Aaron Levie on enterprise AI in 2026: token shock, agent diffusion, and the rise of the internal FDE

Summary

Aaron Levie, CEO of Box, discusses what AI is actually doing inside Fortune 500 and Global 2000 enterprises in 2026, based on hundreds of CIO conversations. Levie argues the industry is still at "day one" of agents: most firms have only rolled out chat, and agents demand data infrastructure, access-control fixes, and workflow changes most have not done. His key themes: token costs are escaping IT budgets into line-of-business spending; faster lab breakthroughs paradoxically slow adoption because they obsolete each customer implementation; coding agents work well because of technical users and clean access controls; and the gap in other knowledge work is a data and access problem. Read the full conversation via the original episode.

Key Claims

• Token costs are breaking the subscription model and escaping IT budgets. A single agent run can cost $1,000, far above the $20-per-user-per-month ceiling that worked for chatbots. Labs have pricing power because a 10-year hardware cycle compressed into 18 months has created a capacity squeeze; frontier token prices are rising, not falling. The result: AI spending must flow out of the capped 3–7% IT budget and into line-of-business allocations (marketing, sales, operations), creating new friction between finance, IT, and business owners over compute spend.

• Enterprises are at "day one" of agents. Most large companies have only just finished rolling out the chat paradigm for knowledge work, and the capability has already moved on to agents that take actions and produce work, not just answer questions. The agent conversation is now the primary topic in every CIO engagement Levie has, and demand is pulling from the business side, not just IT.

• A "capability overhang" is paradoxically slowing diffusion. When Levie asks a hypothetical where the labs freeze model progress today, he argues enterprises could complete the change management in two to three years. Instead, breakthroughs keep arriving faster than customers can standardise an architecture, and each new model often obsoletes the implementation customers just shipped. The result is more reference architectures, longer sales cycles, and CIOs signing one-year deals with the labs rather than locking in.

• Coding works; the rest of knowledge work does not. The gap is structural. Coding agents have highly technical users who fix mid-run failures and absorb best practices fast, verifiable outputs (code runs or it doesn't), a single source of context (the codebase), clean access controls, and purely digital work. Knowledge work has context strewn across 20 systems (many not digital), scattered access permissions, and no single truth. The barrier is not capability but infrastructure.

• The "Bob and Sally problem" is the access-control blocker for agents in the rest of the business. Bob has too much access, Sally has too little, and the agent either bounces off an entitlement wall or, worse, answers questions using data it should not have seen. Until companies clean up identity, permissions, and data scoping for every system the agent touches, agents cannot be deployed safely outside narrow use cases.

• Data is the bottleneck, and it always was. Levie calls out the 20-year-old semantic layer problem: contracts stored in five different places, roadmaps across 30 locations, no consistent definitions of metrics like net retention or FX-adjusted growth. When only the data science team needed to answer those questions, humans could compensate. When every employee gets an MCP connection to the data warehouse, bad definitions become a company-wide problem. The same pattern applies to unstructured content, which is the wedge for Box's own headless play.

• The internal FDE is now the highest-demand hire in tech. A hyper-technical person embedded in a business function maps workflows, wires agents, and manages the data, permissions, and skills the agent needs. This is permanent work, not a one-time setup: each model upgrade creates fresh work to capture gains or redeploy scaffolding. Most enterprises lack this talent and must hire or retrain.

• Headless software is inevitable, but it does not kill the GUI. Levie expects every enterprise software vendor to operate a hybrid model three years from now: a seat business model for the human end-user interface, and a consumption business model for the agent caller. By volume, headless queries will dwarf human interface interactions, but the GUI remains useful for complex document work, data rooms, and cases where a person wants to be hands-on. For Box specifically, going headless is mostly a matter of adding agent-aware features (account provisioning, search that returns context for the agent) on top of an API-first stack that has existed for years.

• The AI job-doomers are wrong, for three structural reasons. First, knowledge work has an irreducible last-mile human loop: a lawyer must attest that an agent-drafted contract is valid. Second, Jevons paradox: when one engineer becomes five-times more productive, firms take on five times the projects, not one-fifth the staff. Third, SMEs that could not afford a designer or marketer before can now hire the first one, because the agent handles bulk work and the human supervises. Box is hiring across all its traditional functions.

• Startups can still win where scale becomes dysfunctional for labs: the bridge layer. Labs cannot build change management, vertical data integration, and ongoing support for every industry. The dynamic mirrors hyperscalers deciding where to compete versus where to be pull-through. Some use cases will be absorbed by labs; others will need vertical players because account control cuts both ways: labs want no wrapper-swapping; wrappers don't want to be undermined by labs' own applied products.

Quotes

• "The problem is the breakthroughs keep happening faster than the customer can implement any kind of standard architecture, and those breakthroughs often make obsolete the last thing you implemented."

• "You couldn't do totally dramatically crazy stuff with your data. Obviously it has to be implemented by somebody hyper technical."

• "Most agentic challenges are inversions of a data challenge. The agent can't get access to the right information to do the work."

• "This isn't a one-time thing. Once the model changes, there's another set of work to be done. Did you get the gains of that model improvement, or did you have to leave behind scaffolding you built for the prior model?"

• "I remain very confident in the need for a bridge layer from AI capability to the end-user workflow. Unless the labs build hundreds or thousands of people for every vertical and line of business, there's going to be opportunity in that area."