Two numbers frame this week, and they point in opposite directions.
The first: frontier inference costs have fallen roughly 50x year-over-year, with GPT-4-class capability now priced under $1 per million tokens (Berkeley AI Research). Intelligence is effectively free. That inverts the ROI calculus that has kept most enterprise agent programs in pilot purgatory — and the index registers it, with ROI Impact climbing to 59 and Workforce AI Access to 65. The economic objection to automation is dead.
The second: Gartner projects 40% of agentic AI projects will be cancelled by end of 2027, and fewer than 10% of enterprises have scaled agents to measurable value (AI Hive). The cancellations are not model failures. They are governance failures — integration complexity, absent risk controls, no prototype-to-production clarity.
Here is the argument: the cost of thinking collapsed, but the cost of trusting a thinker did not. That gap is the whole game now. When intelligence is free, competitive advantage stops accruing to whoever has the best model and starts accruing to whoever can safely put an agent in production. The bottleneck moved from the model layer to the Identity Control Surface — provenance, authorization, audit, and the machinery that lets a risk committee say yes.
The market is pricing this in real time. Cisco now sells agent identity as a security product, treating non-human agents as first-class principals requiring cryptographic authorization (Cisco Newsroom). Visa and Mastercard standardized on agent identity protocols in the same cycle, wiring Cloudflare's Web Bot Auth into payment rails (Digital Commerce 360). And Anthropic's Jacobian technique makes agent reasoning auditable rather than opaque (MIT Technology Review) — the technical substrate that turns a Governance & Ethics score of 76 from rhetoric into approval.
Notice the tension the index exposes: Governance scores highest, Brand scores lowest at 40. Enterprises have written the policy. They have not built the operational identity architecture that makes the policy executable. Deutsche Telekom's cross-functional OpenAI deployment (OpenAI) shows what closing that gap looks like — but it is the exception, not the pattern.
So the move for principals this quarter is not to shop for a better model. It is to answer three questions before your next agent reaches production: How does this agent authenticate? What can it authorize? And can you reconstruct why it decided what it decided? If you cannot answer, you are building on the foundation Gartner says collapses for four in ten.
Watch item: US officials now estimate unauthorized model distillation costs AI labs $6B annually — a signal that model weights are becoming sovereign capital assets. Monitor for regulatory action that converts this estimate into enforceable IP claims, because weight provenance and derivative-deployment authorization are about to become a board-level liability, not a procurement footnote.
WatchUS Officials Estimate Unauthorized Model Distillation Costs AI Labs $6B Annually — The Trump administration's $6B distillation-loss estimate signals that model weights are being treated as sovereign capital assets, not open infrastructure. If IP enforcement tightens, enterprises building on distilled or third-party-fine-tuned models face retroactive compliance exposure. Identity Control Surface for the AI supply chain — provenance of model weights, lineage of training data, authorization to deploy derivatives — becomes a board-level liability question, not a procurement footnote. Monitor for regulatory action that converts this estimate into enforceable IP claims against enterprise deployments.