The real cost, security, and culture problems behind enterprise AI agents

Red Hat’s Brian Gracely says scaling agentic AI moves costs from a technical issue to a recurring board-level concern, driven by orders-of-magnitude higher token/usage spend versus the chatbot era. He argues enterprises can cut agent costs fastest by right-sizing models (via semantic routing and caching) and adopting FinOps-like token discipline, while also accelerating security patch cycles to a ~7–14 day window as AI speeds vulnerability discovery. Adoption success further depends on subject-matter expert buy-in and incentives, implying operational friction can stall deployments beyond pilots.

Analysis

The key market implication is not that agentic AI slows down, but that monetization migrates away from premium model layers toward control planes, orchestration, and governance. If enterprises standardize semantic routing and caching, the first-order loser is per-token pricing power; the second-order winner is any vendor that sits between the app and the model and can enforce policy, cost controls, and auditability. That favors hybrid-infra incumbents and security tooling more than headline AI brands.

The more underappreciated effect is budget ownership shifting from product teams to CFO/procurement within one to two quarters. Once token spend becomes visible at scale, enterprises tend to squeeze variance before they expand volume, which usually compresses usage-based growth rates for AI-enabling software even as adoption headlines stay strong. In other words, the market may be overestimating how much near-term revenue accrues to frontier-model ecosystems and underestimating how quickly customers will arbitrage down to cheaper models for routine tasks.

On security, faster vulnerability discovery shortens the value of slow, calendar-driven patching and rewards vendors that can automate validation, rollout, and compliance workflows. That is a structural tailwind over 6-18 months for exposure-management and secure-infrastructure names, but it also raises execution risk for enterprises with fragmented IT estates: the adoption bottleneck is now process discipline, not model quality. The contrarian view is that the bottleneck may actually accelerate consolidation, because firms that cannot operationalize cost and patch governance will simply buy more managed platforms rather than build internally.

Key Decisions for Investors

• Long IBM vs. short MSFT, sized modestly, as a 1-3 month relative-value expression of enterprise demand for model-agnostic control and cost governance; cut the pair if MSFT’s AI revenue commentary shows no sign of margin pressure or if IBM fails to convert AI messaging into bookings.
• Buy QLYS or TENB on pullbacks for a 1-3 month trade into budget season, on the thesis that compressed patch windows force spending toward automated exposure management; the thesis breaks if vulnerability volumes do not translate into higher renewal rates or if security budgets get frozen.
• Underweight or hedge the most token-intensive AI application names in the software basket on rallies, since route-to-cheaper-model behavior can cap usage growth and multiple expansion; treat this as an alert until next-quarter disclosures show whether token spend is being managed or simply passed through.
• Monitor enterprise earnings calls for mentions of semantic routing, caching, and model mix optimization; if those terms remain absent by the next reporting cycle, the market is probably still pricing agent adoption as pure upside and the cost-discipline thesis is not yet investable.