The Takeaway: Enterprise AI is no longer a demo problem; it’s a budgeting, security, and workflow redesign problem.

• The real shift isn’t “AI everywhere,” it’s AI leaving IT budgets and colliding with line-of-business spending, where finance teams don’t yet have the tools to manage token burn.
• The biggest bottleneck is not model quality but rollout speed: every new breakthrough can obsolete the architecture companies just finished implementing.
• Coding is ahead because it’s verifiable and technical users can self-correct; the rest of knowledge work is harder because context is scattered and access controls are messy.

Aaron Levie, CEO of Box, has spent years building the bridge between bleeding-edge tech and enterprise reality, and his read is blunt: the industry is still early, even if it feels saturated. Chat got enterprises comfortable, but agents are the next fight. As he puts it, the question now is how to move from “I ask a question, I get an answer back” to systems that actually do work across the business.

What’s changed most is the mood. CIOs are no longer debating whether AI is real; they’re watching engineering teams use Cursor, Codex, and Cloud Code to ship faster, and business leaders are asking for the same lift. But the excitement comes with a hard reset on economics. Token costs are no longer a rounding error. Levie notes that one coding agent can burn “$1,000 of compute on a single task,” which makes old subscription pricing models look absurdly outdated.

His bigger point: the enterprise will end up with a mosaic of models, not one magic system. Frontier models will handle the hardest tasks, then lower-cost models or even plain software will take over once the workflow is stable. The winners won’t be the loudest AI believers; they’ll be the companies that can measure value, control spend, and decide when not to use an agent at all.