Aaron Levie on Enterprise AI, Software Value and the Future of Work

Box founder and CEO Aaron Levie joins Harry Stebbings to discuss enterprise AI, employment, software value and the pace of adoption. His claims about Box and its products come from a company executive; his market and employment forecasts are opinions, not established outcomes.

Levie discusses Jensen Huang’s interview about AI and China, saying he agrees with much of Huang’s framing. He sees a commercial and economic competition with safety considerations, rather than believing a lead of a month or two necessarily determines the entire outcome. This is his geopolitical judgment, not a verified assessment of national intentions.

Levie argues that software security upgrades take time and that defense and attack will continue to leapfrog. He does not see a single moment of model access as automatically securing every system. He also says discouraging people from studying engineering or healthcare on the assumption that AI will eliminate those roles could be harmful.

In Levie’s view, current AI often changes where a person reviews work rather than removing humans entirely. He argues that the relevant unit of review may become larger as systems improve. This describes his current assessment of workflows, not a guarantee that all jobs or tasks will remain unchanged.

Asked whether Box will have more engineers in five years, Levie says it will and predicts engineering demand will broaden beyond technology companies. He cites industrial, banking and pharmaceutical organizations that he says need more automation. His projection about Box and the wider economy is not evidence of future hiring levels.

Levie suggests that easier creation of legal material could increase demand for lawyers to review and use it in regulated processes. He predicts there will be more lawyers in five years. Stebbings challenges the view by pointing to pressure on junior roles; the conversation does not establish the eventual number or distribution of jobs.

Levie acknowledges a separate challenge: if AI handles traditional entry-level tasks, firms must find new ways to train and mentor the next generation. He describes a customer working on automated patient referrals, but notes that appointment capacity and clinician availability remain constraints. The example is an anecdote, not a measured system-wide result.

Levie proposes an “agent operator” role combining technical understanding with business-process knowledge. He imagines such people connecting systems and helping marketing, legal, operations or research teams use agents. He is uncertain whether this remains a distinct job; his estimate of hundreds of thousands to a million roles is a speculative forecast.

He contrasts a startup’s ability to design fresh workflows with a large regulated organization’s fragmented data, existing processes and compliance obligations. In his view, using agents effectively requires redesigning parts of the workflow and deciding where human review belongs. His organization, Box, sells enterprise content and workflow tools.

Levie says a new model can change the behavior of an agent workflow, requiring ongoing testing and adjustment. An operator would need to understand systems, prompts and the business process rather than simply install a model once. This is a proposed implementation role, not evidence that every model update breaks every workflow.

Stebbings asks whether software products become low-value databases when agents work across them. Levie agrees that some interface-heavy categories may face pressure but distinguishes a database from software containing useful APIs, permissions and business logic. He sees the value of an interface shifting rather than all application value disappearing.

Levie uses enterprise resource planning as an example of software that can embed accounting and supply-chain rules above the database. An agent may interact through APIs rather than a person clicking each button, but the underlying logic and controls can still matter. This is his framework for comparing products, not a valuation of a particular vendor.

He expects agents to work across several systems in the background while people review some outputs. Products may need usable APIs and pricing that reflects automated use. He says the impact will differ by category and by how much human-agent collaboration a workflow requires.

Levie predicts that agents will create and consume more unstructured data, including contracts, reports and marketing material. He argues that companies will still need ways to manage, secure and govern that material. This thesis is directly relevant to Box’s commercial interests as a content platform.

Asked whether that increases Box’s value, Levie says the company already has extensive API usage and can support workflows where users do not see Box directly. He expects more agent activity to create opportunities but acknowledges the revenue per agent interaction could differ from past application usage. He supplies no API counts or financial forecast.

On security, Levie worries that AI-generated code can increase the volume of changes that require review, and that attackers can also use automation to look for vulnerabilities. He says defensive agents may help review code. These are risk scenarios, not evidence that automated review is sufficient or that a specific attack has occurred.

Levie expects organizations to spend more compute per employee but declines to provide a five-year Box financial projection. He treats the direction as likely while leaving the magnitude unknown. Stebbings notes that technology suppliers have their own interests in encouraging compute spending.

On token budgets, Levie says spending should relate to the value generated by a use case. He describes examples such as teams pitching for a budget or allocating different model tiers to different users. He explicitly makes up the illustrative percentages; they are not Box policy or measured results.

He notes that large enterprises operate within annual budgets, earnings commitments and approval cycles. Unlimited token use across an organization can conflict with those constraints. In his view, companies must learn which teams and workflows justify the spending rather than simply copy a startup’s usage pattern.

Levie predicts some AI spending will shift from a central IT budget into operating budgets for business functions. He suggests a line-of-business leader may compare an agent with another campaign or process. His estimate that technology spending could double is a forecast, not an observed or guaranteed outcome.

Stebbings asks whether current enterprise interest in AI could be a temporary period of experimentation. Levie accepts that risk but expects adoption to take longer and reach further, drawing an analogy to cloud computing. He is forecasting a diffusion pattern, not demonstrating that demand will persist.

Levie says regulated companies cannot simply deploy agents into consequential workflows without addressing review, compliance and security. He sees product-led adoption and hands-on enterprise implementation as compatible, and expects services firms to have substantial work helping organizations integrate systems and change processes.

He gives an example of contracts spread across multiple systems, some old or difficult to connect. An agent can retrieve an outdated or wrong document if information is not organized and the task lacks context. As Box’s CEO, he has a commercial interest in the content-management problem he describes.

Stebbings raises accountability when an automated workflow causes harm. Levie argues organizations still need owners who review outputs and accept responsibility for decisions, security and contractual obligations. This is his business view; liability depends on the facts, agreements and applicable law.

Asked about open models from Chinese developers, Levie says such models are used in Silicon Valley but provides no data to establish the scale or funding of that use. He mentions hypothetical security risks while saying he does not view the issue solely through that lens. The interview does not establish a backdoor or the motives of any developer.

Stebbings relays a claim that public companies have not built strong agent products. Levie does not fully accept it and says Box’s agent is strong for content work. His comparative claim is self-interested and the interview provides no independent benchmark against competing products.

Levie says keeping up with AI requires leaders to follow technical practitioners and respond to changes on a much faster cycle than traditional news summaries. He describes simultaneously adapting product, strategy and partnerships while helping customers manage the transition. He says this has made his job harder.

On monetization, Levie describes a newer Box plan that includes workflow, application and agent capabilities. He says the offering contributed to an inflection in revenue growth. This is the CEO’s characterization of company results; the interview does not isolate causation or supply the underlying financial data.

Levie believes software stocks are being treated too uniformly and predicts investors will distinguish companies as agent adoption affects categories differently. He says some valuations may be too low and others may previously have been too high. These are his market opinions, not current prices or investment advice.

Stebbings questions whether established CEOs are ready for AI. Levie says he knows public-company leaders who actively experiment with coding tools, while acknowledging that technology shifts produce winners and losers. His examples do not establish how prepared any particular leadership team is.

Levie expresses a favorable opinion of Atlassian and believes tools supporting engineering may benefit if engineering activity expands. He also says products must evolve and that multiple vendors can serve a large market. His views on valuation and future demand are speculative and not a recommendation.

He says software for agents needs effective APIs, pricing and surrounding controls. In a Box sales example, he describes retaining and governing regulated documents while agents interact with them. This is an illustration of Box’s positioning, not independent proof that a particular workflow complies with a regulation.

Levie says he has become more convinced that software should work without a visible user interface because agents have improved at tool calls and searching systems. He still recognizes the importance of user interaction and human review elsewhere in the conversation. His claim about past and present agent accuracy is experiential.

Asked about acquisitions he missed, Levie says he has no major regret and recalls being glad that some deals he wanted were not completed. He declines to name them. The answer is a personal reflection, not information about any specific transaction.

On whether OpenAI or Anthropic will win enterprise AI, Levie says he cannot predict market shares. He compares the question with cloud computing, where several providers grew as the overall market expanded. He expects enterprises to use multiple vendors to reduce dependence, but this is a forecast.

Levie cautions against assuming the speed of progress in AI coding will immediately carry over to every form of knowledge work. Different tasks have different data, review and organizational constraints. He offers that distinction as a hypothesis about adoption, not a measured timeline.

If he were a venture investor, Levie says he would still consider frontier-model rounds because he thinks the market could grow further. He also recognizes that investors have repeatedly misjudged how far valuations can rise. This is a hypothetical speculative view, not an endorsement of a specific valuation or a recommendation.

Levie interprets some market movement as investors shifting attention toward chips and data-center infrastructure rather than traditional software. He acknowledges unusual trading dynamics and uncertainty. His account is a market interpretation, not evidence of a particular security’s future return.

He sees potential in companies that bring AI into real enterprise workflows, including evaluation and observability tools. He names Braintrust as an example and explicitly says he is not an investor. His enthusiasm is a prediction about a category, not an independent assessment of any company’s performance.

Stebbings asks how Levie balances a demanding CEO role with marriage. Levie says he is fortunate to have a supportive family and that they try to make time for enjoyable activities despite constraints. He avoids presenting himself as a model husband, instead describing a long shared experience of work and family.