The headline number on enterprise AI has finally turned encouraging, which is exactly why boards should pay closer attention. Lenovo's IDC-backed CIO Playbook 2026 finds that nearly half of AI proofs of concept, 46%, have now moved into production, with organizations anticipating returns of up to $2.79 for every dollar invested.¹ After two years of headlines about pilot graveyards, enterprises have learned how to put conventional AI into production. The blockage has moved. It has not disappeared.

The same study, however, contains a warning that should stop any CIO mid-celebration. While 60% of organizations are in late-stage AI adoption, only 27% have a comprehensive AI governance framework in place.² That gap becomes acute as soon as enterprises turn to agents. Only 21% report significant use of agentic AI today, with most organizations still exploring or piloting it, and three in five say they are more than a year away from being ready to scale it across their operations.²

Agentic AI, it turns out, follows a separate path to production, and most enterprises have not completed it, let alone reached scale. A 2026 Wakefield Research survey of 1,000 senior technology and data leaders, commissioned by Teradata, found that only 7 percent of enterprises had reached the final stage, where tangible business outcomes occur, while 68 percent remained in the experimentation or development stages.³ India was not among the six markets surveyed, so these agentic-maturity findings should be read as a directional benchmark from comparable global markets rather than as India-specific evidence.

Sachin Chedda - Vice President, Strategic Alliances and Business Development at Nutanix

What makes agents different from GenAI's first wave

For most enterprises, the 2023 to 2024 wave of generative AI was additive and relatively low-stakes. A model drafted an email, a human checked it, and nothing in production changed. Agents break that pattern in four specific ways. They persist instead of answering once. They act instead of suggesting. They consume compute and model calls continuously rather than on demand. And when they are wrong, the consequences are real. An agent that reconciles a ledger, reroutes a shipment, or adjudicates a claim is not simply a productivity feature. It is part of the operating business. The failure that matters is no longer a wrong answer. It is a wrong action, taken autonomously and at scale, with no one watching.

If the model is not the problem, what is?

Frontier intelligence is becoming close to a commodity. Any enterprise can rent a capable model. What it cannot rent is a specific bank's risk logic or a specific manufacturer's plant-scheduling reality. That proprietary knowledge, the context of how a particular business actually works, is information the........

© Forbes India