Box CEO Aaron Levie: The adoption of AI agents is currently polarized, with programming and programming-adjacent tasks already deployed
Box CEO Aaron Levie stated that the adoption of AI agents is currently polarized. Programming and programming-adjacent tasks have been deployed, while other knowledge work is still early. The post did not provide customer numbers, revenue, or launch dates.
He further divided programming into a continuous spectrum. A small portion of developers may have enabled backend agents to work on projects in parallel, while the majority still operate one agent per person. He did not specify the proportion of this small group, only stating it is likely very small.
The reason other knowledge work has not progressed is that processes have not been redesigned for agents. This is not the same as deploying a chat system to improve retrieval efficiency. Redesigning includes workflows, data access, accountability and responsibility after agent participation in decision-making, governance and compliance, as well as security upgrades.
His conclusion is that, aside from a few categories of work, the external forms are still very early. Therefore, relative to the adoption already seen, a 100-fold increase can still be expected. The 100-fold figure is his assessment of future potential, not achieved growth.
Box sells enterprise content management and workflow on documents. This statement did not announce new agent products nor did it mention the company's own adoption rate. It shifts the bottleneck from model capabilities to the processes and accountability systems that enterprises need to change.
In market mechanics, buyers are enterprises and developers still using one-on-one chat, while sellers are agent runtime and content platforms. The event is driven by the emergence of parallel backend agents on the programming side, not a full industry switch. Budgets are first flowing to agents that can write code, not yet to departments that need to rebuild approval, compliance, and security. Beneficiaries are programming tools, while enterprise software that treats chat boxes as agents is under pressure.
Source: Public information
ABAB AI Insight
Levie founded Box in 2005, selling enterprises the ability to store files in the cloud, then apply permissions and processes. The company's growth post-IPO has repeatedly stalled in large enterprise procurement cycles, not in consumer viral spread. This time, by breaking down agents, he essentially admits that the content layer he operates in does not yet have the ready-made workflows of the programming layer.
There is no new financing or acquisitions. Resources are directed towards redesigning, rather than launching another chat entry point. Data needs to be reconnected, decision-making responsibilities need to be redistributed, and compliance and security need to change paradigms. The motivation is that if Box's files are only used for retrieval, agents will remain focused on research efficiency. Files need to become operational objects for agents, and permissions and audits must be modified first.
Similar forks appeared in the early days of cloud computing. Engineers first moved testing environments to Amazon Web Services, while finance, legal, and HR systems migrated years later. Programming agents correspond to that batch of testing environments. One-on-one chat corresponds to fitting old software into browsers. The industry position is expanding on the programming side, while knowledge work is still in the transformation phase.
Structurally, this is a case where technology substitution has not yet crossed responsibility boundaries. Code can be rolled back, but payments, contracts, and HR decisions cannot. The reason is that once agents execute, responsibility shifts from the operator to the deployer. Without new processes for accountability, compliance, and security, enterprises will not integrate parallel agents into core systems. Thus, the 100-fold potential depends on the speed of reconstruction, not on the release of another model.
ABAB News · Cognitive Laws
- Use agents for work that can be rolled back first, and for work that cannot be rolled back later.
- Chat is fitting old processes; agents need to dismantle old processes.
- The 100-fold potential lies in process reconstruction, not in model releases.