Are SaaS Companies Cooked: Which Thrive & Which Die | Aaron Levie
Summary
Levie’s central labor bet is that AI will produce more engineers and lawyers over the next five years, not fewer. Tech represents perhaps 8-15% of GDP; giving the other 85% Silicon Valley-grade engineering shifts developers toward John Deere, Caterpillar, and Eli Lilly. AI also floods lawyers with draft contracts and memos while courts, patents, and professional approval remain constrained: “We haven’t removed humans from the loop. We’ve just changed where they enter the loop.”
Enterprise adoption’s bottleneck is organizational redesign, creating perhaps 500,000 to 1 million “agent operator” jobs. These technically fluent operators will understand MCPs, CLIs, skills, and AGENTS.md while rebuilding regulated workflows around agents. The resulting software workload is less another phone app than “an unlimited amount of software” connecting fragmented data and automating background processes.
SaaS is not uniformly cooked: agents strip value from button-heavy interfaces but amplify systems rich in APIs, proprietary business logic, governed data, and human review. An ERP is more than its database because supply-chain and accounting logic remain valuable; Box similarly says agents could drive API activity up by potentially 100x or 1,000x. Levie’s mandate: become “the best place where agents want to work with that data.”
Token spend escapes the IT budget and becomes regular operating expense, potentially doubling global technology spend rather than increasing it tenfold. Enterprises could trade a marketing campaign against automation, allocate premium models to the highest-value 5-10% of users, or make teams pitch for compute like contestants on Shark Tank. Unlike startups, large enterprises cannot “token max” indiscriminately: they have annual budgets and, for public companies, EPS commitments.
AI creates a structurally larger cyber market because agents both generate more vulnerable code and let attackers scan faster. As AI moves from writing most code toward 90% or 95%, every additional feature becomes another opportunity to open the wrong port or introduce a flaw. Defensive agents can review that output, but Levie’s formulation captures the circular opportunity: “Agents are the solution to the problem that agents have caused.”
Public software is being treated as an indiscriminately bucketed sector, but Levie expects separation over the next one to two years. Three-times free cash flow can be “aggressively low” for companies that respond effectively, although parts of software were previously valued beyond plausible terminal outcomes; winners will possess deep workflow ownership, valuable data, agent-ready APIs, and monetizable automation. He views Atlassian as potentially oversold because more engineering should expand demand for engineering infrastructure, even as its products must evolve.
Levie would still “load up” on frontier-model rounds despite the host’s test of an $850 billion valuation needing roughly a 3x path to $2.1 trillion. His cloud analogy is AWS at $500 million of revenue in 2010 versus a couple-hundred-billion-dollar ecosystem 15 years later: markets that work can become vastly larger than expected. OpenAI and Anthropic can both win in a multi-model world, while cross-lab infrastructure such as agent evaluations may spawn “a dozen, two dozen, five dozen” new categories.
Deep dive
1. AI is an economic race, not a one-month existential sprint
Levie was “probably 80%” with Jensen Huang against the post-interview commentary. His framing was a “commercial and economic race,” with safety built in—not a binary contest where one or two months of advantage permanently determines the outcome. He also said Jensen had oversimplified some components.
The host’s opening view was that Jensen had come off badly; Levie called the reaction a “Rorschach test” for one’s AI beliefs. Upgrading complex systems takes years, so there is no magical instant when early access secures everything: defense and offense will keep leapfrogging “till the end of time.”
Jensen’s underappreciated warning, as Levie relayed it, was that scaring people away from engineering, radiology, or health care harms society. AI may eliminate review of tiny intermediate steps, but someone still reviews the larger work product: “We’ve just changed where they enter the loop.”
2. Automation exposes bottlenecks—and increases demand for scarce experts
Levie’s categorical prediction: “We will” have more engineers in five years. Tractor manufacturers, banks, and pharmaceutical companies universally say they lack enough engineers; cloud code, Codex, and similar tools let the roughly 85% of the economy outside tech automate at Silicon Valley’s level.
The destination changes more than the skill’s relevance. A computer-science graduate may join John Deere, Caterpillar, or Eli Lilly rather than Google—automating pharmaceutical research, farming, or industrial equipment instead of “building a little app with little buttons.”
The legal example carries the mechanism: AI lets clients generate more contracts, memos, and case material, inundating lawyers with review work. Courts do not approve filings faster, and patents still require accountable professionals, so Levie would “take the other side”: there will be more lawyers within five years.
The host pressed on disappearing junior legal roles; Levie conceded a genuine apprenticeship problem for law firms and banks when AI removes traditional entry-level tasks. But patient-referral automation illustrated his distinction: if the specialist appointment remains 18 months away, the binding constraint is still doctors and institutional capacity.
3. “Agent operators” become a major new enterprise profession
Levie cautiously workshopped a role that might remain full-time or diffuse into other jobs, but was unequivocal on demand: some kind of “agent operator” could create 500,000 to 1 million positions. These workers will understand MCPs, CLIs, skills, AGENTS.md files, and the underlying AI ecosystem.
Embedded inside marketing, legal, operations, or life-sciences research, the operator’s job is to redesign the business process itself: “The workflow needs to be redesigned for agents, not for people.” The key decision becomes where humans review, approve, and assume responsibility for agent output.
Fortune 1000 deployments require change management, organized data, connected systems, and continuing technical care. A new model can break a workflow because its prompting or indexing preferences changed; the talent may therefore come from IT, operations, or engineering rather than one established profession.
4. Agent-ready business logic, not interface acreage, determines SaaS value
Levie accepted that some button-heavy SaaS products are vulnerable: software with “93 features” may have derived value from human familiarity with its interface. When agents replace those clicks, value migrates toward APIs—but API count alone matters less than proprietary logic, security, and permissions.
An ERP is not merely a database; it encodes supply-chain automation and accounting rules that do not disappear. Agents may instead work headlessly across ERP, CRM, HR, and document systems, while applications persist wherever people need to inspect, collaborate on, or approve the work.
Levie’s explicit change of mind was becoming far more convinced over the past year that software will be headless. Two or three years ago, agents routinely found the wrong document and could not reliably open and interpret it; tool-calling and cross-system search improved faster than he expected.
For Box, agents are a force multiplier on an API-heavy model already handling far more machine activity than visible user interaction. The exact dollars per agent remain uncertain, but 100x or 1,000x more calls would expand the opportunity as agents generate contracts, marketing assets, and reports requiring security, governance, and FINRA-compliant retention.
5. Compute moves into operating budgets while cyber risk compounds
AI-generated code introduces risk through sheer volume: as agents approach writing 90% or 95% of code, every shipped feature can accidentally expose a port or create another vulnerability. Attackers also gain faster internet-scale scanning, leaving defensive code-review agents as the one offsetting benefit.
Levie was not newly alarmed; he considered the cyber consequences “priced in” once GitHub Copilot began generating code roughly five or six years ago. Still, the asymmetry supports substantial spending on agentic security: “Agents are the solution to the problem that agents have caused.”
Token allocation will follow enterprise value rather than egalitarian access. Examples included Shark Tank-style pitches for compute and tiering users: unlimited best-model access for the highest-value 5-10%, constrained efficient models for the next cohort, and the cheapest adequate model for general productivity.
The decisive budget shift is from corporate IT into regular OpEx. An enterprise may trade off its next marketing campaign against automating the marketing engine; Levie argued this new access to labor-like budgets could “certainly double” technology spend, while explicitly rejecting a simplistic 10x assumption.
6. Enterprise diffusion creates a decade of services and accountability work
The host worried today’s demand pull might be an 18-month burst driven by every company needing an AI story. Levie took the opposite wager: cloud looked explosively early yet expanded for 20 years, and AI diffusion will take longer than Silicon Valley expects because regulated workflows cannot remove review overnight.
A bank cannot let an agent autonomously produce every client proposal if bad advice could cost its license. Levie’s latest-model experience still required changing about 15% of the output—evidence for substantial human-agent workflow design rather than immediate labor elimination.
He agreed it “rounds to being true” that enterprise sales require an FDE-style model, while rejecting any conflict with product-led growth. AI services should be larger and more durable than assumed because enterprises need systems integration, data curation, workflow mapping, and change management.
One contract-risk query might touch 10 systems, including network shares and legacy repositories; fragmented data can make an agent retrieve the wrong document or contract. Levie called the cleanup and redesign “10 years of work for Accenture”—or for next-generation specialists focused on particular industries and workflows.
7. Liability keeps humans, firms, and institutions firmly in the loop
The host supplied the blunt services thesis: companies need “someone to blame,” just as clients hire lawyers partly because responsibility must attach to someone when an NDA fails. Levie agreed that a customer will not accept “Anthropic made a mistake” after corrupted data or a security breach.
Once liability exists, ownership and reporting structures follow. Human behavior, contract law, and regulatory regimes have not fundamentally changed; computers have merely been handed “a machine gun to go generate way more information and work with all of our data.”
On Chinese open models, Levie said Silicon Valley benefiting from them “must be empirically true.” He preserved a caveat around potentially triggerable backdoor weights or parameters but did not share the host’s alarm—and stressed that even the best frontier models still require accountable human review.
8. Software will separate, while frontier labs and agent infrastructure expand
Jason Lemkin’s provocation was that public companies produce mediocre agents, with Palantir the exception. Levie would not fully endorse it, argued Box has the best content agent, and said leadership now requires following practitioners discussing memory and harnesses—not waiting two weeks for a conventional publication’s recap.
The pace makes Levie’s job harder than ever: technical changes can alter product, strategy, and partnerships several times a week. Public-company CEOs are building with cloud code and Codex on weekends, while simultaneously giving slower-moving customers “a bridge into the future.” For builders, this is a year of “complete, unrelenting execution.”
Box says its new plan tier—combining workflow, automation, application development, and agents—already produced a revenue-growth inflection last year. Yet Wall Street is waiting to see where everyone lands; Levie expects indiscriminate software pricing to separate over one or two years, with three-times free cash flow “aggressively low” for companies that respond effectively.
Calling OpenAI versus Anthropic was “impossible”: AWS had $500 million of revenue in 2010, Azure had just launched, and GCP was still Google App Engine; 15 years later cloud is a couple-hundred-billion-dollar ecosystem. Levie expects multi-vendor adoption and would still fund frontier rounds, plus cross-lab infrastructure such as Braintrust evaluations that detect when agents suddenly mishandle loan-origination documents.