Aaron Levie on AI Agents, Jobs and the Future of Enterprise Software

Matthew Berman visits Box to talk with CEO Aaron Levie about AI agents and their implications for work. Levie says agents have become a major focus of his attention.

Berman asks what Box means by becoming AI-first. Levie says the goal is to consider AI across business tasks to expand the company’s output, rather than simply reducing headcount. Examples range from writing code and running more tailored marketing campaigns to customer support and sales.

Levie says Box wants to compress projects that would have taken weeks into days. An executive group identifies strategic opportunities, while teams use Box AI and Box Hubs to ask questions across internal sales and marketing knowledge.

Experimentation is decentralized within approved platforms and access controls. Levie says the company seeks systems it trusts, with protections against training on customer data and access restricted to information employees are authorized to see. Weekly all-hands demonstrations let teams share their own AI workflows.

Levie acknowledges that Box has an advantage in adoption: its employees work at a Silicon Valley software company already engaged with AI. Other organizations may first need basic training about the technology and its uses.

Employees may resist if they think AI adoption means their jobs will disappear. Levie says Box has tried to be explicit that its goal is greater output. One exception he highlights is deflecting support cases, potentially redeploying people or money toward proactive customer success.

Levie suggests rewarding teams that use AI effectively with more budget or headcount, rather than allocating new staff only where AI cannot be used. He hopes a cycle of adoption, productivity and reinvestment can make the rollout feel safer to employees.

Berman asks why companies would stop at automating existing jobs instead of pursuing opportunities that were previously uneconomic. Levie argues a competitor that uses AI to serve more customers could outgrow one that focuses narrowly on cutting costs.

Levie predicts that some small businesses may eventually employ more people because agents let them do work they could not previously afford. A small team might use AI to improve products or marketing, then hire sales or adjacent staff as it grows. He presents this as a hunch, not a measured employment outcome.

Berman raises Dario Amodei’s warning about white-collar job losses and Andy Jassy’s comments about corporate headcount. Levie distinguishes technical capability from economic outcomes: he thinks agents might be able to perform many tasks in a hypothetical vacuum, but expects adoption and employment effects to unfold differently.

Levie argues that a last mile remains: people must direct agents, review their work and integrate outputs. He also expects interpersonal activities such as negotiating and customer relationships to remain important. Those are his forecasts about work, not guarantees that particular jobs will be unaffected.

Levie invokes the lump-of-labor fallacy and Jevons paradox. If AI makes legal, healthcare or marketing services cheaper and more productive, demand might expand enough to create more work. Whether demand actually exceeds the efficiency gain is the question, and he believes many categories have room to grow.

Berman notes the objection that demand may not scale enough to offset productivity gains. Levie concedes that the outcome is unknowable on both sides, but expects substantial unmet demand. He offers long waits for healthcare appointments as one example and suggests administrative automation could free clinicians’ time.

Levie describes a friend’s online balloon business as another hypothetical example: automate marketing or inventory work, reach more customers and then hire for supply chains or product development. The actual effect would depend on demand, costs and the business’s ability to expand.

Levie applies the logic to Box sales: if AI makes a sales representative more productive, a company can keep the gain or invest more in sales because the return becomes attractive. He says Box would choose reinvestment when the market opportunity justifies it.

Berman asks whether people will manage agents or review their output. Levie says enterprise software traditionally helped workers produce at their own speed, whereas agents may perform tasks in parallel and shift people toward planning, delegation, review and integration.

Levie describes an inversion of the old autocomplete model: instead of software merely accelerating each keystroke, people may spend their time making agents productive. They must define the broader goal and choose tasks that are specific enough to succeed but large enough to create a productivity gain.

Levie recounts a Box engineer using Cursor to generate code and then review it, finding errors regularly. The engineer estimates 50% more output in that example. Levie treats the anecdote as a sign of changing work, not a controlled productivity study.

Levie points to Codex as a task queue: ask an agent to build something, let it work, then inspect the changes and incorporate useful pieces. He expects similar systems to appear across other work. Asked about Box’s own deployment, he says engineers are using Cursor and he does not know the official Codex status.

Berman asks whether AI productivity and capital requirements could concentrate power among a few large firms. Levie acknowledges the risk but sees continuity with existing trends. At the same time, he argues that small AI startups can compete by finding the right product experience with fewer employees.

Levie says startups must choose their markets carefully: competing directly with established general-purpose assistants may be difficult, while other categories are newly open. He likens the moment to earlier waves that created consumer services in transportation, food and travel.

Levie expects agents to create markets in professional services, including legal work, healthcare, tutoring and research, where much work was not previously digitized in the same way. He argues that the absence of an established software incumbent in some niches gives startups an opening, even though established vendors can also enter.

Berman agrees that agent scaffolding is a promising area. Levie notes his own bias as an enterprise software executive and says he sees substantial opportunity in building on top of a small number of powerful models.

Berman asks how personal agents with long memories might interact with workplace agents and company intellectual property. Levie sees a difficult privacy and governance trade-off: personal health information, for example, should not unexpectedly influence a corporate system.

Levie also notes that personal conversations may be hypothetical or temporary, so an assistant must decide what is useful to remember. He worries about portability if years of context are locked into one provider, and suggests a standardized personal data card that users could carry between systems.

Berman asks whether an employee could take some agent memory to a new job while company information remains with the employer. The pair brainstorm an agent résumé describing personal working preferences. Levie stresses this is an idea, not a Box product announcement, and says usability would be as important as a portable format.

Berman asks who is responsible if an agent makes a critical mistake. Levie says that for the foreseeable future people remain responsible for reviewing code, legal documents and customer messages. He expects a transition period with mistakes and training while systems are not reliable enough for full autonomy.

Levie imagines a later stage in which a provider could bear responsibility for a sufficiently autonomous service, especially when customers cannot inspect its operation. He presents this as a future governance and liability question, not a statement of existing law.

Berman asks whether caution and limited guardrails will slow innovation. Levie argues that organizational change is often much slower than technology development: companies may need years to adapt workflows, permissions and employee practices. He uses the continuing migration to cloud services as an analogy.

Berman asks whether AI is nevertheless moving faster than cloud. Levie agrees that the technology is advancing much faster, but maintains that adoption still requires people to discover improvements and incorporate them into countless workflows. His proposed timeline is an analogy and prediction, not a measured forecast.

Berman asks who will own agents: SaaS companies, model providers or frameworks. Levie expects all of them to play a role, citing the variety that emerged in cloud and enterprise software. He favors interoperability standards such as agent-to-agent protocols and MCP so systems can hand work to one another.

Berman raises Satya Nadella’s suggestion that applications might collapse into agents. Levie finds the idea useful as a provocation: a conversational layer could abstract much of the interface and change where value accrues. But he argues users will still want dashboards, buttons and tables for recurring tasks.

Levie believes SaaS providers have an advantage deploying agents in the domains they understand: service management, sales or documents. Existing APIs can make those systems accessible to agents, while a higher-level assistant might coordinate work among them. As Box CEO, he has a direct business interest in this view.

Berman asks whether agents could generate custom interfaces on demand. Levie thinks that is possible and is enthusiastic about AI-assisted coding for IT teams, but doubts most finance employees want to own the risk of custom code for critical financial processes. Many will prefer an accountable software vendor, he predicts.

Berman asks whether browser use or APIs will dominate agent interaction with the web. Levie expects both: agents may spur more APIs, but browser use can handle long-tail tasks and combinations that APIs do not cleanly cover. He imagines agents observing a video feed as one possible application.

Berman worries that AI-generated content could degrade the signal-to-noise ratio online. Levie shares concern about synthetic content and expects verification of human-created material to matter. He views agent intermediaries as a broader consequence of automation, not solely a response to low-quality content.

Levie imagines agents with budgets paying data providers for information as more sources become restricted. Berman adds a concern that an agent between a person and the web could amplify the echo-chamber effects seen in social media. These are possible futures, not demonstrated outcomes.

Berman asks where companies overestimate and underestimate AI. Levie says CIOs he speaks with more often ask when particular use cases will be ready than assume everything already works. He reports that many examples focus on new capabilities, such as analyzing large contract collections or customer conversations, rather than replacing an existing worker.

Levie thinks people underestimate the change from using software directly to delegating work and reviewing a draft. He advises treating an imperfect output as a starting point that can save work, provided the human inspects and edits it. The value and error rate depend on the task.

As an example, Levie says AI drafted the skeleton of a survey for IT leaders, saving him an estimated three to five hours. He still made substantial edits, but the draft helped a project proceed that he might otherwise have deferred.

Berman asks whether AI writes Levie’s frequent social posts. Levie says he writes his own thoughts, aside from occasional spelling help, because writing forces him to understand a fast-changing field. Conversations with engineers and customers often trigger the ideas, and he says he does not use AI to brainstorm those posts.