The investment committee approved the annual AI budget three months ago. Line items agreed, forecasts attached, a figure the organisation could absorb over the financial year. Since that approval, AI agents in procurement, customer service, and operations have been committing spend against those lines in increments of cents, thousands of times per day. No human sits in the approval chain. No mechanism reports the cumulative total back to the executive who owns the budget.
In Brief
- AI agents commit spend in real time against budget lines governed by annual review cycles that cannot see or cap the spend.
- Three control gaps — visibility, spending caps, and accountability — exist because annual governance was never designed for autonomous consumption.
- A single uncapped agent can exhaust a quarterly budget allocation overnight without triggering any procurement control.
- The funding governance decision is setting a maximum spend threshold per agent before a human must review and reauthorise.
That is not a technology failure. The agents process documents, triage queries, draft responses. They are doing what they were deployed to do. The failure sits in the funding governance wrapped around them: governance designed for commitments that moved slowly enough for a quarterly review cycle to catch, and for expenditure that required a human to approve each purchase before the money left the organisation. Neither condition holds when AI agents are the ones committing the spend. The result is a funding governance structure that governs everything except the fastest-growing line of technology expenditure in the organisation.
Three gaps annual governance cannot close
Three control gaps sit between the annual budget approval and the real-time spend it is supposed to govern.
Gap #1: The running total is invisible
Most organisations report financials monthly or quarterly. AI agent spend (application programming interface (API) calls, compute consumption, third-party model fees) accumulates continuously. The executive who approved the budget line cannot see what has been spent against it until the period closes and finance reconciles the data. By the time the number surfaces in a report, weeks of unmonitored spend have already passed. The oversight mechanism is a rear-view mirror. Real-time dashboards exist for cloud infrastructure costs. Most organisations have no equivalent spend visibility for AI agent consumption.
Gap #2: Nothing caps a single run
Traditional IT expenditure comes with procurement controls: purchase orders, vendor contracts, approval thresholds. An AI agent making API calls against a cloud-hosted model triggers none of them. There is no purchase order for the ten-thousandth call. A single agent processing a backlog of documents overnight can consume a full quarter’s budget allocation before anyone checks the output the next morning. The absence of a real-time spending cap is not an oversight in most organisations’ governance design. It is a category of control that never existed, because no prior technology required one.
Gap #3: Accountability sits between two parties
Annual budgets attach accountability to the executive who owns the line. That accountability assumes the executive had the opportunity to approve each material commitment against it. When an AI agent commits spend autonomously, the budget owner did not approve the individual commitment. The team that deployed the agent may not have anticipated the spend volume it would generate. Accountability lands between two parties, neither of whom made the spending decision that produced the overrun.
Prior technology never spent this way
These control gaps exist because the financial governance embedded in annual budget cycles worked reliably for every technology investment that preceded AI agents. Organisations purchased software licences annually. Infrastructure went through procurement processes with defined approval gates. Even cloud computing, for all its pay-as-you-go flexibility, was managed early on through reserved instances and negotiated enterprise agreements that fit annual budget cycles. The governance matched the cadence of the spend, and for decades that cadence was slow enough for the match to hold. No finance team questioned it, because nothing the organisation had deployed ever broke it.
AI agents are the first enterprise technology that commits real spend autonomously, continuously, and at a speed that outpaces every financial control that requires a human. The governance gap is not negligence. It is the predictable consequence of a consumption-based cost model running inside a funding governance structure built for commitment-based costs, one that never needed to accommodate anything else.
Uncapped agents can exhaust quarterly budgets before anyone checks
Gartner (2025) projects worldwide generative AI spending will reach US$644 billion in 2025, a 76.4% increase from 2024. That figure matters less for its size than for what it reveals about deployment velocity. Organisations are putting AI agents into production faster than they are building the financial controls to govern what those agents spend.
A department approves a $200,000 annual budget line for AI-assisted document processing. The agents deployed against that line make API calls to a cloud-hosted large language model, each costing between one and four cents. At 10,000 documents per day, the daily spend runs between $100 and $400. At the lower bound, the annual budget lasts the year. At the upper bound, it is exhausted by September. The executive who approved the line will not see the variance until the Q3 financial review. By then the budget is gone and the agents are still running.
That is the exposure on a single departmental line. Across an organisation running multiple agent clusters in production, the same dynamic operates on every line that funds AI consumption. Each quarter without real-time spend visibility and automated spend caps, the gap between approved budget and actual consumption widens. The CFO’s ability to reforecast accurately erodes with it. The risk is not that one line overspends; it is that no one knows which lines are overspending until the variance is already in the accounts. Every new agent cluster deployed into production without a spend threshold adds another unmonitored line to the CFO’s exposure.
One decision for this month
The funding governance most organisations have inherited cannot see AI agent spend, cannot cap it, and cannot hold anyone accountable for it, because it was never designed to try.
That is not a criticism of the leaders who built it. Annual budget cycles were designed for annual commitments, and they worked effectively for every technology investment the organisation made before this one.
What this means for senior leaders
- The AI program owner in every organisation deploying agents into production needs to make one funding governance decision this month: the maximum any single agent or agent cluster can spend against a budget line before a human reviews and reauthorises the commitment. A rough threshold is better than none.
- That threshold is a funding governance decision, not a technology decision. Until it is made, the organisation’s AI spending is governed by whatever the agents happen to consume.
- Setting the threshold creates the spend visibility and accountability that annual funding governance currently lacks. The executive who sets that threshold before the next quarterly review closes is working with a different cost structure to the one who discovers what happened after the accounts are reconciled.
References
- Gartner. (2025, March 31). Gartner forecasts worldwide GenAI spending to reach $644 billion in 2025. https://www.gartner.com/en/newsroom/press-releases/2025-03-31-gartner-forecasts-worldwide-genai-spending-to-reach-644-billion-in-2025
