Box CEO: How AI Is Reshaping Software and the Future of Agents

Box CEO Levie says AI splits the software stack into agents and context, urging enterprises to build structural substrates for superintelligence.
Box CEO Aaron Levie explains how AI is reshaping software and enterprises. He argues the software stack is splitting into an agent layer and a real-world context layer, with systems of record evolving into systems of intelligence centered on data governance, access control, and metadata routing. Box launched Box Architect to let frontier models design workflows autonomously. Levie calls himself a Bitter Lesson disciple, advocating discarding features obsoleted by new models and building only the security and data substrate for superintelligence. He predicts explosive token growth and sees open weights and closed-source frontier as complementary, not zero-sum.
From Systems of Record to Systems of Intelligence: Re-layering the Software Stack
Two years ago, Satya Nadella proposed that software would split into two layers: agents and real-world context. Levie believes this prediction is becoming reality. He cites Jensen Huang's claim that the number of agents in the future could be 10 to 100 times the human population, reaching tens of billions. This means there must be systems to manage what data these agents can access and what processes they can operate.
You can't just dump this stuff into one giant information pool. It has to be carefully curated, with the right governance, sandboxing, and isolation.
Levie positions Box as a "modern system of record," but he prefers to call it a "system of intelligence." Such systems control the access layer, data governance, and metadata routing while retaining a user interface, because humans still need to input data and interact with the system. This rationale explains why Box remains relevant in the AI era—unstructured data will only grow as a proportion of enterprise data, and agents need a trusted way to access it.

System of Record is a classic concept in enterprise software, referring to the authoritative system that stores core business data (such as customers, orders, and financial records). Typical examples include Salesforce (CRM), SAP (ERP), and Box (content management). The key attributes of a system of record are trusted data, controlled permissions, and auditability. Levie's notion of a "System of Intelligence" emphasizes that in the agent era, storing data alone is no longer enough; the system must also understand the data, determine which data can be accessed by which agents, and dynamically route requests. In other words, a system of record focuses on "where the data is and whether it is correct," while a system of intelligence focuses on "how the data is understood and used securely." The latter becomes a more critical competitive moat when agents access enterprise information at scale.
Partner-First, but Also Growing Upward
Box's DNA is partner-first. Levie emphasizes that information needs to flow to where work happens—contracts should appear in Salesforce, supply chain documents in SAP or Oracle, and product marketing plans in ChatGPT or Claude. Therefore, openness and interoperability are principles Box will not abandon.
But Box is also extending upward. The newly released Box Architect is an autonomous system for designing, running, and ultimately optimizing agents, workflows, and applications built on Box. Enterprises can submit requirements, such as a contract management solution or an invoice processing solution, and Box Architect will invoke frontier models (such as Astra, Opus, or Fable) to design a complete solution and use Box's internal building blocks to implement it.
Levie does not see this as a violation of the partner strategy. He admits that many customers do not have the time or ability to assemble modular solutions themselves, and end-to-end solutions are a necessary complement. At the same time, Box remains neutral on whether users choose Box Architect or call Box headlessly within other technologies.
A Disciple of the Bitter Lesson: Ready to Throw Away Your Own Code
The most candid part of the interview concerns the frequency of "throwing things away." Levie admits that Box often has unreleased features that are simply scrapped because of new models.
We had features that we hadn't managed to release before a new model made them completely useless. This happens constantly, and it's accelerating.
He gives an internal example: about six months ago, an engineer demonstrated with the then-latest model how to design a complete Box application solution. The team's reaction was, "Oh my God, the agent did in 30 seconds what takes a customer two days to design and architect on a whiteboard." This directly inspired the idea for Box Architect—rather than automating today's processes as they are, let AI tell us what better processes should look like.
Levie calls himself a "deep believer in the Bitter Lesson." The internal principle at Box is: assume that future superintelligence will be smarter than everyone on every topic, so what needs to be built now is the substrate required to safely deploy AI in that world—data governance, security, high-speed high-throughput infrastructure components, and schemas that agents can easily invoke.

The Bitter Lesson was proposed by AI researcher Rich Sutton in 2019. Its core point is that approaches in AI that rely on human prior knowledge, hand-engineered features, and rules are, in the long run, always surpassed by "general methods that leverage massive computation." Levie's self-proclaimed deep belief extends this principle to product strategy: do not bet on any current process or model remaining effective long-term, because model capabilities will rapidly evolve and crush old assumptions. Box therefore chooses to frequently scrap unreleased features and invest only in structural capabilities needed regardless of how models evolve, such as data governance, security, throughput, and standardized schemas. This thinking explains why Levie believes "building the substrate for superintelligence" is more sustainable than "optimizing today's processes."
The Explosion of Token Consumption and the CFO's Fear
Levie predicts that token usage will grow by orders of magnitude. His logic is straightforward: imagine that in three years, every piece of information generated in the world will be viewed, reviewed, and processed by one, three, five, or even ten agents. In sales, every customer will have a continuously updated agent profile; in security, every anomalous login will trigger agent review; in engineering, agents will review code, test code, and simulate risks; in supply chains, every document will have key information extracted by agents and written back to the ERP.
We may currently have attached agents to only one percent of all interactions. Many of the things I do on my computer are not seen by agents. That won't be the case in three years.
But Levie does not subscribe to simple "token maximization" logic. He believes the correct application of tokens matters more; infinite tokens pointed in the wrong direction will also waste resources. However, he admits that how much compute to invest in problems worth solving may become one of the key factors defining competitive advantage in the future. This creates tension: capital becomes a dependency for competitive advantage, which may lead to concentration of power and may also stifle innovation.
When asked how CFOs react to hearing "token spending cannot be predicted in advance," Levie laughs and says he can hear a thousand CFOs trembling. But he believes that in three to five years, token spending will become a normal budget item like marketing or travel budgets. What Box can do is provide lower-cost solutions, such as a new context retrieval engine built with NVIDIA, so that agents only retrieve the data snippets they actually need, thereby reducing unnecessary token consumption.

Open Weights and Closed-Source Frontier: Not a Zero-Sum Game
Levie strongly supports open-weight models, seeing them as an "escape valve" to prevent AI labs from becoming economic gatekeepers. But he does not see open source and closed source as opposites. He uses the database market as an analogy: after 2005, most Silicon Valley startups used open-source databases, yet Oracle still grew into a giant during the same period.
You can have both. Open source lets users deploy and control things themselves, and platforms can do the work you don't want to invest in specifically.
He further offers a counterintuitive point: open weights lower the overall cost of AI, which may actually make enterprises more able to afford frontier closed-source models. Because enterprises can use cheap open-source models for high-volume, repetitive tasks, such as reviewing transaction-by-transaction fraud signals, and then use frontier closed-source models for orchestration and planning layers. Without open-source models, these high-volume tasks would be economically infeasible.

Levie also points out the difficulty of the open-weight business model in the United States: training costs are enormous, and if models are released for free, emerging cloud providers can directly offer inference services without bearing training costs. To sustainably subsidize the development of open-weight models, one usually needs to be an inference provider, or profit at a lower layer like NVIDIA. Meta, Google, and NVIDIA are positioned to both train and provide inference.
Regarding the gap between open-source models and frontier closed-source models, Levie believes that if open-weight models can stay three to six months behind, that is already valuable enough. Because tasks that frontier models could accomplish a few months ago often already cover most routine use cases in the world. Open-weight models will continuously eat away at token work that has already been proven.
Open-weight Models refer to models whose parameters are publicly downloadable, self-deployable, and fine-tunable, such as the Llama series and Mistral. They are not fully equivalent to "open source": training data, training code, and license terms may not be public, but open weights allow users to deploy privately, control data flows, and post-train for specific tasks. Levie sees them as an escape valve to prevent a few AI labs from becoming "economic gatekeepers." At the same time, open-weight models typically lag frontier closed-source models by several months, but they are sufficient to cover a large number of already-standardized tasks. Combined with closed-source frontier models responsible for complex orchestration and planning, enterprises can form a layered cost structure: cheap open-source models handle high-concurrency, repetitive work, while closed-source models are used for high-value judgment, thereby achieving frontier capabilities with controllable total costs.
The Future: Post-training and Adaptive Routing
Box's collaboration with NVIDIA focuses on the context retrieval engine, using NVIDIA's embedding models to improve retrieval relevance ranking, making answers returned to agents more token-efficient. This capability will be offered as part of the platform substrate, available transparently to customers or via API to retrieve document content by relevant snippets, with a plan to roll out gradually in the Enterprise edition.
As for whether Box will train its own models, Levie believes post-training is almost inevitable. Certain enterprise knowledge or high-frequency customized workflows are worth post-training to improve efficiency and accuracy. But he also mentions that Box is still reaping the baseline benefits of improvements in existing model capabilities, so it will not invest too early in building its own models. He is optimistic about several startups in the post-training ecosystem and envisions a future where enterprises can train their own enterprise data models on the Box platform and run them directly.
On model selection, Levie's ideal state is an adaptive routing layer: users submit queries, and the platform automatically selects the most appropriate model based on cost and accuracy, whether open source or closed source. This aligns with his identity as a "Bitter Lesson disciple"—Box should not bet on any single model always leading, but should build structural capabilities needed regardless of how models change.
Post-training refers to adapting a pretrained foundation model to specific tasks or enterprise knowledge through supervised fine-tuning, reinforcement learning from human feedback (RLHF), instruction tuning, or continued training on domain data. For enterprises, post-training can help models more accurately understand internal terminology, contract clauses, or high-frequency workflows without training a foundation model from scratch. Levie believes Box will eventually enter post-training because certain enterprise knowledge and customized workflows are difficult to achieve sufficient efficiency and accuracy with general models and prompts alone. At the same time, an Adaptive Routing Layer refers to a scheduling mechanism where the platform automatically dispatches requests to different models (open source or closed source) based on query type, cost budget, and accuracy requirements. It frees enterprises from locking into a single vendor and aligns with Box's strategy of not betting on any single model always leading.
One-line Summary
Box CEO Levie believes AI is splitting the software stack into an agent layer and a context layer, and enterprises should build structural substrates for superintelligence.
Paragraph Summary
Box CEO Aaron Levie systematically laid out his views on how AI is reshaping software and enterprises. He believes the software stack is splitting into an agent layer and a real-world context layer, and systems of record will evolve into "systems of intelligence," with data governance, access control, and metadata routing becoming core capabilities. While sticking to partner-first and open interoperability, Box launched Box Architect to let frontier models autonomously design enterprise workflows. Levie calls himself a "Bitter Lesson" disciple and advocates frequently discarding features obsoleted by new models, building only the security and data substrate for the superintelligence era. He predicts token consumption will grow by orders of magnitude, but correct allocation matters more than unlimited investment; open weights and closed-source frontier are not zero-sum—open weights lower total costs and actually make frontier models more affordable for orchestration layers.
Related articles

E-Readers Enter Their Best Era Yet: How to Choose Between Kindle, Kobo, and Chinese Android E-Ink Devices
The Vergecast guests compare the new Kindle, Kobo, XTE Ink mini e-ink devices, and Android readers like the Boox Palma—covering rising prices, color vs. black-and-white, the KDP Select ecosystem trap, and the page-turn buttons and stylus annotations readers really want.

A One-Person IT Security Team: How ESET Covers Protection All Day
Follow Alex, an admin handling IT and security alone, through a full day: threat briefing, third-party patch management, cloud sandbox analysis, AI governance, and MDR response—see how ESET enables a one-person team to cover all-day protection at a hundred-person company.

Building an AI Second Brain with n8n: Agent + Memory + Vector DB in Practice
An n8n tutorial breakdown: build a self-hosted second brain with AI Agent, Simple Memory, Qdrant vector DB, and Notion—Telegram multimodal routing, structured ingestion, Gmail scheduled channel, and configuration details.