Enterprise AI Budgets Ran 2.5x Over Plan as Spending Shifts to Production

Budgets underestimated production deployment costs by 150%

Enterprise AI spending reached $176.8 billion by September 2026, up from $33 billion in 2025, according to Menlo Ventures' State of Enterprise AI report surveying nearly 400 U.S. decision-makers overseeing more than $4 billion in AI spending. The significant finding for procurement: surveyed companies spent 2.5 times their original budgets.

The overspend reflects a shift from pilot programs to production infrastructure. Companies initially budgeted for model API costs and experimentation, but production deployments required additional spending on data-security controls, governance platforms, deployment services, and enterprise-grade monitoring. This pattern creates a procurement problem: AI business cases built on pilot economics fail when scaled.

The spending mix shows where money actually went. AI applications consumed approximately $77 billion, up from $19 billion. Within that category, coding tools accounted for $34 billion and AI assistants plus productivity tools represented $25 billion—together making up 76% of application spending. That concentration matters because these categories increasingly compete with existing software budgets rather than drawing from separate innovation pools.

Existing software budgets now fund 22% of AI growth

New budgets supplied 55% of spending growth, but 22% came from existing software budgets, largely through incumbent vendors converting customers to AI tiers. This changes the buying analysis. Enterprises should compare incremental AI pricing against displaced software licenses, usage-based API costs, implementation fees, and platform lock-in—not just model quality in isolation.

ServiceNow's October 28 earnings call illustrated this dynamic. The company raised its 2026 target for AI annual contract value to more than $1.7 billion and reported that deals involving five or more AI products increased more than 8.5 times year over year. ServiceNow is bundling deployment expertise into selected large contracts, which positions it against not only Salesforce, Microsoft, and agent vendors, but also systems integrators and internal AI teams.

The bundled-services approach may shorten time to value, but it obscures software-versus-services economics. Buyers should request separate pricing for licenses, implementation, usage, and ongoing model consumption, and should test whether multi-product commitments create meaningful discounts or simply increase platform dependence.

Model competition concentrates around three providers

Anthropic, OpenAI, and Google captured 69% of deployment spending, up from 50%. Anthropic's share of enterprise LLM API spending rose from 40% to 55%, while Google's fell from 21% to 9%. That concentration increases pricing risk and reduces buyer negotiating leverage.

National Australia Bank's deployment of its Ruby AI tool, built on Anthropic's platform, to more than 40,000 staff provides a concrete example of regulated enterprise adoption at scale. Large-scale employee rollouts make identity controls, data residency, audit logs, prompt retention, and measurable productivity outcomes central purchasing criteria. A 40,000-user deployment also raises the risk of uncontrolled consumption costs unless access tiers, usage limits, and business-unit chargebacks are defined before expansion.

Security and governance create separate budget category

Data-security company Sentra announced a $150 million strategic investment led by Capital One Ventures on November 4. The company focuses on protecting data at rest and controlling what AI agents can access. The financing reflects a shift in enterprise-AI risk spending from model selection toward the control plane around agents.

Amazon's November public preview of Bedrock Managed Agents, powered by OpenAI, places AWS directly against Microsoft Azure AI Foundry, Google Vertex AI, Salesforce Agentforce, and ServiceNow's agent products. Managed agents can reduce engineering and operations costs, but buyers should validate portability, model-switching support, tool-permission controls, observability, and per-task pricing before treating the service as a production standard.

Model selection becomes an economics problem

A GDPx benchmark published November 2 reported the following results for professional-domain agent tasks: Claude Opus 5.5 scored 12.0% at $51.41 average cost per task; Gemini 4 Carbon scored 10.9% at $16.29; GPT-6 Astra scored 9.6% at $53.99. Gemini's cost per task was roughly one-third of the two higher-cost alternatives while scoring relatively close on the benchmark.

Enterprises should evaluate models using task-level cost, success rate, latency, human-review requirements, and failure remediation—not generic benchmark rankings. A cheaper model can be preferable for high-volume workflows even if a more expensive model leads on absolute task performance. Conversely, a higher-cost model may be justified where errors create regulatory, financial, or safety exposure.

What to watch

The 2.5x budget overspend pattern indicates that enterprises are underestimating deployment, governance, and security costs when building AI business cases. Procurement teams should budget separately for model API costs, security controls, deployment services, and productivity measurement—and should expect existing software vendors to convert contract renewals into AI-tier pricing rather than positioning AI as incremental spend. Model consolidation around three providers increases pricing risk and reduces negotiating leverage unless contracts include portability provisions and usage caps.