Aaron Levie on AI Agents, SaaS Pricing and the Race Between Startups and Incumbents

Harry Stebbings asks Box co-founder and CEO Aaron Levie what the transition to cloud can teach founders about AI. Levie sees a temporary opening, comparable to earlier shifts to PCs, the web, mobile and cloud, in which new large companies can emerge. He does not know whether the window will last two years or five, or whether it has already closed.

This time, Levie expects incumbents to compete as vigorously as startups because they already possess data and workflows. Startups can still find use cases or delivery models incumbents overlook. His message to companies pursuing the opportunity is to focus intensely on survival and execution.

Levie expects far more large application companies than independent foundation-model companies. He argues that the cost of competing with large training runs and well-funded open models makes horizontal models difficult to differentiate. Specialized domains may offer openings, but his estimate of only a few independent horizontal players at scale is a prediction, not an established outcome.

At Box, Levie says, the team is building an architecture that connects customer content securely to different models. Customers may choose one model for legal questions and another for document metadata. He doubts completely automatic substitution will always work because models vary in response style and behavior; his account reflects Box’s own product strategy.

Stebbings reports hearing enterprise fears about putting sensitive information in the cloud. Levie says he sees some former cloud holdouts moving in the opposite direction because they want to use AI on their data. His view is that data must be in cloud-ready form to unlock the full benefit, a position aligned with Box’s cloud business.

Box benchmarks models on business documents, Levie says. He compares an approximately 4,000-token GPT-3.5 context window with Google’s announcement of a two-million-token Gemini version, calling it roughly a 500-fold increase over 18 months. He distinguishes context-window growth from Moore’s law, which concerns chip density or performance, and sees no sign of an innovation slowdown at the time of the interview.

Levie sees agents as a possible next step beyond chat. A chat interface can make information easier to access, but often changes the interface more than the underlying job. An agent, in his formulation, would take an instruction and perform the task rather than simply return information for a person to act on.

Stebbings asks how that differs from robotic process automation. Levie says RPA has long promised automation but can be brittle when screens or procedures vary. More capable models could handle a wider range of business tasks; examples he imagines include outbound sales, product testing and customer support. These are proposed applications, not a claim that all are reliably autonomous today.

Levie expects broad organizational structures to endure, with AI labor added to particular functions. Support agents might handle initial inquiries and escalate difficult cases to people, while AI-generated leads could flow to human sales representatives. Stebbings presses him on whether fewer support or outbound staff would nevertheless change headcount and the organization.

Levie says that, in Box’s experience, productivity gains have often led the company to reinvest in an area because more work remains to be done. His examples include engineers building more software and support staff shifting toward proactive customer success. His hypothetical improvements in sales or ticket handling are illustrations, not reported measured outcomes.

For a new company, he predicts AI could generate leads earlier and help it grow enough to hire more people in other functions. He accepts that very small, highly valuable companies may appear, but expects them to be relatively unusual. This is his forecast about employment and scaling, not a guarantee that every organization will reinvest productivity gains.

On enterprise spending, Levie says both experiments and production applications exist, but he cannot reliably quantify their shares. After roughly a dozen AI events across the United States, he reports seeing many companies with multiple experiments and some production uses. Spending may appear inside existing software budgets or as a separate consumption charge.

Stebbings asks what happens to the seat-based SaaS business model if companies sell completed work. Levie says the answer is still unsettled: vendors are testing consumption units such as support tickets, leads or emails, as well as value-based prices. Bespoke value pricing may be difficult to scale, so he expects some common unit of consumption to emerge.

Looking five years ahead, Levie imagines AI services performing functional work such as reviewing contracts, processing invoices and generating leads, with category winners in several jobs. He also anticipates management tools for digital workers: coordinating providers, setting guardrails and organizing their work alongside human teams. These are scenarios, not current capabilities he has verified.

Incumbents bring customers, workflows and data to the competition, and in some categories adding AI to an existing platform may work well. Startups can exploit blind spots or business-model conflicts. Levie offers the hypothetical example of a support software vendor reliant on seats failing to adapt to an agent-based product; he stresses that the disruption will not necessarily play out exactly that way.

Stebbings raises the fear that OpenAI will absorb startups’ products. Levie welcomes clarity about the platform’s ambitions and advises founders to avoid simple offerings a horizontal chat interface or the next model training run could subsume. He sees more room in complete workflows that help people run business processes.

For language learning, Levie thinks a general model might be capable of tutoring, while a specialized product could still deliver a structured journey over years. His advice to incumbents is to adopt superior technology quickly even if it comes from a potential competitor. He cautions that he cannot give Duolingo specific advice on a podcast; the example is an illustration.

Stebbings asks how a larger company can move with startup urgency. Levie says his earlier call for intense work was directed at Box employees. With about 700 people, he says, the company can still discuss strategy together in an all-hands meeting; his concern is that a company unable to cross the AI transition could lose its business.

Scale makes coordination more costly. A small team can test an idea and pivot quickly with everyone aware it was a hypothesis. At a large company, other teams wait for a dependable direction, so leaders must decide when to expose a developing product: too early creates confusion, too late wastes time. Levie speculates that some early Gemini problems reflected such coordination challenges.

Levie views recent Google announcements as evidence that AI has become a top company priority across search, cloud and workspace. He is optimistic about Google in an AI era, while saying he does not buy or sell stocks. His assessment is an opinion at the time of the conversation rather than investment advice.

On regulation, Levie says he has become less worried that proposals will halt progress. He opposed calls to pause advanced AI development for six months, arguing that deep philosophical disagreements would not be resolved in that time. He supports discussion of more targeted questions such as copyright, training data and intellectual-property protection.

Levie acknowledges national-security concerns about future highly capable AI, even though he is less focused on that risk than some people he respects. He welcomes an open discussion of how to guard against those risks.

Asked about RPA providers, Levie thinks incumbents are well positioned because they already sell workflow automation and often have consumption-based pricing. AI could also make automation easier and expand the market for new entrants. His suggestion that the market might ultimately grow 100-fold is speculative and depends on substantial improvements in cost and ease of use.

Box is pursuing automation of work that was previously difficult to automate, Levie says. He imagines searching a company’s marketing assets for red dresses or asking which contracts with a particular clause come up for renewal next month. These examples express Box’s product vision and commercial interest; they are not independent evaluations of performance.

The longer obstacle to adoption, Levie argues, may be implementation and human change management rather than technical breakthroughs. He expects substantial spending on AI services over the next five years, alongside spending on chips and data centers. If AI becomes as significant as he expects, infrastructure and software spending could eventually exceed human implementation services once systems are running.

Levie is interested in whether AI lowers the barrier to starting businesses, as Shopify, AWS and Stripe did for online selling, computing and payments. He gives the example of a friend who sells balloons online using Shopify. In his scenario, a small company in a country without a mature sales workforce might use AI to help find customers or provide support.

Stebbings worries that unlimited AI-generated outreach and content could overwhelm recipients and diminish value. Levie expects a familiar pattern of competition: things that do not work die off and those that work succeed. His answer is an optimistic market hypothesis, not evidence that saturation will be harmless.

Asked what most concerns him, Levie points to an armed robot using multimodal AI as a dangerous use worth addressing. He believes existing legal frameworks can cover many harms, with new rules where necessary, and is less concerned about runaway AI self-replication.

In the rapid-fire round, Levie identifies delegation as a leadership challenge. He likes involvement in many parts of the business but has learned to select the critical areas, such as AI, end-to-end experience and product strategy, while giving strong leaders room to execute elsewhere.

Levie credits Mark Zuckerberg’s position in AI to access to computing, engineers, users and data, as well as a desire to build an open platform. He would add NVIDIA CEO Jensen Huang to Box’s board if he could, because Huang’s position in the AI supply chain gives him unusual insight into future technology demand.

Levie is bullish on Apple. He argues that AI use on Apple devices does not necessarily threaten its existing business and could make the phone a powerful interface for tasks such as booking travel or arranging rides. He says the monetization is uncertain and that Apple could partner for underlying models; reported partnerships were rumors at the time of the interview.

Looking back at Box, Levie wishes he had focused on cash flow earlier. He says financial discipline helps a company control its destiny and makes constraints useful for better decisions. On balancing work and marriage, he mentions making time for dates and family alongside demanding schedules.

Box had passed $1 billion in revenue, and Levie says its next goal is $2 billion as quickly as possible. Stebbings challenges its roughly $3.89 billion market capitalization. Levie argues the valuation is too low, says the core business had slowed and describes AI, workflow automation and business processes as a developing next act; his valuation view is that of the company’s CEO.

For the next five years, Levie wants companies to treat their documents and other content as a form of digital memory that AI can help them use. He imagines better decisions, automated processes and faster employee onboarding. It is a product vision for Box, rather than a demonstrated outcome for every customer.