- IBM’s 2025 Cost of a Data Breach Report found shadow AI adds an average $670,000 to breach costs per incident, placing unsanctioned AI use among the top three breach cost drivers.
- Forcepoint’s September 2026 report identifies unbounded LLM consumption as a budget vulnerability: costs spike past approved thresholds with no security alert triggered, leaving an invoice as the first signal.
- A 2024 Writer survey found 61% of companies reported accuracy problems with in-house generative AI, with only 17% rating their projects as excellent, a reliability gap that surfaces as a production blocker, not just a development annoyance.
Shadow AI’s $670,000 per-incident premium on breach costs, documented in IBM’s 2025 Cost of a Data Breach Report, is only part of what unsanctioned AI use costs enterprises. A September 2026 Forcepoint report adds a second exposure: LLM deployments with no compute caps are draining budgets without triggering a single security alert, a pattern Forcepoint calls “denial of wallet.” Together, they describe a cost problem that most security budgets were not built to handle.
The Hidden Cost Stack
IBM’s breach data puts numbers to what many finance teams already suspect. Organisations with high shadow AI exposure faced average breach costs of $4.74 million, a $670,000 premium over those with low or no unsanctioned AI use. That figure covers only the breach itself. Unmanaged AI use also generates quieter costs: underutilised enterprise licences, employees buying duplicate personal subscriptions, and data flowing through tools that were never reviewed for compliance.
A November 2025 analysis by MILL5 estimated that between 30% and 50% of AI-related cloud spend goes to idle resources, overprovisioned infrastructure and poorly optimised workloads. Enterprise AI budgets were projected to average $85,521 per month in 2025, yet only around half of organisations could confidently evaluate whether those investments were returning value. Inference spikes and training jobs with unpredictable runtimes push teams to overprovision compute as a hedge against degraded performance, which compounds both problems.
Denial of Wallet
Forcepoint’s September 2026 report focuses on what OWASP ranks sixth in its 2026 Top 10 for LLM Applications: unbounded consumption. The vulnerability is a missing control over how much compute, cost, or resource a single request can use. Without a cap, runaway processing is possible through volume alone, misconfigured automation or long sessions, no technical expertise required, no malicious intent necessary.
The financial exposure is asymmetric. Services typically stay operational while costs spike past budgeted amounts, so the first signal is often an invoice rather than an incident alert. No single request looks anomalous. Cumulative consumption escalates without crossing any threshold that conventional monitoring would flag. Computational costs scale with model parameter count, and continuous data pipeline overhead creates cost multipliers that do not appear in upfront licensing comparisons, a structural difference from traditional enterprise software that procurement teams are still pricing in. The OWASP 2026 framework’s shift toward blast radius containment reflects exactly this gap between where security tooling looks and where AI costs actually accumulate.
Developer Trust Is Sliding
A 2024 survey by Writer found that 61% of companies reported accuracy problems with in-house generative AI solutions, and only 17% rated their projects as excellent. High-profile failures across 2023 and 2024, image generation controversies, AI-generated electoral misinformation, voice clone fraud and biased hiring tools, gave those concerns public visibility. The pattern has carried into enterprise procurement: reliability questions that begin in developer tooling surface later as blockers in production deployments. With Gartner projecting AI agent inference costs rising fivefold by 2028, the reliability-per-dollar question becomes harder to defer.
Regulatory Costs Are Already Live
The EU AI Act’s penalty structure for prohibited practices reaches 7% of worldwide annual turnover or €35 million, whichever is higher, with enforcement for high-risk AI systems phasing in from 2026 through 2028 depending on system category. That is a recurring compliance cost, not a one-time remediation.
Governance structures have not kept pace with regulatory exposure. A 2024 Gartner poll found 55% of organisations reported having an AI board, but accountability within those structures remains diffuse. A 2026 Gallup survey found only 27% of Americans trusted businesses to use AI responsibly, down from 31% a year earlier, a public confidence deficit that creates reputational risk alongside regulatory risk.
Governance as Cost Control
The companies best positioned to contain these costs share a common trait: centralised visibility over what AI tools are running, what they cost, and who approved them. Clear procurement policies, approval gates for new generative AI solutions, and regular audits of existing deployments address the shadow AI premium IBM quantified. Compute caps and consumption monitoring address the denial-of-wallet exposure Forcepoint flagged. Neither requires a new platform; both require someone to own the outcome.
Centralised oversight also closes the licence redundancy problem. When departmental teams procure tools independently, duplicate subscriptions accumulate and approved models get bypassed in favour of faster, unreviewed alternatives. The IBM breach data suggests the cost of that bypass averages $670,000 per incident before any remediation work begins. Governance applied at the procurement stage is cheaper than governance applied after a breach or a budget overrun, a straightforward calculation that AI board structures, where they exist, are still working through.
