Enterprise AI shifts from pilots to production controls

Enterprise AI programs are entering a harder phase as pilots give way to production systems that must be sustained, governed and defended. MIT’s NANDA initiative found that 95% of enterprise generative-AI programs have produced no measurable P&L impact despite $30–40 billion in spend, while 42% of firms abandoned most AI initiatives in 2025, up from 17% a year earlier.

The central operating metric is shifting from token costs to cost per useful task completed. Workloads need to be matched to the simplest reliable model and routed according to business value, latency, sensitivity and control needs, with public cloud suited to experimentation and private or sovereign environments better aligned to high-volume or regulated use cases.

Governance is becoming inseparable from deployment as agents move from answering questions to taking action across workflows. Effective programs require clear permissions, shared policies, audit trails, orchestration and human checkpoints for higher-risk actions. Defense means designing for failures through tiered autonomy, traceability, kill switches, policy overrides and selective human review, so oversight protects trust without undermining efficiency.