Aaron Levie, CEO of Box, on Box AI, Enterprise Enthusiasm, and the Evolution of SaaS
Summary
Enterprise AI demand is running far ahead of cloud-era enthusiasm, but production remains in “very early innings.” Cloud required reluctant companies to surrender physical infrastructure and trust unfamiliar vendors; AI instead has executives proposing “almost as many use cases as possible,” sometimes more than practical. Levie sees a relatively narrow window in which large technology companies, enterprise-software vendors, and new agent startups can capture that demand.
AI changes IT from a software-enablement function into an operator of digital labor. Business units will ask IT not merely to deploy CRM or HR systems, but to provision agents that run sales campaigns, review contracts and invoices, or execute onboarding. Borrowing Jensen at NVIDIA’s framing, “the IT department becomes the HR department of AI,” requiring much deeper business knowledge and strategic authority.
Box’s RAG advantage came from architecture it built before ChatGPT, partly through luck. Box Hubs lets users curate authoritative documents without copying or changing permissions, reducing the risk that draft-heavy corporate repositories contaminate retrieval when users query a bounded hub. Because enterprises lack the public web’s PageRank-like authority signals, Levie argues this human curation makes Box’s RAG “at least a hundred times better” than indiscriminate search across all company data.
The strategic destination is a new “system of intelligence” that combines structured control with probabilistic judgment. Box’s current agents include model access, tools, skills, instructions, and enterprise data, though Levie concedes many would have been called assistants two years ago. The larger prize is agents that review content, make context-dependent decisions, route work, and coordinate with Salesforce, ServiceNow, Microsoft, or humans.
Being human-equivalent or 40% cheaper is insufficient to overcome enterprise adoption friction. A buyer still faces 17 competing projects, AI-council approval, testing that might take six months, privacy and security concerns, and workforce transition; consequential workflows may also demand “99.99999% reliability,” not 98%. Levie’s commercial threshold is an order-of-magnitude gain in cost, quality, or capability, especially for work that was never automated before.
Incumbents should retain their natural application domains, while startups win in cross-platform or previously unserved workflows. An “AI-first CRM” must assume Salesforce becomes AI-first too, just as Workday and ServiceNow will defend their domains; thin layers over an incumbent or OpenAI are therefore vulnerable. Independent compliance, agent red-teaming, cross-application workflows, and products requiring substantial non-AI interfaces remain credible startup territory.
AI may reopen SaaS pricing while improving productivity before the aggregate statistics clearly register it. Levie expects outcome pricing, compute consumption, and subscriptions all to coexist after two decades dominated by per-seat SaaS; Box itself is still testing. Internally, coding-tool gains range from 5–10% for some engineers to possibly 50% for a new hire, but Levie would take the over on economy-wide productivity forecasts if the clock starts in a few years.
Deep dive
1. Enterprise enthusiasm has inverted the cloud-adoption pattern
Levie would not pin down AGI because its definition remains amorphous and he is “downstream of whatever Ilya, Sam, or Greg are talking about.” Conditional on the current pace continuing, he points to reasoning gains in math, logic, and coding; continued benchmark improvement could provide building blocks for models that learn on their own, making “some form of whatever we would have previously defined AGI as” plausible within a few years.
Personal adoption offers one small demand signal: Levie added at least five AI applications to his home screen in six months, versus perhaps one meaningful addition every year or two over the preceding decade. He regularly talks to Gemini Voice and ChatGPT voice and video, uses Perplexity, plays with xAI and Grok, and built a prototype with Artifacts and Claude.
Cloud’s first years brought skepticism over moving infrastructure out of company data centers, trusting unfamiliar vendors such as Amazon, and replacing physical control with APIs, dashboards, and audit reports. At the same point in AI’s cycle, customers are inventing use cases “probably in many cases more than is actually practical.”
The caveat is deployment: excitement has not yet translated into much at-scale enterprise usage. Still, Levie sees a “relatively narrow window” in which large technology companies such as Microsoft, Oracle, and Google, enterprise-software vendors such as Box, Salesforce, and ServiceNow, and entirely new agent companies can address problems enterprises previously could not attack.
2. IT becomes the operating department for AI labor
Traditional IT selected, deployed, secured, and managed systems while business functions such as sales, finance, marketing, and HR remained responsible for execution. A CRM made salespeople productive; it did not itself own the sales campaign.
AI “flips that on its head.” A sales leader might ask IT to spin up agents for a campaign, while operations teams request agents to review invoices, process contracts, or manage client onboarding—the technology organization now helps perform the work, not merely support it.
Jensen at NVIDIA’s formulation captures the organizational shift: “the IT department becomes the HR department of AI.” IT must understand operating processes, model capabilities, agent vendors, and the wider ecosystem well enough to provision digital labor, making the function more strategic but forcing dramatic transformation.
Box, which Levie says serves about 15,000 customer companies and stores well over 100 billion files, is positioning enterprise content as that labor’s substrate. Box AI preserves permissions, privacy, security, and access controls while connecting models to documents, extracting structured metadata, and eventually enabling agents to operate on contracts, financial records, research, media, and product plans.
3. Authoritative curation—not embeddings alone—makes enterprise RAG work
Labenz’s pushback concerns the familiar “trough of disillusionment”: many RAG systems fail because vector search retrieves the wrong source before generation even begins. Similar embeddings cannot reliably distinguish an authoritative earnings report from “Earnings_Final,” “Earnings_Draft_1,” and “Earnings_Draft_1_Sally_Edits” variants accumulated over years.
Box happened to begin Hubs roughly a year before ChatGPT. Its many-to-many architecture lets one canonical document appear in 20 topic-specific hubs without moving it, duplicating it, or changing permissions; updating the source automatically updates every hub. Levie’s candid assessment: “In this case we got totally lucky.”
Public search benefits from PageRank-like evidence that one article or site is more authoritative than another; messy corporate content generally has no comparable signal. By choosing what belongs in a sales, product, regional, or HR hub, employees identify both the trusted corpus and the questions appropriate to it.
That bounded, curated retrieval layer is the claimed breakthrough. Rather than ask generic questions across 100 million heterogeneous files, a user queries authoritative sales material inside a sales hub; Levie estimates the resulting service is “at least a hundred times better” than broad-based RAG across everything.
4. Agents evolve from branded assistants into probabilistic workflows
Box deliberately uses a broad definition of agent to avoid confronting customers with “17 different versions of a thing.” An agent combines one or more models, platform tools, underlying skills or capabilities, system prompts, proprietary architecture, and permissioned access to enterprise data.
The first generation is modest: users can converse with one document, query many files, generate content, extract metadata, or create a sales agent with specialized language and instructions. Levie acknowledges that “many of these agents would be what we would have called assistants two years ago.”
The destination is closer to Labenz’s autonomy test: multi-step workflows with meaningful decision-making discretion. An agent reviews a contract, identifies risky clauses, determines what subsequent process to trigger, and routes the result to another agent or human; browser operation could bridge systems where clean APIs do not exist.
Levie maps the shift across three eras: systems of record deterministically changed database rows; systems of engagement made human collaboration fluid and less structured; systems of intelligence combine record-like control with engagement-like adaptability. Because most real work “does require judgment,” probabilistic workflows expand what enterprise software can digitize.
5. Reliability matters, but enterprise inertia demands a 10x result
Levie agrees that organizations underuse capabilities already available. Even while immersed in AI, he must remind himself that instead of asking somebody to handle a task, he should “go and try” creating it with AI; the awareness gap outside the industry is likely much larger.
He nevertheless rejects 98% reliability for consequential processes: nobody accepts arriving at an airport to discover that 2% of booked tickets do not exist. Enterprises may require “99.99999% reliability,” and even a 0.01% error rate is disqualifying when applied to a billion financial transactions.
Cost is another constraint. Customers may want 10,000 agents aimed at a problem, then discover they cannot yet afford the required AI; privacy, security, workforce transitions, and retraining add more friction. Levie therefore expects a minimum decade-long change for enterprises to become genuinely AI-first.
Labenz’s sharper challenge: use deterministic code whenever an explicit algorithm exists because it is faster, cheaper, and more reliable; reserve intelligence for fuzzy work, where a well-contextualized model may already match a human. Levie agrees that an additional commercial obstacle is that “incrementally better, incrementally cheaper, incrementally faster” rarely clears organizational prioritization.
6. Startups need orthogonal markets, not AI replicas of incumbents
A pitch offering the existing process at 40% lower cost sounds compelling in an economics model, yet may rank ninth behind 17 other projects. The buyer must secure AI-council approval and perhaps run a six-month validation, so Levie believes vendors need one-tenth the cost, 10x the quality, or some comparable order-of-magnitude improvement.
Net-new augmentation faces less resistance than replacement. Copilot and Cursor work because users keep coding while receiving immediate productivity gains. Levie separately argues that a 10-times-better result—such as more effective cancer discovery—can justify adoption; work that was not automated before can add capability without first dismantling a staffed process.
Levie’s “timeless” startup rule is to pursue what incumbents cannot naturally absorb. Thin layers over OpenAI or Salesforce are bad bets: “you should anticipate that Salesforce is an AI-for-CRM system,” while Workday and ServiceNow will build AI products for their domains.
Openings remain where workflows cross applications, require extensive non-AI interfaces, or sit orthogonally to incumbent strategy. Levie says independent agent safety, compliance, and red-teaming can be viable, while a vendor spanning Salesforce, HR, ERP, and Box can occupy a layer no single application necessarily owns.
7. Pricing and productivity will change, but diffusion sets the clock
On reports that Klarna shut down a couple of systems of record, Levie’s view is deliberately mixed: the account may be overplayed, but the technical claim is not impossible. Building a homegrown AI Workday replacement to save a few hundred thousand dollars is “super provocative,” yet he doubts 90% of corporations would prioritize it.
AI’s proximity to completed outcomes reopens pricing. A lead-generation agent could charge per lead, per unit of compute consumed to create 10,000 leads, or through a fixed subscription that absorbs volume variation; Box is testing models rather than offering universal guidance after more than 20 years of per-seat SaaS convention.
Inside Box, a new employee learning to sell the product can query the sales hub “like you’re talking to a top expert in the company,” available 24/7. Coding gains reportedly range from 5–10% to possibly 50% for a new hire; the practical KPI is simply shipping more software.
Levie would take the “over” on AI eventually adding more than half a percentage point to annual productivity if the clock starts in a few years. Diffusion through “human-mediated environments” will take longer than expected, even as instant access to roughly “90th-percentile expertise on any topic” makes the underlying technological break unmistakable.