The Inflection Point Aaron Levie Named Specifically
Aaron Levie , CEO of Box, one of the earliest enterprise software companies to bet heavily on AI , put a specific date on when enterprise AI agents became real: November.
Not a specific product launch. A threshold. The models crossed a capability level in reasoning and intelligence where agents could reliably complete multi-step enterprise tasks without constant human correction. Before that threshold, agents were impressive in demos and unreliable in production. After it, the calculus changed.
The view from OpenAI's enterprise team: the shift was the result of two things moving simultaneously , the underlying model intelligence and the tooling that lets agents access enterprise data. When both moved at once, enterprise deployments that had been stalled in pilot phase suddenly started working.
The Two Things Enterprise Agents Actually Need
Capability one: model intelligence. This is advancing rapidly and is largely outside the enterprise's control. You pick the model and benefit from improvements.
Capability two: data access. This is entirely within the enterprise's control, and it is the one that most companies are behind on.
An agent without access to your data is a general-purpose tool. An agent with access to your CRM, your document management system, your historical tickets, your contracts, and your operational data is something qualitatively different. It knows your business. It can answer questions a generic AI cannot answer. It can take actions that only make sense in your context.
The connectors that enable this , MCP servers, internal APIs, content management integrations , are the actual infrastructure investment that determines whether enterprise AI works. The model is a commodity. The data plumbing is the moat.
Why Document Organization Is the AI Readiness Work
Box's position in this conversation is not accidental. Content management , structured storage of documents, contracts, presentations, and unstructured knowledge , is the primary form of enterprise data that agents need to be useful in knowledge work.
The argument: companies that have already organized their content well will get dramatically more value from AI agents than companies that have not. An agent connected to ten years of well-organized contracts, policies, and institutional knowledge is not the same as an agent connected to a chaotic shared drive.
The implication for most enterprises: the AI readiness work is not primarily about model selection or infrastructure decisions. It is about data organization. The companies doing that work now are building the foundation for agents that will be substantially more capable than competitors' , because the agents will have better information to work with.
What Enterprise Software Looks Like Going Forward
The working hypothesis from both Box and OpenAI's enterprise team: the interface layer of software is going to change more than most people expect, and faster.
Today's enterprise software is built around humans navigating interfaces , dashboards, forms, search boxes, menus. AI agents do not need those interfaces. They need APIs, data access, and clear action surfaces. The software that survives will be the software that is genuinely useful to agents as well as humans.
The software that exists purely as a wrapper around a UI , with no real API, no clean data model, no designed integration surface , is in a more precarious position than most of its vendors would like to discuss publicly. The enterprise buyers who are thinking about this now are the ones who will have more use in vendor negotiations over the next three years.