emerging approach to this problem is intelligent routing — Meta’s Switchboard system for routing simpler AI tasks to cheaper models represents exactly the kind of cost-optimization architecture enterprises are beginning to evaluate.
Traditional FinOps frameworks compound the problem. They were built for infrastructure workloads with relatively stable consumption patterns and were not designed to handle model inference pricing, dynamic token consumption, or multi-model routing, Forrester found this month. AI coding platform Cursor encountered this directly in 2025, transitioning from a flat subscription model to usage-based pricing after increasingly capable models made predictable flat-rate pricing unsustainable, Lumenova reported. What was true for a developer tool vendor is now a planning reality for every enterprise IT organization deploying agents at scale.
How scale and shadow AI widen the AI agent cost management gap
Individual task variance becomes a budget overrun when agents run at volume. Lumenova’s July analysis identifies the highest-cost patterns in sanctioned deployments: inefficient execution loops, oversized model selection for routine tasks, verbose prompt construction, excessive tool calls, and always-on or multi-agent workflows running without human checkpoints. A 30x variance at low volume is a rounding error. At hundreds or thousands of daily tasks, it is a material line-item problem.
The scale risk has a documented ceiling. Earlier this year, a research project creator disclosed that approximately 100 autonomous coding agents had consumed more than $1.3 million worth of OpenAI API tokens in 30 days, driven by millions of requests running without spending controls, Lumenova reported. OpenAI covered the cost because it was a supported research project. Enterprise deployments do not carry that backstop. The broader AI platform race is accelerating this risk — Google’s Gemini reaching 950 million monthly users signals how rapidly organizations are adopting AI at a scale that outpaces existing financial controls.
Unsanctioned AI tool use adds a separate and compounding visibility problem. Employees used between three and five unsanctioned AI tools per person per month in 2024, and 69% of organizations suspect prohibited AI tool usage is occurring on their networks, according to ShadowLock’s State of Shadow AI 2026 report. Network-only detection misses roughly half of that activity. Even organizations with well-instrumented sanctioned platforms face an incomplete picture of total AI spend until endpoint-level controls are deployed alongside network monitoring.
Shadow AI is not a marginal risk. ShadowLock’s research finds that effectively every organization that has measured for unsanctioned AI activity has found it. That spend does not appear in the sanctioned platform invoices finance teams review each month.
Governance and visibility controls that close the gap
Fewer than 25% of organizations have technical AI governance controls in place, per Gartner data cited in ShadowLock’s 2026 report. The most common audit finding across 2025 and 2026 is a written AI policy with no enforcement mechanism behind it. SOC 2 Type II audits now routinely include AI-specific control questions approximately 55% of audits surveyed asked about written AI policy, technical enforcement, and audit logs, the same report found. Organizations without enforcement behind their policies face both cost and compliance exposure.
Three operational controls address the gap directly.
Visibility: Instrument agent deployments to capture per-agent, per-task token consumption specifically input token volume by workflow and team before the monthly cloud invoice arrives. Per-task logs, tool-call frequency, and model selection records are what make chargebacks and rolling budget revisions possible. Without that granularity, governance defaults to aggregate monthly caps that arrive too late to change spending behavior for the period already billed.
Attribution: Assign AI spend to business units, use cases, and individual agent workflows. Spend that cannot be attributed cannot be governed or contested. Attribution also produces the data record needed to negotiate enterprise contracts, benchmark future deployments against pilot-stage baselines, and answer the ROI question directly high AI spend does not confirm high value, and teams need per-workflow cost data to evaluate whether accuracy and productivity gains are materializing at current cost levels, Lumenova’s analysis notes.
Runtime guardrails: Apply spending limits and automatic throttling at the agent-workflow level, not only at the account level. Governance built during the proof-of-concept stage is significantly easier to enforce than controls retrofitted after production deployment. For unsanctioned usage, endpoint-level controls are necessary because network-layer monitoring will not surface roughly half of shadow AI activity. The same discipline applies to broader enterprise software governance — the credential security failures that exposed the Chick-fil-A account through password reuse illustrate how enforcement gaps at the endpoint level create compounding exposure across unmonitored access points.
The budget model shift finance leaders need to make is from fixed annual license assumptions to rolling, usage-based forecasts anchored to pilot-stage spend baselines by workflow type. Forrester’s July examples are instructive not because those organizations mismanaged AI, but because they all applied a procurement cost model to a workload that behaves like cloud infrastructure. AI agent spend is consumption-based by design. Until organizations instrument it, attribute it, and constrain it at runtime, the budget gap compounds with every new deployment.